EP0702819A4 - Procede de traitement numerique de l'intensite pixel par pixel ameliorant la lisibilite des informations d'imagerie - Google Patents

Procede de traitement numerique de l'intensite pixel par pixel ameliorant la lisibilite des informations d'imagerie

Info

Publication number
EP0702819A4
EP0702819A4 EP95919731A EP95919731A EP0702819A4 EP 0702819 A4 EP0702819 A4 EP 0702819A4 EP 95919731 A EP95919731 A EP 95919731A EP 95919731 A EP95919731 A EP 95919731A EP 0702819 A4 EP0702819 A4 EP 0702819A4
Authority
EP
European Patent Office
Prior art keywords
hysteresis
image
intensity
image data
cursor
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP95919731A
Other languages
German (de)
English (en)
Other versions
EP0702819A1 (fr
Inventor
Klaus-Ruediger Peters
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Connecticut
Original Assignee
University of Connecticut
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from US08/207,488 external-priority patent/US5563962A/en
Priority claimed from US08/207,489 external-priority patent/US5592571A/en
Application filed by University of Connecticut filed Critical University of Connecticut
Publication of EP0702819A1 publication Critical patent/EP0702819A1/fr
Publication of EP0702819A4 publication Critical patent/EP0702819A4/fr
Withdrawn legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/73—Deblurring; Sharpening
    • G06T5/75—Unsharp masking

Definitions

  • This invention relates generally to the processing of digital image data. More specifically, this invention relates to image data processing using digital techniques for image enhancement and enlargement.
  • Digital image processing has taken an increasing importance as a result of the technological advances in image acquisition and image communication; and can provide advantages over conventional analog image information handling, e.g., undisturbed access of the "raw data set", objective image evaluation, quantitative analysis of the image information, and reduced costs and increased flexibility of image data handling.
  • a complex infrastructure network is in place for high-speed image communication with local, national and international access
  • the general use of digital image processing is hindered through lack of universal standards for identifying image information.
  • visual image perception varies amongst individuals and depends strongly on the image perception and pattern recognition ability. This is the reason why the same image is evaluated quite differently by more than one person.
  • the lack of consistency in image information analysis and display is extremely problematic and creates serious concerns for image evaluation.
  • image communication There are many kinds of information contained in images, but only a few classes may be important in image communication, i.e., detail-oriented (scientific, technical and medical images), composition-oriented (arts, martials science) or information-oriented (binary and CAD, computer assisted drawing). Most important are detail containing images which describe two-or three-dimensional data sets dealing with spacial features. The lack of proper pixel accurate tools for objective description of image details as well as image imperfections produced by acquisition and transmission (noise) limit image communication at this time to information-oriented contents only.
  • imaging e.g., microscopic images derived from SEM or TEM; or medical imagery such as mammograms or x-rays
  • image analysts are limited most by their own visual system (e.g., the human eye) regarding image perception and pattern recognition, since most modern imaging instruments (e.g., microscope, x-ray device, mammography device) provide more data than the eye can process.
  • imaging instruments e.g., microscope, x-ray device, mammography device
  • such data are spacial information documented with certain contrast mechanisms and translated into images. I ⁇ espective of the kind of imaging technique used, the imaging information must be communicated to the visual system for evaluation of its information content at the level of imaging instrument contrast resolution as well as spacial resolution.
  • IR intensity range
  • PW pixels per picture width
  • Image evaluation is primarily a process of pattern recognition which works at a much lower information density than the eye can perceive. Only patterns of large image components of high contrast, high edge sharpness and a few intensity levels (4 bit) are recognized. All other small detail information can only be partially recognized and therefore is commonly generalized as image background or texture. However, in microscopy, radiology, x-ray and other imaging sciences, such background contains a wealth of information of acquired image detail data which is commonly lost in visual analog evaluation.
  • digital image processing methods can be applied for separating image contents on account of certain intensity criteria.
  • the separation of noise and detail structures can be defined by a threshold of intensity variations below which spacial variations are thought to represent noise and are eliminated (smoothed) but above which the intensity variations are defined as significant and are maintained.
  • Conventional image processing methods apply spatially extended processing masks or Fourier filters for the determination of local intensity fluctuations (either in the space domain or in the Fourier domain) and use various methods for determination of the threshold intensity value.
  • the utilized pixel area (mask) and an often used weighing factor applied to the local intensity distribution within the mask will shift the spacial boundary between (smoothed) background and (maintained) detail, altering the spacial dimensions of details.
  • spacial artifacts are produced when structural features are similar in size or smaller than the effective spacial filter area.
  • This problem occurs in all conventional processing modes using spacial kernels, Fourier filters or statistical approaches.
  • certain significant spacial intensity distributions may be seriously altered by eliminating or adding structures, indicating a strong dependency of the processing result on the image content.
  • Such alteration of the spacial content of the original image (raw data set) is a serious limitation of all conventional noise filters in cases where the structural integrity of the image data is important ⁇ i.e., in structure characterization and quantitation.
  • the extent of spacial artifacts in G7 conventional processing depends on the image content. Therefore, complex time-consuming determination of optimal processing parameters are required for each image in order to reduce processing artifacts.
  • conventional image processing speeds are so slow that visual control of intensity threshold adjustments and recognition of processing artifacts are significantly hindered. Therefore, conventional noise smoothing techniques are applied strictly to full frame images, thereby reducing the visualization and recognition of the produced artifacts.
  • novel "smoothing" and “detail enhancement” techniques for processing digital images create a new quality of image perception (centering on enhancement of objective image information) which have wide application in the general field of image enhancement and more particularly in the fields of microscopy, radiology, remote sensing, astronomy, robotics and machine vision and general image communication.
  • the new image processing technology of this invention makes possible a separation of the image information into three objective distinct non-random intensity groups of "large image features" and small “spacial detail” and “intensity detail", and one group of random intensity information of the "image noise”.
  • “Smoothing” allows an elimination of intensity-defined image components and the concomitant reduction of image contrast can be restored by linear contrast stretching of the smoothed image.
  • “Detail enhancement” allows a recovery of intensity-defined image components eliminated by the smoothing process through subtraction of the smoothed image from the original image.
  • the overall contrast range of the recovered information is equal to the applied smoothing factor and can be automatically brought to the full visual intensity range by linear contrast stretching. This has the advantage of maintaining detail contrast proportions which closely reflect the raw data characteristics. If high pixel accuracy processing is provided, "detail slicing" becomes possible. Two smoothed images, each processed with a different smoothing factor, can be subtracted providing the information difference between least smoothed and the most smoothed data set. Again, precise linear contrast stretching can present the extracted information at full visual intensity range.
  • the smoothing technique of this invention comprises a two dimensional digital hysteresis filter which utilizes a variable, automatically adjusting "two-dimensional mask".
  • the filter is independent of the image size and content and cannot alter the size of any structural (significant intensity) features, even if they are as small as only one pixel in size.
  • the two-dimensional hysteresis smoothing technique of this invention calculates smoothed pixels using a set of one-dimensional hysteresis lines at various angles running through each pixel in an image. Each one-dimensional hysteresis line calculates a new value for each pixel in an image.
  • the technique of this invention is preferably implemented on a desktop massively parallel processor that has a large number of 16-bit processing elements (PE's) connected in a ring.
  • PE 16-bit processing elements
  • Each PE has a small, high-speed on-chip data memory and a large off-chip data memory.
  • the PE's all have access to a shared sealer data memory and a shared program memory.
  • a single program is executed by all the PE's in lockstep (SIMD processing).
  • smoothing thus constitutes a method for reducing or smoothing selected intensities in a digitized image data a ⁇ ay comprising a matrix of pixels a ⁇ anged in columns and rows, including the steps of:
  • a somewhat different smoothing technique in accordance with this invention constitutes a method for reducing or smoothing selected intensities in a digitized image data a ⁇ ay comprising a matrix of pixels a ⁇ anged in columns and rows including the steps of:
  • step (e) repeating steps (a)-(d) for at least some of the other pixels in at least a portion of the matrix to define a smoothed digitized image data a ⁇ ay.
  • the hysteresis lines comprise spatially neighboring pixels. More preferably, the hysteresis lines are linear lines radiating at preselected angles through said selected pixel.
  • the smoothing technique of this invention which utilizes a "two-dimensional hysteresis filter" provides many features and advantages relative to conventional digital imaging techniques.
  • conventional image enhancement using fixed small processing masks may not be pixel accurate (as is the technique of the present invention).
  • Spacial artifacts may be as large as the size of the mask and depend strongly on the image content.
  • the processing parameters must be optimized for each image.
  • the image itself is used for the definition of a local processing mask (as in the present invention) such spacial artifacts can be eliminated and the processing result will be independent from the image content.
  • the processed pixel must be related to all other pixels within the "mask" assessing the spacial significance of intensity differences by the smoothing factor.
  • “variable automatically adjusting local mask” produces a processing characteristic equal to point processing since each processed point has a specifically adjusted spacial mask.
  • High precision processing reduces processing artifacts to a level not visible in enhanced images.
  • “detail enhancement” refers to an enhancement of the contrast of image details and must include the spacial details as well as the intensity details in order to maintain the image character (image accuracy).
  • “Spacial details” constitute intensity variations over a short distance (a few pixels long), and “intensity details” constitute intensity variations of a few intensity steps independent of their spacial extent.
  • a desirable enhancement procedure must preserve the unrestricted possibility of image quantitation not only of the spacial content but also of the intensity content; that means the image processing technology must maintain the accuracy of the image at the level of individual pixel's intensity.
  • Such pixel accurate intensity processing (PAIP) for image enhancement is not possible with any conventional technology, but is achievable utilizing the detail enhancement processing technique of the present invention which can fulfill the latter requirements and which is therefore inherently suitable (trustworthy) for scientific and medical applications.
  • the basis for its spacial accuracy is the utilization of pixel-accurate intensity processing; which preferably utilize the "smoothing" technique discussed above.
  • the detail content of the image is reduced on the basis of intensity variations defined by a single processing parameter consisting of the "cursor width" using the two dimensional hysteresis smoothing technique described above (or any other "pixel-accurate” smoothing technique).
  • significantly larger cursor widths are used than in conventional noise management since the image content must not be maintained but instead must be reduced.
  • the detail image is created by subtracting the smoothed image from the original image (or from another smoothed image). The maximum intensity difference of the detail image between any points is equal to the applied cursor width (or the difference of the larger minus the smaller cursor width).
  • the contrast range of the detail image is enhanced by linear contrast stretching with maximum enhancement obtained by utilizing the full width of the intensity range available for visual perception (typically 8 bit).
  • the evaluation of the processing result is dramatically enhanced if the processing occurs in "near-real time" (less than one second) so that an interactive change of the processing parameter becomes possible. This enhances the perception of the image details and their co ⁇ elation with the original image.
  • tail enhancement thus constitutes a method of enhancing the detail in a digitized image data a ⁇ ay comprising a matrix of pixels a ⁇ anged in columns and rows including the steps of:
  • the digital enhancement technique of the present invention provides many features and advantages relative to conventional digital imaging techniques.
  • the detail enhancement filter works principally different from other conventionally used detail filters, i.e., the Oho filter which is described in Oho E., Automatic Contract Adjustment for Detail recognition in SEM Images On-Line Digital Image Processing, Scanning 14: 335-334 (1992).
  • the Oho filter is a highlight filter for edge enhancement of spacial image details only (it specifically suppresses intensity detail enhancement). It extracts small area intensity changes (highlights) from the original image by use of a fixed large mask size median filter (19 x 19 for lKxlK images).
  • the use of a median filter has many limitations.
  • the filter maintains local intensity variations and when the median is subtracted from the original image, the intensity details are completely removed from the selected detail data.
  • the filter cannot select a certain intensity range. This has the serious implication that the intensity range of the detail image depends fully on the image content. Therefore, only histogram equalization can be used for contrast enhancement of the filtered details which may cause spacial contrast artifacts.
  • the Oho filter erodes fine structures at the level of its mask size at edges and at small multiple intensity variations. As a consequence, the enhanced detail image must be added to the median filtered image in order to restore (in part) the eroded fine structure.
  • the new PAIP detail enhancement filter of this invention produces quite different information as compared to that obtained by the Oho filter and, presents all image information in an exhaustive fashion with linear contrast enhancement.
  • the new detail filter produces detail images which characterize the intensity distributions within an image. It provides a tool for a novel method of image information analysis and classification based on the concept that any image communicates information only through image contrasts which are intensity variations between certain pixel a ⁇ ays within the total data matrix. An area of certain contrast is defined by the differences between its average intensity and the su ⁇ ounding intensity i ⁇ espectively of the overall intensity variations (background).
  • the new detail filter of this invention provides the only tool available for selecting such local intensity variations independently from the other intensity variations within the full data matrix through only one parameter which is the intensity range (equal to cursor width or significant intensity range).
  • Application of the filter with increasing intensity ranges selects from the image contrast components of specific visual information contents in a defined and principle manner which matches the visual pattern recognition mechanisms.
  • digital images represent two- (or three-) dimensional intensity maps which characterize the spacial x/y(/z) location of all contrast information.
  • the described detail filter of the invention Since the described detail filter of the invention generates all basic visual information classes it is best suited for information analysis, quantitation and communication of such technical images.
  • the ability of categorizing and quantifying image information also provides a tool for objective measurement of image quality.
  • the intensity ranges (significant intensity ranges) of each image intensity class are proportionally distributed within the overall intensity range of the raw data in the sequence as generated and displayed by the detail filter: first the noise, then the spacial detail, the intensity detail, and the image feature/background.
  • a graphic display of the image information components and their proportion within the intensity range as schematic intensity maps in an "information cube" facilitates visual image quality assessment (see FIGURE 16). Either the proportion of the noise component to the adjacent intensity information of any width, or the relation of any component to any other or the overall intensity range provides an objective tool for categorizing images and image quality.
  • Such graphic display of the image information content will facilitate image analysis and communication.
  • enhancement of image detail contrast is accomplished by adding a differential hysteresis pattern to a digital image.
  • the differential hysteresis image processing described hereinbefore, utilizes the persistence of intensity variations (i.e., hysteresis) as means for data reduction and image detail contrast enhancement.
  • Image hysteresis is determined by a hysteresis cursor of an interactively chosen hysteresis range, as described with regard to the above embodiments.
  • FIGURE 1 A is a block diagram illustrating the process of the present invention
  • FIGURE IB is a block diagram of a system in accordance with the present invention
  • FIGURE 1C is a flow chart depicting the smoothing technique of the present invention which utilizes a two-dimensional hysteresis filter for noise reduction;
  • FIGURE 2 is a diagrammatic example of 45 degree hysteresis lines covering an image
  • FIGURE 3 is a flow chart of the "group" processing operation used in the smoothing technique of the present invention
  • FIGURE 4 is a diagrammatic representation of the data structures used in the smoothing technique of the present invention
  • FIGURE 5 is a graph depicting the relation of digital image information with visual pattern recognition parameters
  • FIGURE 6 depicts in Section "A” the principle approach for intensity information extraction; and in Section “B” a group of graphs depicting different types of intensity defined information including “noise”, “spacial detail”, “intensity detail” and “feature/ ackground” which are contained in digital images and in Section “C” a schematic representation of the intensity extent of the various information groups;
  • FIGURES 7A-F depicts a test pattern of Gaussian noise having been processed using the noise smoothing technique of this invention with and without prior randomization of the image
  • FIGURES 7G-H depicts a test pattern of Gaussian noise and conventional noise reduction techniques
  • FIGURES 8A-H are images of low magnification SEM data depicting the information classes and pixel-accuracy of the detail enhancement using the detail enhancement techniques of this invention
  • FIGURES 9A-B are images of high magnification SEM data having been processed using the detail enhancement techniques of this invention
  • FIGURES 10A-B are images of high magnification field emission SEM data having been processed using the detail enhancement techniques of this invention.
  • FIGURES 11 A-B are low voltage field emission SEM data having been processed using the detail enhancement techniques of this invention.
  • FIGURES 12 A-B are environmental SEM data having been processed using the detail enhancement techniques of this invention.
  • FIGURES 13 A-B are high resolution field emission SEM data having been processed using the detail enhancement techniques of this invention
  • FIGURES 14A-D are cryo-TEM energy filtered, phase contrast data having been processed using the detail enhancement techniques of this invention
  • FIGURES 15 A-B are high voltage dark field TEM data having been processed using the detail enhancement techniques of this invention
  • FIGURES 16 A-B are high resolution TEM data having been processed using the detail enhancement techniques of this invention
  • FIGURES 17 A-B are high magnification scanning transmission electron microscopy data having been processed using the detail enhancement techniques of this invention
  • FIGURES 18 A-B are images of atomic force microscope or AFM data having been processed using the noise smoothing and detail enhancement techniques of this invention
  • FIGURES 19 A-B are images of light microscope data having been processed using detail enhancement techniques of this invention
  • FIGURES 20A-D are images of confocal laser light microscopy data having been processed using the noise smoothing and detail enhancement techniques of this invention
  • FIGURES 21A-D are images of mammogram data having been processed using the noise smoothing and detail enhancement techniques of this invention
  • FIGURE 22A-D are images of chest x-ray data having been processed using the noise smoothing and detail enhancement techniques of this invention
  • FIGURE 23 is a group of graphs depicting quantitation of image data "information cubes" using the smoothing and detail enhancement techniques of this invention
  • FIGURE 24 is a series of plots diagrammatically illustrating hysteresis line processing
  • FIGURE 25 is a plot diagrammatically illustrating hysteresis image processing radially in each pixel
  • FIGURE 26 is a block diagram illustrating differential hysteresis image processing in accordance with the present invention.
  • FIGURE 27 A is an original image of a human face
  • FIGURE 27B is an extracted hysteresis noise pattern image obtained from the original image of FIGURE 27 A
  • FIGURE 27C is an extracted spacial hysteresis details image obtained from the original image of FIGURE 27 A;
  • FIGURE 27D is an extracted intensity hysteresis details image obtained from the original image of FIGURE 27A;
  • FIGURE 27E is an extracted hysteresis image feature obtained from the original image of FIGURE 27 A;
  • FIGURE 27F is a composite hysteresis pattern image obtained from the images of FIGURES 27B - E;
  • FIGURE 28 A is an original image of a human face
  • FIGURE 28B is an enhanced hysteresis details image of the image of FIGURE 28 A;
  • FIGURE 28C is an enhanced intensity hysteresis details image of the image of FIGURE 28A;
  • FIGURE 28D is an enhanced spacial hysteresis details image of the image of FIGURE 28A;
  • FIGURE 29A is a raw data image of an ultra-thin section of plastic embedded retina tissue;
  • FIGURE 29B is a differential hysteresis pattern image of a DHR 1-65 contrast range of the image of FIGURE 29 A;
  • FIGURE 29C is a differential hysteresis pattern image of a 20% DHR 1-64, 80% DHR 1 -9603 contrast range of the image of FIGURE 29 A;
  • FIGURE 29D is a differential hysteresis pattern image of a DHR 1-3 contrast range of the image of FIGURE 29 A;
  • FIGURE 29E is a differential hysteresis pattern image of a 15% DHR 1-3, 85% DHR 1-9603 contrast range of the image of FIGURE 29 A
  • FIGURE 29F is a differential hysteresis pattern image of a 40% DHR 1-3, 60% DHR 1-64 contrast range of the image of FIGURE 29A;
  • FIGURE 30 A is a computerized tomography cross section image
  • FIGURE 30B is a differential hysteresis pattern image of a DHR 1-256 contrast range of the image of FIGURE 30 A;
  • FIGURE 30C is a differential hysteresis pattern image of a DHR 27-35 contrast range of the image of FIGURE 30 A;
  • FIGURE 30D is a differential hysteresis pattern image of a 40% DHR 27-35, 60% DHR 1-256 contrast range of the image of FIGURE 30A;
  • FIGURE 30E is a differential hysteresis pattern image of a DHR 21-23 contrast range of the image of FIGURE 30 A;
  • FIGURE 3 OF is a differential hysteresis pattern image of a 40% DHR 21-23, 60% DHR 1-256 contrast range of the image of FIGURE 30A;
  • FIGURE 31 A is a digital Fugi plate image
  • FIGURE 3 IB is a differential hysteresis pattern image of a DHR 1-256 contrast range of the image of FIGURE 31 A;
  • FIGURE 31C is a differential hysteresis pattern image of a DHR 9-15 contrast range of the image of FIGURE 31 A;
  • FIGURE 3 ID is a differential hysteresis pattern image of a 25% DHR 9-15, 75% DHR 1-256 contrast range of the image of FIGURE 31 A.
  • the present invention comprises several related digital image processing techniques including a novel "smoothing" or “data reduction” technique which utilizes a two-dimensional hysteresis filter for noise reduction and a novel “intensity enhancement” technique which enhances the “smoothed” or “reduced” data for selected spacial details and intensity levels.
  • a novel "smoothing” or “data reduction” technique which utilizes a two-dimensional hysteresis filter for noise reduction
  • a novel "intensity enhancement” technique which enhances the “smoothed” or “reduced” data for selected spacial details and intensity levels.
  • the novel “smoothing” technique (which may also be used in the "detail enhancement” technique) will be described first. I. TWO-DIMENSIONAL HYSTERESIS SMOOTHING
  • top streak is produced as well as a “bottom streak” at the level of low intensity variations (bottom of cursor).
  • the length of these streaks depends upon the linear characteristics of the data set.
  • the linear (e.g., one-dimensional) smoothing technique of the prior art is performed in reversed direction, and both new data sets are arithmetically averaged maintaining their cross-registration.
  • the dual direction processing of the prior art has significant advantages and disadvantages. Since the maximum intensities of a structure ("top" and "bottom") are maintained when read in both directions, the height and position of the structure is maintained in the averaged data (only reduced or increased from the raw intensity by the value of the reference point on the cursor). At both shoulders of the structure, within the streaks, the slope of the raw data set is modified along the streak until again the raw data are read providing for the continuity of the smoothed data set. In some regions, the streak intensity may be different in both directions and an averaged "background" intensity will result. In order to minimize border artifacts at the beginning and end of the linear data set, at each start of reading, the cursor reference point is positioned on the original data value.
  • the two-dimensional hysteresis smoothing technique of this invention calculates smoothed pixels using a set of one-dimensional hysteresis lines at various angles running through each pixel in an image.
  • the technique requires an input image and two parameters, (1) the number of hysteresis lines per pixel, and (2) the cursor width for the hysteresis algorithm.
  • the program calculates an output image of the same size as the input image.
  • a suitable computer program was written in assembly language for the AP X desktop massively parallel processor manufactured by Visionary Systems Inc. of New Haven, Connecticut and described in detail in the paper entitled "The AP X Accelerator", E.
  • the AP X parallel processor has up to 256 16-bit processing elements (PE's) interconnected in a mesh topology.
  • PE has a fast 256 word on-chip data memory and a 65536 word off-chip data memory.
  • the PE's all have access to a shared sealer data memory and a shared program memory.
  • a single program is executed by all the PE's in lockstep (SIMD processing).
  • the AP X uses a PC-AT clone as a Host computer, and the AP X off-chip PE memories are memory-mapped into the Host computer's address space.
  • FIGURE 1 A flowchart of the overall two-dimensional hysteresis smoothing program of this invention is shown in FIGURE 1. It performs eight passes over an input image, with each pass co ⁇ esponding to hysteresis lines in a particular angle range. Table 2 shows the characteristics of each angle range. The actual hysteresis lines are at equally spaced angles around a 360 degree circle. For example, if the number of hysteresis lines is 16, there will be hysteresis lines at angles of 0, 22.5, 45, 67.5, 90, 112.5,...,270 and 292.5 degrees and they will be partitioned into eight angle groups of two lines each.
  • FIGURE 2 shows an example of 45 degree hysteresis lines covering an image. Each hysteresis line calculates an output value for each pixel in the image. The final output value for a pixel is the average of all the output values for that pixel. TABLE 2
  • the program does one pass over the input image for each of the eight angle groups, doing the calculations for all the hysteresis lines within that group.
  • a flowchart of the group processing is shown in FIGURE 3.
  • the program steps sequentially through the image rows beginning with the Starting Edge from Table 2. If the Starting Edge is either Left or Right, the input and output images are transposed before and after the pass, so that the program can step from row to row by incrementing or decrementing its memory address. Each row is read in turn and each hysteresis line in the group is applied to it. Hysteresis midpoints are then conditionally shifted to the left or right neighboring PE's, to maintain the co ⁇ ect angle of the hysteresis line, and the process repeats for the next line. The conditional shifting is based on the line patterns stored in the sealer memory.
  • the hysteresis calculation is as follows: if the input pixel value is less than the cu ⁇ ent cursor midpoint minus half the cursor width, the midpoint is changed to the input pixel value plus half the cursor width; if the input pixel value is greater than the current cursor midpoint plus half the cursor width, the midpoint is changed to the input pixel value minus half the cursor width. The output value for the pixel is the resulting cursor midpoint. This is added to the midpoints for other hysteresis lines applied to the same pixel. At the Starting Edge, the cursor midpoints are initialized to the input pixel values.
  • cursor midpoints When cursor midpoints are shifted off the edge of the image, they are re ⁇ initialized with the value of the edge input pixel. Initialized cursor midpoints are clamped to their valid range; from the minimum possible pixel value plus half the cursor width to the maximum possible pixel value minus half the cursor width.
  • the input and output images are stored in the large, off-chip PE data memories.
  • Each PE stores at least one column of the image in its memory as shown in FIGURE 4.
  • the number of PE's is the same as the number of columns in the image to be processed and that each PE stores one column from each row. If there are more PE's than columns, the extra PE's are simply disabled and do not participate in the calculations. If there are more columns than PE's, each PE does the processing and storage for multiple columns.
  • each output pixel holds a sum of the values calculated by the hysteresis lines running through the pixel. The final output image is calculated by dividing the pixel sums by the number of hysteresis lines.
  • Each on-chip PE data memory holds cursor midpoints for all the hysteresis lines running through the cu ⁇ ent pixel in the cu ⁇ ent angle group. It also temporarily stores each input row as it is being processed, and the output sums for the pixels in that row.
  • the sealer data memory holds line patterns for sets of eight hysteresis lines. One line pattern represents eight hysteresis lines whose angles are offset in increments of 45 degrees. This pattern represents the angle of the hysteresis line in terms of vertical/horizontal and diagonal steps of one grid unit. It is used to control the conditional interprocessor shifting of cursor midpoints in the hysteresis processing.
  • the optimal number of iterations (of new reading frames necessary for artifact suppression) will depend on the raw data set and/or the cursor width. Insufficient averaging will produce visually destructive contrast jumps along the reading directions. Such artifacts are easily seen in straight (radial) reading directions and in images of non linear structural characteristics. Eight to thirty two different linear reading directions may be sufficient in reducing the artificial linear background intensity fluctuations so as to be non-recognizable by contrast analysis in average images even if small features of high contrasts are present. In addition, for the rare cases where streaks are persistent, sixty four to two hundred fifty six (or more) iterations are provided.
  • the number of iterations required can be set automatically from the chosen cursor width according to a semi-empirical evaluation of effectiveness or from calculations using the selected cursor width and contrast properties of the image.
  • other non-linear reading directions may be provided in order to distribute the background intensity fluctuations in such a manner that it will not be recognized by visual perception.
  • Such a non-linear method may be applied to an unsatisfactory result of the linear method for the purpose of redistribution of the background intensity fluctuations into a non recognizable or non distracting pattern.
  • the information content of a digital full frame image can be categorized into a non-random (structural) class of features and details, and a random class of noise.
  • a feature is an image component accessible to visual pattern recognition, i.e., it is larger than - 10% of the picture width, has a contrast range of more than -5% of the visual intensity range, and each must fall within no more than 15-20 intensity levels in order to be recognized as a pattern. All other structural image components are summarized as details.
  • small high contrast components ⁇ 10% PW, >5% BW
  • low contrast components of any size ⁇ 5% BW
  • digital images contain noise of various origins. Only the high frequency noise components at the level of a few pixels are visually perceivable as random intensity fluctuations. Other low frequency noise components, which have the character of intensity details, may be recognizable only after eliminating (averaging, smoothing) the high frequency components.
  • noise is indistinguishable from small spacial detail lacking a dominant pattern. Noise will also disrupt the integrity of larger structural components in proportion to its strength. Thus, if visually distracting noise components are reduced by smoothing, some detail will be lost. Additionally, if the noise has a Gaussian characteristic, some noise pixels of extreme intensities will always remain as well as some low frequency components.
  • Image evaluation of microscopy, radiological and other similar data requires that all image details are made to be identifiable by visual pattern recognition while maintaining the spacial relation of details within the overall image. This implies that low intensity details be contrast enhanced by a factor of 10-100 and that the smallest spacial details be enlarged by a factor of 10-20. However, for co ⁇ elative image component evaluation, only three image processing tasks are required.
  • Noise management In general, the evaluation of an image's information content will require a sequence of image processing steps: first, in the full frame image after contrast enhancement, details will be recognized and cross co ⁇ elated with image features; then interesting, enhanced details will be enlarged and analyzed; and finally, distracting noise may be reduced. If the image contains a high noise level, the noise is first reduced to a level at which the detail information can be easily recognized.
  • image processing must fulfill some stringent conditions in order to be effective and practical.
  • the image processing techniques must avoid processing artifacts inherent to most common image enhancement procedures which distort the spacial and most of the intensity characteristics of details. Only pixel accurate enhancement techniques promote closer visual inspection by digital enlargement.
  • the processing must be fast (close to real time) in order not to disturb the visual recognition process; it must be able to automatically accommodate all images independent of their size, depth and content, it must be exhaustive and objective to avoid missing any existing detail; and it must be simple and without any other input than a single factor, i.e., a "visibility enhancement factor".
  • digital image processing in accordance with the present invention can provide visual access to acquired digital image data at the level of instrumental image resolution by extracting and imaging intensity defined image information classes. This data reduction fosters visual pattern recognition.
  • Digital image data sets are intensity maps showing the intensity at each pixel in a two dimensional a ⁇ ay of pixels.
  • an intensity profile graphs this intensity along a line of pixels (FIG.6, upper left box), and a series of intensity profiles outlines in a three dimensional graph the three dimensional "intensity profile" surface of the data in x and y direction.
  • the intensities surface may vary in height and spacial extent indicating image components with certain spacial and intensity characteristics. Four different image components can be visually identified since they match basic visual perception patterns. Random intensity fluctuations over smallest distances (single pixels) are characteristic of "noise" (FIG. 6, bottom left box).
  • the noise pixels In the three dimensional intensity profile surface, the noise pixels would appear as slender spikes or holes covering larger intensity components.
  • the average intensity variation of noise in Gaussian noise +/- 2 standard deviations ⁇ SD ⁇ similar to the typical bandwidth display of analog noise
  • IR S intensity bandwidth
  • noise occupies only a few percent ( ⁇ 10%) of the total intensity bandwidth too small to be visually recognized.
  • extracting the noise component FIG. 6, bottom left box; upper intensity profile
  • contrast stretching it to the full visual bandwidth will make the noise pixels visible and identifiable by their spacial random distribution (FIG. 6 bottom left box: bottom intensity profile and map).
  • noise is always dominant, e.g., in low dose imaging or in high magnification imaging.
  • lack of contrast prevents acquisition of high quality data and will result in poor data quality (noise JR_ > 20%).
  • Homogeneous intensity variations over areas of many pixels are nonrandom and present the spacial information of the data.
  • three major visual pattern types can be distinguished. Small low contrast components are summarized as "spacial detail" (FIG. 6, bottom middle box). If the components are uniform they will occupy an identical significant intensity range (IR S ) above the background intensity. In the three-dimensional intensity profile, the spacial details will appear as well defined small little "bumps" or "invaginations”.
  • Extraction and contrast stretching provide for easy perception of these details and recognition of the distribution pattern.
  • IR S 2-20% of the total intensity range.
  • spacial details may be a dominant contrast component.
  • features Many weak contrast mechanism produce intensity details and their recognition is an important part of microscopy or other imaging services. Intensity details are the least accessible in analog imaging since the visual system cannot easily recognize large low contrast components which often do not provide shape edges. However, if extracted from the raw data and contrast enhanced, these data become strikingly accessible for image evaluation due to increased edge contrast. Linear contrast stretching will maintain proportionality between different intensity details and thus facilitate interpretation. Finally, most image data have some large features which dominate the full frame image (FIG. 6, top right box). Their dominance comes from a nearly complete occupation of the available data intensity range. Nearly always, other details and noise are dwarfed by the feature's intensities and become hidden from visual perception since only few intensity steps remain for their accommodation.
  • the information extraction program in accordance with this invention creates an output image from two smoothed input images (or the original data set and one smoothed image).
  • the smoothed images are preferably generated by the two- dimensional hysteresis smoothing program described above using two different cursor widths (but alternatively may be generated by a different pixel-accurate smoothing technique).
  • the two input images must have the same number of rows and columns.
  • the detail extraction subtracts the most smoothed image from the least smoothed image and then does linear contrast stretching on the result.
  • a computer program implementing the present invention is written in assembly language for the aforementioned Visionary Systems AP X desktop massively parallel processor.
  • processor has up to 256 16-bit processing elements (PE's) interconnected in a mesh topology.
  • PE 16-bit processing elements
  • Each PE has a fast 256 word on-chip data memory and a 65536 word off- chip data memory.
  • the PE's all have access to a shared sealer data memory and a shared program memory.
  • a single program is executed by all the PE's in lockstep (SIMD processing).
  • SIMD processing lockstep
  • the AP X uses a PC-AT clone as a Host computer, and the AP X off- chip PE memories are memory-mapped into the Host computer's address space.
  • the image subtraction is done in one pass over the two smoothed images.
  • the linear contrast stretching is performed in two passes over the output of the image subtraction.
  • the first pass finds the minimum and maximum pixels values in the image.
  • q(x,y) is the initial pixel value
  • q(x,y) is the final output pixel value
  • minpix is the minimum pixel value in the image
  • maximumpix is the maximum pixel value in the image
  • maximumval is the maximum pixel value.
  • An important feature of the detail enhancement technique of the present invention is the use of the aforementioned process for two-dimensional hysteresis smoothing which avoids commonly encountered processing artifacts.
  • the "smoothing" technique is applied for noise reduction and due to its unique pixel accurate design, maintains the spacial information of the smoothed image.
  • the image details eliminated from the smoothed image may be recovered from the original image without loss of the detail pixel accuracy. Since the contrast range of the extracted detail is limited and falls into the cursor width applied for the smoothing, the contrast enhancement of the detail information becomes predictable. This is an important advantage since it allows the extraction and enhancement for any range of image details.
  • a second important and novel feature of the detail extraction and enhancement technique is in its ability not only to selectively extract details but also in extracting the contrasts of larger image features by as much as its cursor width.
  • This observation has led to the definition of the present invention as detail enhancement in full frame images since the image information eliminated through the smoothing contains both the detail information and feature information. The portion of feature information is proportional to the cursor width.
  • the image information recovered after smoothing (through subtraction of the smoothed image from the original image) can be contrast enhanced through only one step of linear contrast stretching and maintains in principle all image information. Therefore, the detail contrasts are more enhanced than the feature contrasts which are in fact first selectively reduced.
  • Another novel feature of the present invention is the discovery that the contrast range of all extracted information (before enhancing) is exactly known (intensity range of the image minus cursor width), thus the contrast enhancement factor is predictable and easily established.
  • the defined contrast range of the extracted information leads to a third feature of the detail enhancement technique of this invention, "intensity slicing". Smoothing of the original data with two different cursor sizes allows extraction of the information difference between both smoothed images.
  • This intensity slice has the same property as the extracted details: its significant intensity range is known and it contains a proportion of the feature information dependent on the applied cursor sizes. Thus, the information contained in an intensity slice can be displayed after linear contrast stretching.
  • the development of this new process leads to a new unique definition of the information content of digital images (as containing four different intensity defined contents, i.e., noise, spacial details, intensity details, and image features/background) which will have great impact on image communication, image analysis, pattern recognition and image quantitation.
  • the data reduction to individual intensity information classes provides a new and unique tool for the analysis of image information and quantitation of the image information content and image quality.
  • the application of this tool in digital image processing lead to the discovery that all known contrast mechanisms of any technical imaging equipment (microscopes, telescopes, photographic and video cameras, medical imaging technologies, etc.) establish one of the three basic intensity variations found as structural classes in digital images.
  • every image information structural component
  • contrast mechanisms which generate specific intensity differences between image components and their su ⁇ ounding.
  • the contrasts may directly image the complete structural components, part of the components (phase contrasts, diffraction contrasts) or none, requiring a complete spacial reconstruction (interference contrasts of holograms).
  • each of these different image contrast types can only be established in a data set as one of the three non-random information classes. Therefore, the detail enhancement technology described herein can separate and extract different contrast types for the image data and present the specific contrast information as a separate image which is easily visually recognized and objectively quantified by its significant intensity range. Since the contrast types match the visual perception and recognition parameters, such extracted contrast information is easily visualized. Connectional visual quantitation is possible only in a limited and subjective way. However, image contrast quantitation now becomes possible through the measurement of the significant intensity ranges of each contrast type. The quantitation also allows the establishment of the intensity proportion of each contrast type with the other image intensity components including noise and intensity background. This procedure provides the first objective way for image quality assessment describing the proportion and extent of the intensity components of an image.
  • a graphic display of the intensity ranges of each of the image components in an "information cube” allows easy visual access to the image quality.
  • the individual "detail information contrasts” have an intensity range of only a few percent of the overall intensity range, such detail can be extracted from high precision raw data of 10-bit to 16-bit accuracy and visually displayed on an 8-bit level without compromise of the raw-data accuracy.
  • the new information enhancement technology of this invention thus makes possible a quantitation of the image information and provides objective criteria for image quality assessment. It makes possible objective image communication via electronic networks on "high speed highways" since accurate data reduction generates clearly displayed image information without any distortions or artifacts.
  • the novel detail enhancement technique of this invention utilizes two- dimensional hysteresis processing for several unique enhancement purposes which are required for the accurate enhancement of image details and which are not addressed either by the noise management application of the technique or by any other commonly used processing techniques: pixel accurate extraction of spacial details with a defined and limited intensity range, and image-accurate scaling of detail intensity and determining of a single interactively defined output parameter for the enhancement.
  • the detail enhancement filter has fundamental advantages over the conventional spacial processing principles (Oho filter, Sobel filter or homomorphic filters) since it uses a different, pixel accurate intensity processing principles for the selection, extraction and enhancement and which are independent of the image content.
  • the hardware used for image processing provides for adequate high speed processing using a 486/66 MHz PC-AT host CPU and a high level PC based AP X parallel processing technology.
  • the a ⁇ ay processor (AP) technology is based on single instruction/multiple data (SIMD) architecture using an expandable system of 64 to 256
  • 16-bit processors which provide peak instruction rates of 800-3200 MIPs.
  • the individual processors are 16-bit RISC processors which can be software configured to 32-bit mode.
  • IEEE format single precision floating point operations are supported in 32-bit mode with peak ratings from 40-160 MFLOPs.
  • VLSI technology allows fast one-cycle communication of 32-bit numbers.
  • the AP X processor boards fit into PC bus slots and provide supercomputer performance. Workstations of this type are commercially available from Visionary Systems, Inc. 25 Science Park, New Haven, Connecticut.
  • the standard image format is square and the image is displayed together with a simple menu on a 20 inch workstation monitor with 1280x1024 pixel resolution and 120 Hz refresh rate (Hitachi CM2085MU SuperScan20 monitor, allowing full stereo display with a Stereographies CrystalEyes System).
  • the monitors are provided with custom fitted anti-magnetic Nu-metal shieldings.
  • Image enhancement of a typical high quality image requires 3 billion instructions per second (3000 mips). This requires supercomputer processing speed.
  • a PC based parallel processing system such as the aforementioned APx system constitutes an important feature of this invention.
  • the aforementioned APx system processes 1600 mips and is therefore capable of processing an image in 2 seconds.
  • a larger APx a ⁇ ay will process the image much faster.
  • a conventional smaller image 512 pixels 8 bit
  • Such very short processing times are provided only by a parallel processing system.
  • the noise smoothing capability of the digital process imaging of this invention can be demonstrated using a test pattern derived from Gaussian noise of +/- 4 standard deviations (SD) width (1024xl024x8bit, mean pixel value of 127, minimum and maximum pixel values of 43 and 211, standard deviation of 21, and 0.0001% clipping accuracy) (FIG. 7A).
  • SD standard deviations
  • a superimposed spacial test pattern consisted of only one pixel wide features, i.e., two perpendicular double lines of 0 and 255 intensities, and two sets of small crosses of either +/- 2 SD (top half) or +/- 3 SD (bottom half) intensities (FIG.
  • noise reduction algorithms utilize spacial masks (in the space domain or Fourier filters) and may produce spacial distortions of image details at a maximum level set by the mask or filter size.
  • the PAIP technique of this invention maintains the spacial integrity of image details at the precision level of the raw data
  • PAIP image enhancement revealed that all expected local contrasts were generated and collected at low, as well as at high, magnification but that detail contrasts were compressed in inverse proportion to the extent of the feature contrasts or background level.
  • the visual recognition of local, small intensity variations was reduced in close proximity to large bandwidth intensity variations.
  • PAIP image enhancement in accordance with the present invention provides an easy and fast procedure for adjustment of the proportion of detail and feature/background intensity range by interactively reducing the latter and visually evaluating the effectiveness.
  • FIGURE 8F The image quality changed dramatically once the electron beam induced contrast mechanisms were visualized. Especially, intensity details (micro- roughness) contrasts on the specimen support and spacial details (edges along the chitinous plates covering the animal, and smallest hairs and bristles) became dominant. The enhanced image proved that the electron probe was capable of detail imaging despite an overwhelming contrast range produced by other mechanisms. Quick adjustment at the significant intensity range in the overall intensity or an intensity slice was essential for seeing what signals the electron probe was generating and thus promoted contrast interpretation. Detail evaluation by digital magnification was facilitated if the image's intensity range was appropriately adjusted to the size of dominating contrasts.
  • the enhancement revealed truthfully (pixel accurate) fine structures and minute contrasts produced by the electron probe on the sample surface expanding scanning electron microscopy to scanned electron probe microscopy. Besides spacial information, small topography contrasts (relief contrasts) became identifiable.
  • Imaging at high performance conditions (30 kV, field emission electron source) produced only weak detail contrasts (Figure 10 A: 2x bicubic zoom) which did not allow an identification of the device architecture.
  • An intensity slice underneath the noise component of only 4 intensity steps (Il- 7-11) included all spacial detail of the raw data and revealed the cross section's topography in short range contrasts at the precision level of the electron beam (individual pixels) ( Figure 10B).
  • Such enhancement is invaluable for routine FSEM application.
  • the new low vacuum SEM technology images non-conductors and wet surfaces, both often consisting of samples with low mass density and low signal yield. Higher magnification images therefore are limited by a very large noise component in addition to the common large signal background component. At medium magnification
  • the new ultra-high resolution FSEM instruments provide a 0.5 nm probe diameter and are capable of working at magnifications of 100,000 - l,000,000x in order to take advantage of the high theoretical resolution.
  • contrast quality sets the limitation at high magnification.
  • high magnification even in these in-lens microscopes, a large signal background is generated which compresses the high precision short-range contrasts (spacial detail information).
  • High magnification cryo- imaging 100,000x, 30 kV, sample temperature -120°C
  • of a molecular preparation on thin C film (2 nm thickness) shadowed with a 1 nm continuous Cr film, produced only disappointingly low contrasts with no molecular details being visible (Figure 13 A).
  • TEM contrasts are a good example for the superposition of various contrast mechanisms and the difficulty in visual contrast information interpretation in the space domain (the image).
  • the image In the TEM at the level of smallest structural details, both phase contrast and scattering contrast occur and are superimposed.
  • Phase contrasts can be easily analyzed and reconstructed in the Fourier domain using the transfer function of the optical system and microscope imaging parameters. Low-dose imaging is a prerequisite for beam sensitive materials but increased noise and lack of contrast hinders immediate evaluation of acquired raw data. In addition, the image quality may be obscured by beam damage, insufficient dose, and other factors (contamination, instabilities etc.).
  • Cryo-TEM Cryo-Transmission Electron Microscopy
  • the phase contrast transfer function relates a specific intensity characteristic to each of the spacial frequencies.
  • This inherent intensity characteristic of the TEM contrasts can be determined by the processing technology described in this application and used for a separation and imaging of each of the existing particular spacial frequencies of the raw data. Even if the energy filtering is applied for generating an enriched phase contrast data set, other signal component with the same energy range are included and produce a large background signal (up to 80% of the total signal as can be shown here). Intensity background reduction through intensity slicing maintains the spacial character of the data and generates enhanced phase contrast images enriched in certain spacial frequencies.
  • This imaging technology is especially important for high resolution cryo- TEM in phase contrast on frozen-hydrated biological materials.
  • the different spacial phase contrast component were enriched in different successive intensity slices proving the coherence of the intensity profile and validating the new image information processing technology.
  • Interactive visual control facilitated the determination of a significant intensity level at which noise reduction and preservation of fine structural detail were optimized.
  • the latter component contained the knife marks (originating from the cryo-sectioning) as well as ice crystal contamination. Separation of the phase contrasts or its various components will facilitate the reconstruction of the specimen's ultrastructure.
  • High voltage microscopy provides for high resolution but is limited by a reduction of image contrast. Therefore, often dark field microscopy is used for a recovery of some of the low contrasts components.
  • the extent of collected fine structural information in such high precision data was unknown but can be assessed through intensity slicing.
  • a 300 kV TEM image (60,000x magnification in conical dark field illumination) of mineral platelets coated with surfactant reveals only few fine structural details (Figure 15 A: 2x zoom) seen in some darker areas in between plates
  • Scanning transmission electron microscopes are important instruments in R&D due to their ultra high resolution capability, high depth of field and limited demand for specimen preparation.
  • the imaging capabilities of these instruments are limited as well as other microscopies by the proportion of high precision, short range contrasts and their intensity background.
  • the confocal laser light imaging technology aims for a reduction of the signal background produced from light scattering with the sample.
  • low signal yields and high noise levels are characteristic for the CFLM data.
  • noise management is important and can be facilitated through the intensity slicing since random and non-random (structural) information can easily be visually recognized and assessed.
  • Mammogram evaluation assesses the tissue structure of the mammary gland. Two stages of tumor growth pattern can be distinguished: 1. Early indications are seen in widening of gland ducts and in micro-calcifications composed of groups of small high contrast deposits, 2. Late indications are seen in growth of tissue masses in round areas of increased contrast (more water content from dense cell accumulations).
  • the structural diagnostic criteria fall into the two basic image information classes of image details, i.e., spacial details (smallest contrast variations as found in fibers and micro- calcifications) and intensity details (large area contrasts as found in tissue components and alterations). Problems in mammogram evaluation arise from excessive tissue density and concomitant superposition of contrasts.
  • a dense mammogram (FIGURE 21 A) was evaluated by digital PAIP image information enhancement.
  • the intensity detail image revealed the gland ducts and large round areas of decreased contrast indicative of fat depositions (less water content). Nearly all gland ducts can be followed leading to a point of origin (nipple). Some ducts were found dramatically increased in width (FIGURE 2 ID: circle).
  • PAIP image enhancement in accordance with the present invention revealed that all expected local contrasts were generated and collected but that detail information contrasts were compressed in inverse proportion to the extent of the background contrasts. In addition, the visual recognition of local, small intensity variations was reduced in close proximity to large intensity variations.
  • PAIP image enhancement provided an easy and fast procedure for adjustment of the proportion of detail and feature intensity ranges by interactive reduction of the latter while visually evaluating the processing effectiveness.
  • a chest X-ray image (FIGURE 22 A) occupied the full intensity range and left little room for the detail contrasts which had a significant intensity range of only 5%. They were only recognizable in part within the soft tissue background and were fully absent in the high and low intensity areas of the X-ray.
  • Detail enhancement with stepwise reduction of the significant intensity range lead to a reduction of the feature intensity range and to an inverse proportional increase of the detail contrasts.
  • the definition in digital images of any origin and content through PAIP of defined contrast classes and the quantitation of these classes provides a unique and new tool for image quality quantitation.
  • the four intensity information classes can be schematically represented in an information cube ( Figure 23) which depicts the relative significant intensity (z coordinate) over the image (x and y coordinates).
  • Each information class is presented by a simplified three-dimensional intensity profile and stacked upon each other in the sequence of access through PAIP processing; at the top is the image noise, followed by the spacial detail, the intensity detail, and at the base is the image feature/background. The latter is shaded if its spacial information is limited (light gray) or absent (background: dark gray).
  • info-cubes are presented of some of the microscopy data sets depicted in Figures 8-20.
  • the information classes are labelled with symbols along the left vertical axis, the total intensity range (IR) of the data set it indicated in the lower left corner, and the file name and the percentage range of each of the information classes is given in sequence to their appearance in the data. It is evident, that each image shows an individual proportion of its information components, but that common to all images is a reduction of visually perceivable "image quality" when the proportion of the feature component is more than 50% of the maximum relative significant intensity range.
  • the data are a ⁇ anged with decreasing image quality. Another application of the image quality assessment is found in image evaluation at the time of acquisition.
  • enhancement of image detail contrast is accomplished by adding a differential hysteresis pattern to a digital image.
  • the differential hysteresis image processing utilizes the persistence of intensity variations (i.e., hysteresis) as means for data reduction and image detail contrast enhancement.
  • Image hysteresis is determined by a hysteresis cursor of an interactively chosen hysteresis range. The cursor is stepped through the image pixel by pixel, evaluating intensity changes between neighboring pixels. The cursor's intensity position is maintained when the differential intensity values fall within the range, but follows the data, when the values are outside the range.
  • the processing proceeds bi-directionally on continuous lines at various angles and the averaged values of the cursor positions in each pixel generate a hysteresis image.
  • the hysteresis image contains hysteresis-free areas of maintained input data and continuously merged hysteresis areas of modified input data, resulting in, the input intensity variations being replaced by the local hysteresis value or the input intensity variations being reduced in intensity range proportionally to the local hysteresis characteristic, maximally by one half of the hysteresis range.
  • the image maintains its full integrity as well as the spacial position of all remaining contrast components.
  • the scaling of the resulting differential images to full display intensity range produces a differential hysteresis image having unique properties, i.e., they represent visually discrete differential intensity pattern of additive character.
  • a CCD charge coupled device
  • atomic force microscopy image which, e.g., provides data in the 14-bit range
  • CT image computerized tomography
  • radiogram which, e.g., provides data in the 12-bit range.
  • the present invention provides an interactive process for objective and exhaustive visual real-time access to any level of image resolution including the maximum sensor resolution using only one single parameter (i.e., the differential hysteresis range). Importantly, this process allows visualization of any image at the resolution level of the image sensor, rather than the eye, extending the "visual recognition level" to the acquisition level of the imaging device.
  • Present imaging sensors provide data in the 10-16 bit range (i.e., 1,024-65,536 levels) but the visual system (i.e., human eyes) only can "see” part of the data since it is limited in “perception” to approximately the 8-bit range (i.e., 256 intensity levels) and in pattern "recognition” to the 4-5 bit range (i.e., 16-32 intensity levels).
  • This limited information content of the visual image challenges digital image processing in finding general mechanisms for translating the non-perceivable or unrecognizable part of the sensor information into a recognizable image, i.e., patterns of image data details should be presented with a limited number of intensity steps spread over the full visual perception range.
  • high precision microscopes as found in the atomic force microscope (AFM), acquire digitally much more image data (e.g. in the 12-16 bit range) than the human eye can accept and data presentation becomes a limiting factor rather than spacial or contrast resolution.
  • image data e.g. in the 12-16 bit range
  • Similar problems are encountered in other sensor obtained images, e.g., medical imaging, satellite data or non-destructive testing. Accordingly, large data sets of many KBytes in dimension and 16-bit in depth have to be visually analyzed in a very short time and with demanding precision.
  • the image detail components may include small structural components of low contrast, i.e., the spacial details, as well as large components of higher contrast, i.e., the intensity detail. Both detail components are preferably preserved by image enhancement.
  • One approach to overcoming these limitations is the processing of local intensity components instead of local spacial components. Intensity variations can be characterized by their hysteresis properties which are accessible by intensity processing and provide an alternative to conventional spacial image processing.
  • Hysteresis processing has been used for linear spectral data, wherein intensity variations between neighboring pixels were compared by simple binary hysteresis evaluation using a moving "one-dimensional" cursor for sequential reading of pixels intensity values, i.e., either the intensity difference between consecutive reading falls within a given hysteresis range or not, see Ledley RS, Rotolo LS, Golab TJ, Jacobsen JD, Ginsberg MD and Wilson JB (1965), FID AC: Film input to digital automatic computer and associated syntax-directed pattern-recognition programming systems, Optical and Electro-Optical Information Processing (Tippett JT)
  • Hysteresis areas are defined by intensity variations which fall within a pre-selected hysteresis range (i.e., cursor range) and which are removed, and other areas of intensity variations which are outside of the given range and are maintained proportional to the local hysteresis response.
  • hysteresis character of the specific data merges both areas into a continuous hysteresis image.
  • Hysteresis images are visually not very effective because they maintain only the large intensity variations which can be readily recognized.
  • hysteresis images can be utilized for a visualization of the differences between hysteresis images including the original image data (which may be considered to represent a hysteresis image of unity hysteresis) and being characterized only by a single processing parameter, the differential hysteresis range.
  • the differential hysteresis image, scaled to full visual perception range includes both the spacial details and the intensity details and thus provides an efficient tool for image detail enhancement.
  • hysteresis line processing reads the image data in lines using a hysteresis cursor having a pre-selected range (i.e., hysteresis cursor range) and is moved pixel by pixel along the line while its midpoint is read as output data.
  • the cursor follows the input data with one of the cursor's ends if the next pixel value is outside of the present cursor value, or remains unchanged if the next pixel value is inside of the cursor values.
  • the cursor may start reading a line by positioning its midpoint at the input data point.
  • a directional lag of the hysteresis is compensated for by processing the input line a second time but in a backward direction and then averaging both cursor output values generating the final hysteresis line.
  • the hysteresis line has several important characteristics which are essential to hysteresis image processing: (1) intensity variations smaller than the hysteresis range are eliminated and replaced by a hysteresis value which represent the last read hysteresis-free data point, whereby the hysteresis value has a different character than the conventional averaging or median filter values since it relates the output data value to the preserved data portion rather than the replaced data portion; (2) the intensity of input data maxima and minima larger than the cursor range are uniformly reduced by one half of the cursor range independently of their specific intensity values or their height or depth, and intensity variations smaller than the cursor range are reduced in proportion to the local hysteresis character; (3) portions of the hysteres
  • Each successively read line has its own specific hysteresis character which reflects only the one-dimensional neighborhood along the reading direction.
  • the hysteresis character produces a streaky pattern in reading direction.
  • radial processing is prefe ⁇ ed.
  • a hysteresis cursor of selected intensity range (i.e., hysteresis cursor range) follows the input data with one of its ends if the intensity value of the next pixel falls outside of the present cursor endpoint values.
  • the midpoint of the cursor provides the output data (i.e., the solid output data line in FIGURE 24).
  • the cursor output value remains unchanged in the intensity of the next pixel falls within the actual cursor endpoint values (i.e., the hatched output line in FIGURE 24).
  • the hysteresis line contains segments which represent unchanged intensity variations of the input values (e.g., when read in both directions), reduced intensity variations (e.g., when read only in one direction) or eliminated variations (e.g., when read in neither direction).
  • the spacial position of maintained input intensities and the spacial position of reduced intensity peaks and valleys are not changed and provide for the high precision or "pixel-accuracy" of the hysteresis processing.
  • radial hysteresis lines such as described hereinbefore, are formed in many directions and arithmetically averaged in each pixel.
  • the processing requires between 100 - 200 or more different angles per half circle for the generation of the final hysteresis image.
  • hysteresis lines are generated in many directions and then averaged in each pixel.
  • the reading direction of the input lines is changed symmetrically within the half-circle and for each new direction a complete new data set is formed and averaged with the previous hysteresis processed data.
  • the resulting averaged data represents the hysteresis image.
  • a full frame viewing of the hysteresis image often requires only 4 - 8 different reading directions (i.e., a 45-22.5° offset angle between each reading directions) for the suppression of major hysteresis streaking which, if recognizable, would indicate a lack in local hysteresis co ⁇ elation (i.e., precision).
  • reading directions i.e., a 45-22.5° offset angle between each reading directions
  • streaks are observed when a large hysteresis cursor range (e.g., > -5% of the intensity range of the data) is used, high contrast image components are processed, or the image border is considered, wherein it is assumed that the hysteresis persists over a much larger area than that included in the averaging.
  • high precision hysteresis images maintain all the characteristics of hysteresis lines, i.e., "pixel-accuracy" of preserved intensity components which were larger than the hysteresis range, by maintaining their spacial position and their intensity character, but reducing their individual maximum intensity range by as much as one half the cursor range. Also, eliminated intensity components which were smaller than the hysteresis range are replaced by a base intensity level which was determined by the bordering preserved data, see FIGURE 26 designated "a".
  • Differential hysteresis processing recovers those intensity components which were lost after hysteresis image processing through subtraction of the hysteresis image from the input image. Since the original image (i.e., raw data) can be interpreted as a hysteresis image of a hysteresis cursor range equal to one (i.e., the top and bottom values, as well as the reading point, all equal 1), differential images can be described as being processed with a differential hysteresis range equal to the bottom values of both hysteresis ranges.
  • the differential image has a maximum intensity range equal to the differential hysteresis range and contains all the intensity components removed from the original image by the hysteresis image processing. This includes all small intensity components of the original image which were larger than the top value of the differential hysteresis range and a representation of all original large intensity components which were smaller than the bottom value of the differential hysteresis range. More importantly, the differential image can contain spacial details as well as intensity details if included in the differential hysteresis range, see FIGURE 26 designated "a". The limited intensity range of the differential image allows for linear scaling to the 8-bit range to produce the final differential hysteresis image.
  • the images do not need any further image processing since they have a balanced contrast range.
  • hysteresis processing reduction of the differential hysteresis range value leads to a reduction of the intensity range of the differential image and stronger contrast enhancement of its information content in the scaled differential hysteresis image, see FIGURE 26 designated "b".
  • the intensity details are excluded from the differential hysteresis image because they are maintained in the hysteresis image.
  • This pattern extraction provides a powerful new way for data reduction because it can separate from the original image, the differential hysteresis pattern of defined minimum and a maximum intensity variations. If the top value of the differential hysteresis range excludes the spacial details and the bottom value includes the intensity details, then the intensity detail can be extracted and the image scaled to the 8-bit range, as a discrete differential hysteresis image, see FIGURE 26 designated "c". Accordingly, it will be appreciated that differential hysteresis imaging allows extraction and display of any contrast level of a given minimum and maximum range as a differential hysteresis pattern.
  • the specific "pixel-accuracy" of the hysteresis images allows a recovery of those intensity variations which were removed from an image (the original or the hysteresis image) by hysteresis processing.
  • Subtraction of a hysteresis image from the original image produces a differential image of an intensity range equal to the hysteresis cursor range used for the generation of the hysteresis image. Due to its reduced intensity range the differential image can be linearly scaled to the full intensity range proportionally enhancing all contrasts of the details which may include spacial details as well as intensity details depending on the hysteresis range applied.
  • Image features of contrasts larger than the hysteresis range are maintained only at a contrast equal to the maximum cursor range thus they will be imaged with reduced contrast contribution in the differential image. Decreasing hysteresis cursor ranges will result in increasing contrast enhancement of intensity variations smaller than the cursor range. In this way smallest spacial hysteresis details may be extracted from the original image.
  • differential hysteresis images can be formed as well between two hysteresis images each processed with a different hysteresis range.
  • Original image components of contrasts larger than the smaller hysteresis range and smaller than the larger hysteresis range will be extracted and contrast enhanced in proportion to the differential hysteresis range.
  • certain hysteresis patterns can be extracted from the original image and displayed as discrete images, i.e., the intensity hysteresis detail component.
  • hysteresis patterns are a powerful tool for data analysis since they are of discrete and additive character.
  • objective visual evaluation of image enhancement is limited by the subjectivity of perception and recognition as well as the familiarity with the image content.
  • these limitations are taken into consideration by using a common visual pattern, e.g., a human face.
  • a portrait image was acquired with 1,024 x 1,024 x 8-bit resolution with a CCD camera and input and output intensities were approximately linear over the full range of 256 intensity steps. While the general image pattern is familiar, the image details are somewhat unfamiliar, but were objectively and quantitatively accessible and presentable.
  • the raw image represents a video portrait (1,024 x 1,024 pixels) of 8-bit intensity range and contained some saturated high lights (i.e., white areas).
  • FIGURE 27B shows the smallest intensity variations within a differential hysteresis range of 1 (i.e., the raw data image) and 9 intensity steps (indicated as DHR 1-9) represented a hysteresis noise pattern.
  • Most images exhibit a hysteresis noise component as a structurally random pattern which may include some non-random components (e.g., in case of non-linear noise).
  • FIGURE 27C shows that successive differential hysteresis analysis with increased hysteresis values reveal a spacial hysteresis detail pattern.
  • This component represented the smallest non-random contrasts which produced a visually useful coherent image.
  • this component included the high precision contrasts of an imaging system at the level of its contrast resolution.
  • This pattern often extended only over short distances representing the smallest spacial image components. In this particular image they were found in small variations of light absorption and reflection on the face and the clothing.
  • FIGURE 27D shows that below the spacial hysteresis detail pattern resided a pattern of larger contrasts, the intensity hysteresis detail pattern. Often, this pattern represented contrasts which extended over larger areas then the spacial details.
  • FIGURE 27E reveals that the largest contrasts often represented the major structural features of an image.
  • This hysteresis image feature pattern was bare of hysteresis detail data. It contained here the major compositional portrait components.
  • FIGURE 27F shows all the hysteresis patterns summed at the proportion at which they were found in the raw data and the resulting composite hysteresis pattern produced an image indistinguishable from the original image (FIGURE 27 A).
  • the additive character of the discrete differential hysteresis patterns provided the basis for an objective visual display of all available data information as enhanced images.
  • FIGURE 28A which is the same as FIGURE 27A, shows that the original image which does not reveal many of the image details that the CCD camera captured due to its high sensitivity and resolution, but that are at or below visual recognition limits due to the contrast range in the final image.
  • FIGURE 28B shows that all of the differential hysteresis details present in the image data (DHR 9-65, differential hysteresis range of 9 and 65 intensity steps) were visually enhanced by adding to the original image (FIGURE 28A) a portion of the hysteresis detail pattern. However, the smallest high precision contrasts still were not readily visible. A further amplification was required by adding selected contrast-enhanced detail patterns to the original image.
  • FIGURE 28C shows that selected intensity hysteresis details (contrast enhanced by a reduced DHR 35 - 37, differential hysteresis range of 35 and 37 intensity steps), were interactively added to the original image (FIGURE 28A) in such proportion (i.e., 40%) that they were clearly recognizable.
  • the enrichment of the original image with high contrast details clearly indicated their relationship to the whole image without interference from the low- contrast details.
  • FIGURE 28D shows that minute surface contrasts were visualized when selected spacial hysteresis details (contrast enhanced by a na ⁇ ow DHR of 15-17, differential hysteresis range of 15 and 17 intensity steps) were enriched in the original image (FIGURE 28 A).
  • This representation of the 8-bit image data visualized the high precision contrasts while maintaining the integrity of the whole image by including all other contrast components of larger contrast ranges.
  • the present invention makes it possible to "see" the image data at the contrast resolution level of the digital camera.
  • An atomic force microscope is a high precision microscope having a contrast resolution 250-times higher than the human eye (i.e., 16-bit versus 8-bit).
  • differential hysteresis imaging provides visual access to the instruments precision imaging capabilities by enriching the precision contrasts patterns in scaled raw data images.
  • FIGURE 29A shows scaled raw data (intensity range (IR)
  • FIGURE 29C shows that micro and macro topography were displayed together by mixing the 8-bit scaled images of each component (80% DHR 1-9,603 + 20% DHR 1-65).
  • FIGURE 29D shows that the highest contrasts resolution was found in a differential hysteresis pattern that represented height information of 0.3 ⁇ (DHR 1-3). Although limited by reduced spacial resolution, at low magnification, the microscopic data revealed at full contrast resolution an expected wealth of topographic details at an astonishing low level of noise.
  • FIGURE 29E shows that the maximum precision imaging capability can be visualized at the 14-bit level by hysteresis detail pattern enrichment (15% DHR 1-3 + 85% DHR 1-9,603) providing both the contrast of the major tilt and the contrast of the minor surface roughness.
  • FIGURE 29F shows that even the closer visual inspection of the extracted hysteresis detail image at the 6-bit level required an enhancement of the precision information (40% DHR 1-3 + 60% DHR 1-65) due to the limited visual perception. Thus, a step wise increase of contrast pattern enhancement was required for the visual presentation and recognition of all the image information in high precision micrographs.
  • FIGURES 30A - F the conventional display of computerized tomography (CT) is shown to discriminate fine structural information and windowing (i.e., selected intensity range of imaged pixels) is used for the extraction of tissue- specific sensor data.
  • CT computerized tomography
  • This reduction data was based on the absorption coefficient for the bulk materials, but ignored that thickness contrasts of the same materials may be present but will be eliminated by the windowing. All available sensor contrasts can be made visible through differential hysteresis imaging in the non- windowed raw data and visual pattern recognition can be enhanced by enrichment of the precision differential hysteresis pattern in the data.
  • FIGURE 30A shows that a CT cross section (IR 1-2048) scaled to 8-bit image revealed little data.
  • FIGURE 30B shows that differential hysteresis image for a large differential hysteresis range of 8-bit improved the overall structural contrasts.
  • FIGURE 30C shows that the enhanced hysteresis intensity details (DHR 27-35) reveals details in all tissues, i.e., the liver region showed a dramatic increase of fine structures.
  • FIGURE 30D shows an improved image with enriched selected intensity details (40% DHR 27 - 35 + 60% DHR 1-256).
  • FIGURE 30E shows that the highest contrast resolution hysteresis detail patterns (DHR 21-23) were found below the relatively small noise component (1% IR). Although noisy, the liver fine structure indicated a zoning and ultra structures at the level of single canaliculae and vessels.
  • FIGURE 3 OF shows that an improved image was obtained by enrichment of the spacial hysteresis details.
  • the differential hysteresis enhancement revealed that the signal-to-noise ration of the spacial details was insufficient for utilizing the surprisingly high spacial resolution of the CT data acquisition system.
  • the extension of image resolution by differential hysteresis imaging provides a powerful tool in optimizing imaging devices.
  • FIGURES 31 A - D the important contrasts of a image data set can be found interactively with ease under computer mouse control by applying the differential hysteresis filters and the pattern enrichment in a "real-time window".
  • IR 1-1,024
  • FIGURE 3 IB shows that in a second step, within the adjusted image the area of interest was defined and diagnostically important contrasts were determined and enhanced.
  • soft was well as hard tissue were analyzed using a na ⁇ ow differential hysteresis range which excluded the hysteresis noise component but including the spacial precision information (DHR 9-15).
  • FIGURE 31C shows that the diagnostically important contrasts were enriched in the adjusted data at an appropriate percentage using the real ⁇ time window (25% DHR 9-15 and 75% DHR 1-64).
  • FIGURE 3 ID shows that the sensor data could be visualized combined in an enhanced image, extending the contrast resolution from that of the eye to that of the sensor. High contrast resolution was found only within 0.5% of the IR which indicated a very low contrast resolution of the imaging plate compared to that of the digitized film radiogram.
  • the differential hysteresis range of the spacial differential hysteresis pattern can be used as a quantitative tool for optimizing the acquisition parameters.
  • arll output base address
  • arl2 input base address
  • arl3 ⁇ rows
  • arl4 columns/PE (depth)

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Image Processing (AREA)

Abstract

Le lissage de deux images selon des coefficients de lissage différents permet d'obtenir par soustraction les différences d'imagerie entre les données les moins lissées et les plus lissées. Selon cette invention, la technique de lissage inclut un filtre numérique d'hystérésis bidimensionnelle utilisant un 'masque bidimensionnel de variables à réinitialisation automatique'. Le filtre est donc indépendant de la taille d'image et de son contenu, et il ne peut pas modifier le dimensionnement de caractéristiques d'intensité significatives. Pour calculer la valeur de l'intensité du pixel lissé, le procédé de lissage par hystérésis bidimensionnelle utilise un ensemble de lignes d'hystérésis unidimensionnelles de pentes différentes, et coupant chaque pixel d'une image. Pour obtenir la valeur résultante concernant le pixel, le procédé totalise ces valeurs d'hystérésis unidimensionnelle et à les divise par le nombre de lignes d'hystérésis. Cette technique de traitement de l'intensité permet la séparation des informations contenues dans une image en groupes d'informations de base.
EP95919731A 1994-03-08 1995-03-08 Procede de traitement numerique de l'intensite pixel par pixel ameliorant la lisibilite des informations d'imagerie Withdrawn EP0702819A4 (fr)

Applications Claiming Priority (5)

Application Number Priority Date Filing Date Title
US08/207,488 US5563962A (en) 1994-03-08 1994-03-08 Two dimensional digital hysteresis filter for smoothing digital images
US08/207,489 US5592571A (en) 1994-03-08 1994-03-08 Digital pixel-accurate intensity processing method for image information enhancement
US207488 1994-03-08
US207489 1994-03-08
PCT/US1995/002962 WO1995024694A1 (fr) 1994-03-08 1995-03-08 Procede de traitement numerique de l'intensite pixel par pixel ameliorant la lisibilite des informations d'imagerie

Publications (2)

Publication Number Publication Date
EP0702819A1 EP0702819A1 (fr) 1996-03-27
EP0702819A4 true EP0702819A4 (fr) 1996-08-14

Family

ID=26902283

Family Applications (1)

Application Number Title Priority Date Filing Date
EP95919731A Withdrawn EP0702819A4 (fr) 1994-03-08 1995-03-08 Procede de traitement numerique de l'intensite pixel par pixel ameliorant la lisibilite des informations d'imagerie

Country Status (4)

Country Link
EP (1) EP0702819A4 (fr)
JP (1) JP3663465B2 (fr)
AU (1) AU2543095A (fr)
WO (1) WO1995024694A1 (fr)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
FR2796740B1 (fr) * 1999-07-19 2001-10-26 Ge Medical Syst Sa Procede et systeme de gestion de la saturation sur une image radiographique numerisee
JP4405716B2 (ja) * 2002-08-26 2010-01-27 ジーイー・メディカル・システムズ・グローバル・テクノロジー・カンパニー・エルエルシー 画像処理装置およびｘ線ｃｔシステム
CN102750673A (zh) * 2011-11-30 2012-10-24 新奥特（北京）视频技术有限公司 一种提高图像对比度的方法
WO2015105980A1 (fr) * 2014-01-09 2015-07-16 Zygo Corporation Mesurer la topographie de surfaces asphériques et autres surfaces non plates
US10147171B2 (en) * 2016-09-21 2018-12-04 General Electric Company Systems and methods for generating subtracted images
CN115035256B (zh) * 2022-05-06 2023-01-06 中国安全生产科学研究院 一种尾矿库事故隐患及风险演化方法及系统

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4907156A (en) * 1987-06-30 1990-03-06 University Of Chicago Method and system for enhancement and detection of abnormal anatomic regions in a digital image
US5151794A (en) * 1989-01-30 1992-09-29 Ezel Inc. Image processing method
US5063448A (en) * 1989-07-31 1991-11-05 Imageware Research And Development Inc. Apparatus and method for transforming a digitized signal of an image

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
EHRICH R W: "A Symmetric Hysteresis Smoothing Algorithm that Preserves Principal Features", COMPUTER GRAPHICS AND IMAGE PROCESSING, vol. 8, no. 1, August 1978 (1978-08-01), USA, pages 121 - 126, XP000573503 *

Also Published As

Publication number Publication date
WO1995024694A1 (fr) 1995-09-14
JPH09500472A (ja) 1997-01-14
EP0702819A1 (fr) 1996-03-27
JP3663465B2 (ja) 2005-06-22
AU2543095A (en) 1995-09-25

Similar Documents

Publication Publication Date Title
US5715334A (en) Digital pixel-accurate intensity processing method for image information enhancement
US5592571A (en) Digital pixel-accurate intensity processing method for image information enhancement
US5563962A (en) Two dimensional digital hysteresis filter for smoothing digital images
Kumar et al. Application of digital image magnification for surface roughness evaluation using machine vision
Johnson et al. X-ray diffraction contrast tomography: a novel technique for three-dimensional grain mapping of polycrystals. II. The combined case
US6891549B2 (en) System and method for edge enhancement of images
Meyer Mathematical morphology: from two dimensions to three dimensions
Starck et al. Weak lensing mass reconstruction using wavelets
CA1315902C (fr) Minimisation de points diriges generes par la methode des cubes diviseurs tridimensionnels
Grande Principles of image analysis
Clarke et al. Tree-structured non-linear filter and wavelet transform for microcalcification segmentation in digital mammography
EP1071042A2 (fr) Méthode pour déterminer les composants de forme de bruit de l'image d'un dispositif d'images et usage de cette méthode dans un dispositif d'images
EP0702819A1 (fr) Procede de traitement numerique de l'intensite pixel par pixel ameliorant la lisibilite des informations d'imagerie
CA2162276C (fr) Methode de traitement numerique de l'intensite au niveau des pixels pour ameliorer la qualite des images
Conan et al. Geostatistical and morphological methods applied to three‐dimensional microscopy
Russ Computer-aided quantitative microscopy
Tovey et al. Practical aspects of automatic orientation analysis of micrographs
Gremillet et al. Dedicated image analysis techniques for three-dimensional reconstruction from serial sections in electron microscopy
Corrias et al. Total Variation-Based Image Decomposition and Denoising for Microscopy Images
Bons et al. Image analysis of microstructures in natural and experimental samples
Douglas et al. An introduction to image processing in medical microscopy
Kriete et al. Computerized three-dimensional reconstructions of serial sections in electron microscopy
Roldan Reconstruction of porous structures from FIB-SEM data
Biggs et al. Subpixel deconvolution of 3D optical microscope imagery
Jiménez Reconstruction of porous structures from FIB-SEM data

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 19951207

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AT BE CH DE ES FR GB IE IT LI NL

A4 Supplementary search report drawn up and despatched
AK Designated contracting states

Kind code of ref document: A4

Designated state(s): AT BE CH DE ES FR GB IE IT LI NL

17Q First examination report despatched

Effective date: 20071026

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN

18D Application deemed to be withdrawn

Effective date: 20110802