JPH03224075A - Aligning method in image processing - Google Patents

Abstract

PURPOSE:To accelerate a required processing processing time and to improve accuracy in binary image processing by dividing the image of a model and a digital image that it an object to be processed into plural areas, and using feature quantities in the areas when they are compared with each other. CONSTITUTION:An image pickup part 1 photographs the object 2 to be processed, and a signal conversion part 6 converts an analog image signal from the image pickup part 1 to a digital image signal, and a feature quantity extraction part 9 extracts the feature quantity of digital image data from the signal conversion part 6. The video signal of an image that becomes the model and that of the image that becomes the object to be processed are digitized, and digitized images are divided into a prescribed number of areas, and the alignment of the model and the digital image that is the object to be processed are performed setting the feature quantity of the area as reference, and after that, the inspection and the identification of the object to be processed are performed. In such a way, it is possible to accurately perform the alignment and to perform the inspection of the object and to accelerate the stroke of the identification.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical field to which the invention is directed] The present invention relates to an alignment method in image processing, and particularly relates to alignment between a model image and a processing target.

[Prior art]

In conventional image processing, the first function is image alignment,
In other words, in many cases, the function of positioning an image serving as a model and a target image to be compared with the model image does not have a positioning function within the image processing apparatus itself, but relies on external material handling technology. One example is setting up a window for positioning that corresponds to a very characteristic part in the image, searching that window in the target image, finding the characteristic part, and adjusting the position in the image by offset from that window. There is a method of searching, finding a characteristic part, and using an offset from that window to find the inspection target part in the image.

However, this method requires the existence of a characteristic part to be searched, and the alignment accuracy is not high. It is also often used in binary image processing.

Further, although the correlation method and the 5SDA method are well-known as alignment methods, these methods require a large amount of calculation time and are not often applied to general-purpose processing devices.

On the other hand, regarding inspection and identification of target images, the most common conventional image processing methods are the in-window pixel counting method, the so-called multi-window method, the SRI algorithm method,
All three ten-blade matching methods use binary image processing, so the main premise is that a good binary image can be obtained. It is not suitable for testing.

In addition, other grayscale image processing devices and color image processing devices handle a large amount of information, so although they are highly functional, they require dedicated processors and hardware to process large amounts of data at high speed, making them extremely expensive. be.

[Object of the Invention] In view of such conventional problems, it is an object of the present invention to improve the time required and accuracy in binary image processing.

[Summary of the invention]

In order to achieve the object, this invention adopts grayscale image processing in place of conventional binary image processing, digitizes the video signals of the model image and the image to be processed, and converts this digitized image into a predetermined image. The digital image is divided into two regions, the model and the digital image to be processed are aligned based on the characteristic quantities of each region, and the object to be processed is then inspected and identified.This enables high-speed processing. be.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings.

That is, in FIG. 1, the imaging unit, for example a camera, is for photographing the processing object 2. In FIG. The processing target 2 is, for example, the printing of a stamp 4 such as the address of a manufacturing factory stamped on the body of a paper container for dairy products, as shown in FIG. The imaging device 1 is then connected to an image processing device 5. In this processing device, the signal conversion unit 6 has the function of converting the analog image signal from the imaging unit 1 into a digital image signal, and also performs preprocessing such as differentiation on the input signal, and converts the digital signal back into an analog signal. It has a function to output to the monitor 7 instead of

Further, the feature amount extracting section 9 has a function of extracting the feature amount of the digital image data from the signal converting section 6.

The frame memory 10 stores the data extracted by the feature extracting section 9. Each of these parts is connected to the calculation control section 13 and controlled by this calculation control section. The input/output section 14 has a function of receiving a trigger signal from the outside and outputting a determination result by the calculation control section 13 to the outside.

Further, a memory 15 necessary for this processing work is connected to the arithmetic control section 13.

Next, a method for processing an image to be processed will be described with reference to FIG.

First, an image of a model, that is, a non-defective sample is photographed by the imaging section 1, and this image data, that is, an analog signal, is sent to the signal converting section 6 (step 1).

The image of the model input at this time, that is, the size of the area of the stamping part 4 used as the model, is the pixel size of nXm when the total image size is NXM, as shown in FIG. is N>n M>m. Then, the image data of the model is converted into a digital signal in the signal conversion section 6. In this way, the image of the model is input.

Next, the digital image of this input model, i.e. n
The image of size xm is divided into small regions of a predetermined size, that is, rectangular small head regions 21 of aXb pixels as shown in FIG. 5, and the feature amount for each small region is calculated (step 2). .

Note that in this invention, the feature amount is the characteristic value of each region obtained by dividing a digital image of a predetermined size into multiple regions with a predetermined size and accumulating the density of multiple pixels within each region. In this case, the cumulative value of the differential concentration within each small region is used. This differentiation uses the operators for each pixel 20 and 20 shown in FIG. 6 and the following equation.

f'(x,y)=l f(x,y)-f(x,y-
1) +1f(x, i-f(x-Ly)) Then, the cumulative value is calculated by digital differentiation in the differential value calculation circuit shown in FIG. The output of the IH delay circuit 25 is input to the difference absolute value circuit 27, and the output of the IPix delay circuit 26 is input to the difference absolute value circuit 27 and 28 as well as to the circuit 26. Furthermore, the outputs of the difference absolute value circuits 27 and 28 are added by an adder 29 and output as a digital differential image.

Note that a detailed explanation of this differential value calculation circuit will be omitted.

The cumulative value of the digital differential image data is calculated by the feature extracting circuit shown in FIG. This feature quantity extraction circuit has a plurality of line memories Lml to Lm5, which include a write tarlock generator 31, a write enable control circuit 32, a read clock generator 33, a read enable control circuit 34, and a pixel product-sum circuit. 35 are connected.

The digital differential image data has a screen having a predetermined pixel size, for example, 256×240 pixels. This screen is then divided into small areas of a predetermined number of pixels, for example, 5×5 pixels. The digital image has a size of 240
The scanning line is composed of the scanning lines of a book, and these scanning lines are written into the line memories Lml to Lm5 in the scanning order from the beginning of each line memory. Subsequently, the scanning lines from the 6th grain to the 10th grain are sequentially written into the latter half of each line memory Lml to Lm5. Next, the scanning lines from the 11th main port to the 15th main port are written in the first half of each line memory Lml to Lm5 again. In this way, writing is performed cyclically in the first half and second half of the five line memories in units of scanning lines (
(Figure 9a).

Next, reading starts five scanning lines later than writing, and successive clocks are used for line memories Lml, Lm2, Lm3, . L
m4. Lm5. Lml Lm2 Lm3, . . . five line memories are scanned pixel by pixel. In this way, the pixel data whose order has changed from that at the time of input (Fig. 9) is input to the 25-pixel cumulative sum circuit 35, and the feature amount for each small area is extracted. Finally, the small area feature amount is calculated in real time with a delay of 5 scanning lines relative to the input pixel data.

Next, as indicated by the X marks in FIG. 10, feature quantities between local cells in the 9th or 25th neighborhood of the digital image of the model are calculated (step 3).

That is, as shown in FIG. 10, a middle region 22 is formed by integrating a plurality of small head regions 21, 21, ie, A horizontally and '@iB. In this figure, there are three small areas 21 horizontally and three vertically.
．． 21 forms one middle region. Then, the feature amount within this middle region is calculated. Furthermore, at this time, data calculation in the nearby medium area is also performed for second positioning, that is, accurate positioning. The neighboring medium area data is the first
As shown in FIG. 1, this data is based on a medium area 22 formed by integrating new small areas shifted in pixel units within a small area 21 that is the source of the medium area. Note that these calculations are based on the first
This is performed by the calculation control section 13 shown in the figure.

Next, a digital image to be inspected or identified, ie, to be processed, is input. That is, a digital image to be processed is input at the full image size of NXM shown in FIG.

Next, the feature amount within the small head area 21 of the digital image to be processed is extracted. That is, in the same way as the feature amount within the small area in the digital image of the model, the feature amount within the small area in the target digital image is calculated using the same small area size. This calculation is performed by the hardware shown in FIGS. 1, 7, and 8.

Next, a first alignment is performed between the model and the digital image to be processed. That is, this alignment is a rough alignment using small areas of both images (step 6).

In general, in image processing, it is difficult to capture images of moving objects at the same timing and position using a trigger input, so software is used to align the parts of the entire image that are of interest, i.e., where to inspect or identify them. It is often necessary to do so. Therefore, in this embodiment, as shown in FIG.
It is done by law. Since this 5SDA method is normally performed pixel by pixel using the density value of the pixel, it takes an enormous amount of time to calculate. One way to shorten the calculation time is to skip pixels using coarse-fine (coarse-f 1ne).
There are some methods, but they have problems with accuracy. In this invention, the sum of feature amounts, ie, differential values, within the capitula 21 is used.

Next, a small area 22 of the target digital image is calculated. That is, after the first alignment, the middle region 22 is calculated, that is, the small head region 21 is integrated using the same middle region size for the model and the digital image to be processed. Note that, unlike in the case of the model, calculation of the neighborhood medium area is not performed here (step 7).

Next, a second alignment is performed using the model and the small area 22 of the digital image to be processed. That is, more accurate positioning is performed using feature amounts in the 9 or 25 local cells of the previously calculated digital image of the model. This second alignment is performed by comparing the mid-region correspondence between the model and the target using 9 sets of mid-region data for 9 neighborhoods, or 25 sets of mid-region data for 25 neighbors, as indicated by the X marks in Figure 10, and finds the one that is most similar to the target. The position of the neighboring model is set as the accurate position. Specifically, for example, the ratio of the absolute value of the difference between the data to be processed and the model data in the corresponding middle region to the model value is calculated, and the cumulative calculation is performed for all the middle regions. A method can be considered in which the position with the smallest cumulative value among the 9 pairs or 25 m neighborhood is determined as accurate alignment (step 8).
).

Next, an inspection is performed by comparing the middle region of the model and the target. That is, after the second alignment, feature amounts are compared between the model and the processing target in the corresponding region. Naturally, the model uses the data of the closest neighbors in the second alignment. The determination level can be adjusted by setting upper and lower thresholds for comparison. Further, the determination threshold value can be set to a different value for each medium region (step 9).

Although we have described the case where the cumulative value of the differential density value is used as the feature quantity in the small region 21 and the medium region 22, in addition to this, the cumulative value of the density value of the model and the target original image and the cumulative value of the differential value by direction are also used. , the number of pixels having original image density or differential density within a certain range, or the feature amount of the original image density histogram, such as the number of peaks or the maximum slope.

〔Effect of the invention〕

As described above, this invention divides the images into a plurality of regions to compare the model and the digital image to be processed, and uses the feature amounts in both regions, so that the alignment of the two is accurate and The process of inspecting and identifying objects can be performed at high speed. In addition, image processing using feature quantities is suitable for surface inspections where the contrast is not very strong; In inspection, defective elements can be easily detected because the differential values of crushed stamps and missing stamp parts are similarly small.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing one embodiment of the image processing device according to the present invention, Fig. 2 is a flowchart of image processing, Fig. 3 is an external perspective view of an object to be inspected or identified, and Fig. 4 is a processing The fifth is a front view showing the target stamping area, the fifth is a plan view of the stamping area in Figure 4 divided into small areas, and the sixth is a 1
A diagram showing the coordinates of pixels 20 and 20 involved in pixel differential processing, Figure 7 is a block diagram of the differential value calculation circuit, Figure 8 is a block diagram of the feature quantity extraction circuit in a small area, and Figure 9 is the line memory Fig. 10 is a relationship diagram showing the positional relationship between an arbitrary pixel of interest and a small area when calculating a 9-neighbor medium area, and Fig. 11 is an explanatory diagram showing the order of input and output. An explanatory diagram showing the relationship, FIG. 12 is an explanatory diagram of the 5SDA method using a small area. 1... Imaging unit 2... Processing target 4... Stamping unit 5... Image processing device 6... Signal conversion unit 7... Monitor 9... Feature extraction unit 10... Frame Memory 13... Arithmetic control section 14... Input/output section 15... Memory 20... Pixel 21... Small area 25... 1 horizontal line delay circuit 26... 1 pixel delay circuit 27. ...Difference absolute value circuit 28...Difference absolute value circuit 29...Adders Lml to Lm5...Line memory 31...Clock generator 32...Write enable control circuit 33...Read clock generation Vessel 349.・Read enable control circuit 35...pixel product-sum circuit

Claims (12)

[Claims] (1) Input a model image to an imaging unit, convert the output from this imaging unit from analog to digital to obtain a digital image of the model, and integrate this digital image with a predetermined number of pixels. dividing the image into regions and extracting feature amounts within each region; inputting the image to be processed into the imaging device; converting the output from analog to digital to obtain a digital image to be processed;
Divide this digital image into multiple regions integrated with a predetermined number of pixels, extract the feature amount within each region,
An image processing method characterized by aligning the digital image of the model and the digital image to be processed by comparing the digital image of the model and the digital image to be processed in the area of both digital images. . (2) The alignment method in image processing according to claim 1, wherein the feature amounts in the digital image of the model and the digital image to be processed are based on the density of pixels in each region. (3) The alignment method in image processing according to claim 1, wherein the feature amounts in the digital image of the model and the digital image to be processed are based on the sum of density of pixels in each region. (4) The alignment method in image processing according to claim 1, wherein the feature amounts in the digital image of the model and the digital image to be processed are based on the differential value of density of pixels in each region. (5) Registration in image processing according to claim 1, wherein the feature amounts in the digital image of the model and the digital image to be processed are based on the sum of differential values of the density of pixels in each region. Method. (6) The feature amount in the digital image of the model and the digital image to be processed is the cumulative value of differential values in each reading direction of pixels in each region. alignment method in image processing. (7) The feature amount in the digital image of the model and the digital image to be processed is based on the cumulative density value of the model and the original image to be processed. Alignment method in image processing described. (8) The feature amounts in the digital image of the model and the digital image to be processed are based on the number of pixels having a predetermined density within a predetermined area in the original image of the model and the original image to be processed. An alignment method in image processing according to claim 1. (9) The feature amounts in the digital image of the model and the digital image to be processed are based on the number of pixels having a predetermined differential density within a predetermined area in the original image of the model and the original image to be processed. An alignment method in image processing according to claim 1. (10) The image according to claim 1, wherein the feature amounts in the digital image of the model and the digital image to be processed are based on density histograms of the model and the original image to be processed. Alignment method in processing. (11) The alignment of the digital image of the model and the digital image to be processed is performed by comparing the digital image of the model within the digital image to be processed. Alignment method in image processing. (12) The alignment method in image processing according to claim 1, wherein the area of the digital image of the model is smaller than the area of the digital image to be processed.