JPH02205985A - High speed picture processor - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a high-speed image processing device suitable for recognizing noisy and unclear images such as cracks on a road surface, etc., and particularly for a general-purpose computer. The present invention relates to a high-speed image processing device that achieves unprecedented high-speed processing.

[Conventional technology]

There are three types of road surface properties that can be used to evaluate road damage: ruts, longitudinal unevenness, and cracks. Various proposals have been made to automate these measurements, which were conventionally done by visual inspection.

As this type of automation technology, the present inventors have proposed
-127490 and Japanese Patent Application No. 62-62242. These techniques are used to recognize cracks, which are difficult to recognize among the above three elements, and image data is collected by a measurement vehicle that can travel on the road. Then, cracked image data is extracted from the collected image data, and the steps are roughly divided into two steps.

One step is "segment extraction," which is a step that detects crack candidate points from the measurement image and outputs compressed and simple segment information with items such as width, length, and direction. For example, the measurement image is 3
After being divided into square small areas of 2×32 pixels, the image is processed.

The other step is "continuity determination," which is a step to determine continuity between extracted segments and recognize cracks in the final line drawing.

By the way, in these image processes, "segment extraction" is local processing performed for each of the small square regions in the whole image, as shown in FIG. As shown in (b), this is global processing performed using all image data.

[Problem that the invention seeks to solve]

In this type of image processing, "local processing"
In many cases, it is necessary to perform a combination of processing and "global processing," but it is difficult to perform both of these two types of processing at high speed in a general-purpose computer system.
A computer system that can perform efficient processing has been desired.

[Means for solving problems]

Therefore, in the present invention, a high-speed image processing device divides an original image into a plurality of regions, and performs local image processing for each divided region and global image processing in which local image processing results for each divided region are correlated with each other. A first image processing device having a plurality of microprocessors coupled by a bus coupling method and performing the local image processing with the plurality of microprocessors; and a plurality of microprocessors coupled by a network coupling method. , and a second image processing device that executes the global image processing using the plurality of microprocessors, and the first and second image processing devices are operated in parallel.

[Effect]

That is, the present invention focuses on multi-microprocessor systems that are used to speed up computers. Multi-microprocessor systems can be roughly divided into bus-coupling systems as shown in Figure 10(a) and bus-coupling systems as shown in Figure 10(a).
There is a network connection method as shown in Figure 0 (b).
In this invention, a bus-coupled multi-microprocessor is used for local image processing that requires a high data transfer frequency and a large amount of data transfer, and a bus-coupled multi-microprocessor is used for global image processing that requires data from adjacent areas but has a low data transfer amount. Use a network-coupled multi-microprocessor.

〔Example〕

FIG. 1 shows an embodiment of the present invention.

In FIG. 1, a data recorder 1 records road surface image data measured by a road surface measuring vehicle using a laser method. The crack data obtained using the laser method is expressed as a gradation image from white to black due to the shadow effect, and the cracks are shown in black (
rM style O), h, road surfaces without rips are represented as gray, and white center lines, etc. are represented as white NL1t1255).

The road image data reproduced by the data recorder 1 is stored in the original image memory 2.

The original image memory 2 has a capacity of, for example, 128 MB (for example, 4 m across and 16 m in length on a road), and by transferring this stored data to the display memory 3, the image data is displayed on the monitor 4.

In accordance with the segment extraction request from the recognition processor 20, the measurement processor 10 performs a process of determining whether or not crack data exists in each divided area of the original image memory 2 (hereinafter referred to as primary determination process), a starting point search process, and Perform tracking processing, etc.

This measurement processor 10 has an address generation section 11 and a primary determination section 12, and in this case, the measurement processor 10 employs the technique disclosed in Japanese Patent Application No. 62-280776. This process will be briefly explained below.

First, the measurement processor 10 (for example 512 in FIG. 2)
) is determined, and the average value D. is used to determine the threshold value δ for determining the presence or absence of a crack, which will be used in the next starting point search process.

Next, the measurement processor 10 applies 16×16 square slits to the image data for one frame transferred from the original image memory 2, for example, as shown in FIG. Divide into regions. In this case, one slit area includes image data of 32×32 pixels.

The measurement processor 10 first performs a "start point search" for cracks on the thus divided image data. This starting point search is to search for the starting point for the next tracking process (downward movement), and in this starting point searching process, instead of checking for cracks in all slit areas, The presence or absence of cracks is determined only in the slit areas where the arrows ■, ■, ■, ■ intersect, in the direction of the arrows and in the order of ■, ■, ■, ■.

Determination of cracks is carried out as follows. That is, a histogram consisting of the relationship between gradation level and number of pixels as shown in Fig. 3 is obtained for each area under the above-mentioned arrow, and based on this histogram, the above-mentioned procedure is performed. The number N of pixels (corresponding to the hatched area in FIG. 3) existing below the threshold value δ is determined for each slit area.

Then, the measurement processor 10 calculates the above N and the set number Na.
The primary house settlement result Id is classified as follows based on the comparison of
Output.

■ There is a high possibility that a crack exists. ■ It is unclear whether there is a crack. ■ The possibility that a crack exists is low. Such determination results Id can be obtained by appropriately changing the set number Na. Then, the measurement processor 10 outputs this primary determination result Id to the sequential extraction processor 30.

In other words, in this starting point search process, the arrows ■, ■
, ■, ■ By repeating the process of determining the presence or absence of a starting point, the tracking starting point within the frame is selected.

Then, the measurement processor 10 assigns serial numbers to the slit regions selected through such a starting point search process, and executes the following "tracking process" in the order of these numbers.

First, regarding the starting point of NO.1, the next tracking direction is determined based on the position of this starting point. That is, when the tracking starting point is at the position shown in FIG. 2(a), the adjacent right three slit areas are determined to be the next tracking area, as shown in FIG. 4(a). Hereinafter, similarly, the tracking starting point is shown in (b) and (
c), (d).

(e), (f), (g), (h), the next tracking direction is shown in Fig. 4 (b), (C), (d).

They are determined as shown in (e), (f), (g), and (h), respectively. In this way, in tracking when the tracking starting point is used as the tracking source, the tracking direction is specified only by the position of the tracking starting point within the frame.

Next, the measurement processor 10 determines whether or not a crack actually exists in the tracking slit region thus determined, using the histogram-based determination method shown in FIG. 3, similar to the origin search process. And this judgment result I
d to the extraction processor 30.

In addition, the measurement processor 10 performs processing to determine the next tracking direction as described below. However, in this case, the tracking direction is determined based on the positional relationship between the previous slit that was determined to have a crack and the slit from the time before the previous one. . That is,
As shown in Figures 5(a) and (b), when tracking is performed right sideways (when two slits determined to have cracks are adjacent on the left and right), the three slits on the right or left If the tracking is performed directly below or directly above as shown in Figures 5(c) and (d) (when two slits are adjacent to each other above and below), the 3 slits below or above Perform the following tracking on the slit, and if the tracking is performed diagonally as shown in Figures 5(e) to (h) (when two slits are diagonally adjacent), the slit in the diagonal direction is the center. The following tracking is performed for the three slits shown.

In the method of this embodiment, the tracking direction is determined based on the positional relationship between the current and previous slits that were determined to have a crack, so the tracking direction is determined at the time when the position of the slit with a crack is determined. can be determined. Therefore, it is no longer necessary to perform this tracing direction determination process after the line segment extraction process performed by the extraction processor 30, and these tracing direction determination processes and line segment extraction processes can be executed in parallel after determining the presence or absence of cracks. Data processing efficiency and speed can be increased.

The measurement processor 10 executes the tracking process starting from the starting point of N011 by repeatedly executing the tracking process as described above until the tracking area reaches the end area of one frame or there are no cracks in the tracking direction. Execute. When this tracking is finished, the next N082
The same tracking process as described above is executed for the starting point. This process is then repeated until the tracking process for all starting points is completed.

The above is the primary determination, origin, point search, and tracking processing in the measurement processor 0.

Next, the extraction processor 30 executes the following line segment extraction process.

Here, as the line extraction method, Japanese Patent Application No. 61-127490,
61-127491, 61-127492, 61
-The "linear pattern recognition method" shown in No. 127493 is adopted. The outline of this line extraction method will be explained below with reference to FIG.

That is, according to this extraction method, first, the projected waveforms in the X direction and the , n(D,: density data) in the projection direction is used.

(2) Next, the slit is sequentially rotated by a predetermined angle in the range of 0 to 90 degrees using the center of each slit area as the rotation center, and a projected waveform is obtained for each rotation angle.

If the projected waveform is obtained for each rotation angle in this way, the projected waveform in the direction along the linear pattern LP will be obtained when rotated by a certain predetermined angle θ, as shown in FIG. 6(C). The peak value P becomes the maximum, and from this, the linear pattern existence direction θ can be determined.

Next, the line width W of the linear pattern is determined by the length W of the peak cut line when the projected waveform is subjected to threshold processing using a predetermined threshold value T1,1, for example, as shown in FIG. 6(C). shall be.

Next, the length L of the linear pattern is limited to the already obtained crack width W in the slit area only in the width direction, as shown in FIG. 6(d), and the projected waveform is determined for the limited slit area. , the line length included in the slit area is determined by threshold processing using a certain threshold value Thd2.

By performing such imaging waveform analysis, the linear pattern included in one slit area is recognized as a rectangular pattern with width W, length, direction θ, etc., as shown in FIG. 6(e). be able to. As mentioned above, this is called a line segment.

Then, when such processing is performed for each mesh, mesh, and region in FIG. 6(a), a fa-shaped pattern is extracted for each mesh as a line segment with a width W and a length, and whose direction θ is known. be able to.

Then, the extraction processor 30 calculates an extraction result consisting of the following parameters.

1) Width W of the line segment 2) Length of the line segment 3) Direction θ of the line segment 4) Aspect ratio (L/W) 5) Area (L-W) 6) Slit brightness (of the density data in the slit) Average value) (7) Coordinates of line segments (8) Ridge line coordinates (A ridge line is a line segment represented by the peak position of the imaging waveform, and is expressed as the coordinates of the two vertices of the line segment) By the way, extraction As shown in FIG. 1, the processor 30 has a plurality of processors EP to EP512 that are expandable up to 512 units connected together using a bus connection method (FIG. 10(a)), and each processor has a Segment extraction for one slit area is performed, and each of these processors EP to EP512 can operate in parallel.

The idle processors sequentially receive the image data of the slit size transferred from the original image memory 2, execute the segment extraction process described above, and transfer the extraction results to the recognition processor 20 via the measurement processor 0. However, bus management in the extraction processor 30 is performed by the aforementioned measurement processor 0.

Further, in this extractor 10 setter 30, processing contents are varied based on the second-order determination result Id inputted from the measurement processor 0, so as to realize processing efficiency. In other words, the −th determination result Id is
In the case of , coarse but high-speed processing is executed, in the case of ■, fine processing is executed, and in the case of ■, segment extraction processing is not executed (the image data
(not forwarded to).

When the recognition processor 20 receives the extraction result for a processing unit (for example, 4 m in width x 2 m in length) from the extraction processor 30, it performs the following continuity determination process, Recognize cracks as line drawings. In this case, the recognition processor 20
The continuity determination process shown in No.-62242 is used. This process will be briefly explained below. Processing is performed according to the following steps.

(I) Competitive Segment Removal If there is a crack in the middle of adjacent slits, segments will be extracted from the same crack overlappingly, so in such a case, the segment with the smaller area is removed.

(II) Using the extraction results from the continuity determination extraction processor 30, several evaluation parameters representing the geometric positional relationship between adjacent slit segments as shown in FIG. 7 are determined, and these evaluation parameters are Continuity is determined using .

(1) Isolated point search (II) Search for isolated segments in the vicinity of the end points of the connected segment string, and connect those that are appropriately placed.

(IV) Search the vicinity of the end point and branch connection segment, and connect to another end point that has already been determined to be a crack, or generate a branch.

By the way, this recognition processor 20 is composed of a processor a, four processors b, and a processor C, as shown in FIG. When the processor a has stored all the data for the processing unit of the recognition processor 20, it performs predetermined preprocessing and then divides the data into four parts and transfers them to the four processors b.

The four processors b each process the data divided into four parts, but these four processors b are connected using the network connection method shown in FIG. When you need data,
It is now possible to receive data using a communication channel, so that there is no problem with overall processing. These four processors b perform parallel processing.

Processor C integrates the processing results of the four processors b and transfers the final result to system controller 40.

Next, the system controller 40 controls the entire system and connects the user interface 45, and displays the recognition results of the recognition processor 20 in the memory 3.
control such as transferring the data to the monitor 4 and displaying it on the monitor 4.

According to this system configuration, the extraction processor 30 that performs local processing employs a multi-microprocessor system using a bus coupling method, and the recognition processor 20 that performs global processing employs a multi-microprocessor system using a network coupling method. We aim to achieve efficient processing.

In other words, for local processing such as segment extraction, the amount of data is large, the data update frequency is large, the independence of the processing is high, and the number of processors required is large, so the bus coupling method is more suitable than the network method for this processing. ing. This is because, with the network connection method, when data is transferred to a distant processor, it must pass through many processors on the way, making it unsuitable for local processing with high data transfer volume and frequency. According to the data transfer time is as long as the processor! This is because the load of data transfer is small because it is not affected by

In addition, in large-scale processing such as continuity determination, the amount of data is small (because compressed data is used after segment extraction), the data update frequency is low, the independence of small processing is low, and the number of processors required is small. , the network coupling method is more suitable for this purpose than the bus coupling method. This is because in the bus-coupling method, the bus goes into a waiting state every time there is bus contention, but in the network type, the communication channel with adjacent processors is independent, so the bus does not go into a waiting state. .

Further, in the configuration shown in FIG. 1, the measurement processor 1
By performing segment extraction only in the selected region instead of all regions with 0, the processing time of local processing (segment extraction) and global processing (continuity determination) can be made equal, and the extraction processor 30 connected serially can be Pipeline processing with the recognition processor 20 can be performed smoothly. In other words, in the case of pipeline processing, efficiency cannot be maximized unless the processing time for each stage is approximately the same, but local processing using original image data (segment extraction) takes less processing time than global processing (continuity determination). In order to equalize these processing times, it is usually necessary to increase the number of processors on the local processing side, which increases the burden on the hardware. Since the primary judgment is performed and segment extraction is performed only in areas where cracks may occur, the time required to prepare segment extraction data for the processing unit of the recognition processor 20 is shortened, and each stage of the recognition processor 20 The processing time of the extraction and measurement processor 30.10 can be made the same as the processing time of the extraction and measurement processor 30.10. That is, the measurement processor 10 and the extraction processor 30
If the processing in is one stage, the processing in the recognition processor 20 will be in three stages in processors a, b, and c, and a total of four stages of pipeline processing can be realized.

Next, the operation of the system configuration shown in FIG. 1 will be explained.

First, the total length of road surface measurement data to be processed is input by the user to the system controller 40 via the user interface 45 . In FIG. 8(a), the distance specified by the user is an area of 0 m to 1000 m.

The recognition processor 20 receives the specified distance from the system controller 40, and outputs a segment extraction request to the measurement processor 10 for each recognition processor processing unit (cross section: 4 m x longitudinal section: 2 m) as shown in FIG. 8(b).

The address generation unit 11 of the measurement processor 10 sends the data transfer request address corresponding to the area specified by the recognition processor 20 to the original image memory 2. The primary determination unit 12 of the measurement processor 10 then processes the image data in units of slits transferred from the original image memory 2 (see FIG.
), the above-mentioned primary judgment process is executed, and the second judgment process result 1d is output to the extraction processor 30, and the above-described predetermined origin search and tracking process is performed based on the second judgment process result 1d. Execute.

In the extraction processor 30, based on the primary determination processing 1d sent from the measurement processor 10, (1) executes coarse but high-speed segment extraction processing; and (2) executes fine segment extraction processing. (2) Select one of ``Do not perform segment extraction'' and execute the selected process. and,
The segment extraction processing result is output to the recognition processor 20 via the measurement processor 10.

This segment extraction processing result is input to the address generation section tJ11 of the measurement processor 10, and the address generation section 11
Now, this processing result is used to efficiently generate addresses.

The segment extraction processing result via the measurement processor 10 is input to the processor a of the recognition processor 20. Processor a starts processing when the segment extraction results for the processing unit of the recognition processor shown in FIG. 8(b) are collected, and the processing results are divided into four parts and sent to four processors b. Processor C integrates the processing results of the four processors b and sends the final result to system controller 40. The system controller 40 displays the final recognition result on the monitor 4 by transferring the final result to the display memory 3, and the 0 or more results presented to the user are the overall operation of the system.

In this way, in the configuration of this embodiment, a cracked image processing system is configured by three types of dedicated processors consisting of measurement, recognition, and extraction processors 10, 20, and 30, and each processor performs parallel processing, while the overall In addition to increasing speed by executing line processing, a bus-coupled multi-microprocessor system is used for local processing such as segment extraction and edge detection, and network-coupled for global processing such as determining segment continuity. By adopting a multi-microprocessor system, we have realized even more efficient and high-speed processing.

In addition, in such an embodiment configuration, the recognition processor 20 that performs global processing receives a rough address from the measurement processor 10.
The segment extraction result can be obtained by simply inputting the data to the measurement processor 10, and the extraction processor 30 can determine the processing content based on the primary judgment result Id supplied from the measurement processor 10. Therefore, efficient processing can be achieved by sharing both hardware and software. This can be done with a simple configuration.

It should be noted that the primary determination process performed by the measurement processor 10 is not limited to that shown in the embodiment, but may also be performed in accordance with, for example, Japanese Patent Application No. 62-27.
The technology of No. 0017 should be adopted.

In other words, in this technique, the next tracking direction is determined based on the line direction θ of the extracted line segment.

〔Effect of the invention〕

As explained above, according to the present invention, local image processing is performed by a bus-coupled multi-microprocessor system, and global image processing is performed by a network-coupled multi-microprocessor system, thereby improving the efficiency of image processing. and speed-up can be achieved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is an explanatory diagram for explaining an example of a starting point search process, and FIG. 3 is a graph for explaining a method for determining the presence or absence of cracks.
4 and 5 are diagrams for explaining the tracking process, FIG. 6 is a diagram for explaining the line segment extraction principle, FIG. 7 is a diagram illustrating parameters for determining continuity, and FIG. 8 is a diagram used to explain the operation of the embodiment of the present invention, FIG. 9 is an explanatory diagram of local and global processing, and FIG. 10 is an explanatory diagram of bus connection and network connection. 2... Original image memory, 10... Measurement processor, 20
... Recognition processor, 30 ... Extraction processor, 4
0...System controller. Figure 1 (Q) (b) (C) (d) (e) (f) <h) Figure 4 Figure 2 Figure 3 (b) (e) (f) (h) Figure 5 Figure 6 Figure (9) b) Figure 8 (C) Kueko, cage - Kogo - 10,000 and Minnobu Sumon Manslogakotno) 11 P'fr9! Riichiku (ku) (b) Figure 9 (ku) *Tofu 7 axis each nine wards (b) Figure 10

Claims (1)

[Claims] High-speed image processing that divides an original image into a plurality of regions and performs local image processing for each divided region and global image processing in which local image processing results for each divided region are correlated with each other. The apparatus includes: a first image processing device having a plurality of microprocessors coupled by a bus coupling method, and performing the local image processing with the plurality of microprocessors; and a plurality of microprocessors coupled by a network coupling method. and a second image processing device that executes the global image processing using the plurality of microprocessors, and the first and second image processing devices are operated in parallel. Processing equipment.