JP2001175852A - Method and device for convolution processing - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and a device which enable high-speed convolution processing irrelevantly to the size of a mask in consideration of the fact that while the reliability of inspection is higher and higher as the mask for convolution is larger and larger, that can not be actualized because of a processing time. SOLUTION: A mask divided into arbitrary areas corresponding to multiples of sampling intervals of image data is set and cumulative image data of object pixels of convolution processing of the image data are generated by a cumulative image data generating means 100. With the output of the cumulative image data generating means 100, addition data of areas corresponding to respective stages of the mask are obtained by a corresponding area addition data acquiring means 200 and outputted to a convolution value output means 300. The convolution value output means 300 multiplies the respective corresponding area addition data by the values of respective corresponding stages of the mask, sums up all the multiplication results, and outputs the result as the convolution value of the object pixels of convolution.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a convolution in image processing performed for an inspection using an image pickup device.
lution) processing.

[0002]

2. Description of the Related Art In image processing, convolution is an operation performed between a kernel (Kernel), which is an array of numbers, and image data. , Just kernel center)
It refers to an arithmetic process of multiplying each value of the image data in the same area as the area of the kernel by each value of the kernel, adding up all of them, and outputting the result as a result value of the corresponding pixel. The kernel is an array of numerical values. When the kernel is used for processing image data such as a filter, the kernel is called a mask. Each pixel value of the image data is represented by I
(X, y), and the size of the mask is (2m + 1) × (2
n + 1) and the value of the mask (hereinafter simply referred to as the mask value) is M (x, y), the operation result value for each pixel can be expressed as in Equation 1.

(Equation 1) At this time, when the index of the image data is out of the area, a value of 0 is generally substituted. The mask value M (x,
The index of y) is described as having a positive value. FIG. 1 shows a one-dimensional example of such a calculation process.

The use of a filter in an inspection process for performing image processing is a general process, and the quality of an output image often varies depending on the size of a mask at this time. When increasing the size of the mask,
The result of the calculation becomes dull to the noise component of the image data, and the value that can be adjusted by the user increases, so that a more accurate result value can be obtained. However, in the calculation process, multiplication and addition of the number of times corresponding to the size of the mask are required for each pixel. Therefore, when the size of the mask is increased, the calculation time is rapidly increased. Therefore, the size of the mask is actually limited.

As is well known, in the convolution operation process, each pixel is multiplied by a corresponding value, and the multiplied values are summed to obtain a result value of each pixel. But,
As the size of the mask increases, the amount of calculation increases rapidly. Therefore, in order to solve such a problem, an invention was proposed in which a two-dimensional mask was separated into a one-dimensional mask, or real-time calculation was attempted using high-speed hardware.

[0005] Among them, in the case of the conventional method using hardware, in US Pat. No. 5,119,444, Laplacian and Gaussian are copied using hardware as a part of the invention. Here, hardware that replicates a 7 × 7 Gaussian is configured using three simplest forms of a Laplacian filter and a binomial form filter in succession.

FIG. 2 shows a conventional method in which data of a convolution target pixel and data of a pixel located around the target pixel (hereinafter referred to as a peripheral pixel) are stored using a delay unit at the time of acquiring image data. FIG.
Here, reference numeral 42 denotes a scanning electron microscope (Scanning Electron Microscope,
SEM), 43 is an analog / digital converter for converting the analog image signal acquired by the SEM into digital image data corresponding thereto, and 44 is a pattern compensator for outputting compensation data for pattern compensation. . Reference numerals 51 and 52 are switches,
Reference numeral 50 denotes a delay unit, and reference numerals 45 to 49 denote output lines for storing, in a latch, a signal of the combined image data passed through the switches 51 and 52 and a signal passed through a plurality of delay units 50, respectively.

FIG. 3 shows that the data of the surrounding pixels are summed up through summing circuits 60, 61 and 64 out of the five data obtained in FIG. 2, and the four data summed up by the summing circuits 60 and 61 are removed. FIG. 9 is a hardware diagram illustrating a process of multiplying the data of the target pixel through a shifter 66 and then subtracting the two values by a subtractor 68 to obtain a Laplacian.

FIG. 4 shows that the digital image signal obtained from FIG. 3 is delayed to a predetermined level by use of delay units 78 and 79 to obtain necessary data. FIG. 3 is a diagram illustrating a process of obtaining a one-dimensional Gaussian by using hardware. FIG. 5 shows a hardware process for obtaining 7n × 7n Gaussian using the circuit shown in FIG.

However, such a conventional convolution process has a disadvantage that the improvement of the processing speed depends on the hardware configuration, so that it is difficult to implement a filter or the like because the change is not easy, and the modification is not easy. Also, in real-time calculation, if the size of the mask must be increased in order to improve the quality of the calculation result, there is a disadvantage that the speed of the hardware becomes a problem and the cost increases greatly.

[0010]

SUMMARY OF THE INVENTION The present invention has been proposed to solve the problems arising from the above-described conventional convolution processing method and apparatus. That is, according to the present invention, the larger the size of the mask for convolution becomes, the higher the reliability of the inspection becomes. It is another object of the present invention to provide a method and an apparatus which enable high-speed convolution processing.

[0011]

According to a first aspect of the present invention, there is provided a convolution processing method for performing a convolution process between image data and a mask, the method comprising a multiple of a sampling interval of the image data. A mask setting process of setting a mask by dividing the image into an arbitrary region, a process of generating cumulative image data for generating cumulative image data of each pixel of image data and storing the same in a memory; A corresponding area combined data acquisition process for acquiring combined data of the corresponding area that is the area of the accumulated image data corresponding to the area, and multiplying the acquired corresponding area combined data by the value of each stage of the mask corresponding to the corresponding area After that, the configuration includes a convolution calculation process in which all of the multiplication results are summed and output as a convolution value of the convolution target pixel. To. According to a second aspect of the present invention, in the configuration of the first aspect, the size of the memory for storing the accumulated image data is such that the entire accumulated image data has a maximum value. , The size of which can store the sum data of them. According to another aspect of the present invention, there is provided a convolution processing apparatus for performing a convolution process between image data and a mask, which converts an image input from an image pickup device into digital image data. In the converted digital image data, the data of the current pixel that is the accumulation target pixel and the accumulation data from the line origin of the current horizontal line that is the horizontal line to which the current pixel belongs to the previous pixel that is the previous accumulation target pixel are added. By adding the result of the addition and the accumulated image data of the pixel located at the same horizontal position as the current pixel in the previous line, which is the horizontal line one line before the current current pixel, the image origin and the current Cumulative image data generating means for generating cumulative image data obtained by accumulating each data of a rectangular area formed by pixels; A predetermined number of partial area data are extracted from the accumulated image data generated by the generating means, and the partial data is selectively added and subtracted to obtain sum data of the image of the corresponding area, and a mask corresponding to the corresponding area is obtained. The corresponding area total data obtaining means for outputting the value of the corresponding step, the corresponding area total data obtained by the corresponding area total data obtaining means, and the value of the corresponding mask level are multiplied. And a convolution value output unit that outputs the result as a convolution value of the convolution target pixel. Further, in order to solve the above problem, the invention described in claim 4 is:
4. The image data output device according to claim 3, wherein the cumulative image data generation means converts an image input from the image pickup device into a digital value and outputs the digital value as image data f (x, y). The output of the previous accumulated image data storage for temporarily storing the data of the current pixel which is the accumulation target pixel in (x, y) and the accumulated value of the data of the horizontal line up to the previous pixel calculated in the previous accumulation. A first adder for summing the data, a buffer for storing the sum of the first adder, and the same horizontal direction as the current pixel in the previous line that is one horizontal line before the horizontal line to which the current pixel belongs The previous line accumulated image data storage storing the accumulated image data of the pixel at the position, the image data output from the buffer, and the previous line accumulated image data A second adder that adds together the output previous line accumulated image data and accumulates each data of a quadrangular region formed by the image origin and the current pixel, and stores the added result of the second adder as accumulated image data. And a cumulative image data storage. Further, in order to solve the above problem, the invention according to claim 5 is as follows.
4. The configuration according to claim 3, wherein the corresponding area sum data obtaining means obtains necessary four partial area data F _xy from the accumulated image data (F _xy (x, y)) generated by the accumulated image data generating means. (X2, y2), F _xy (x
1, y1), F _xy (x1, y2), F _xy (x2, y
First to fourth partial area extractors for extracting 1), four extracted partial area data F _xy (x2, y2), F _xy
(X1, y1), F _xy (x1, y2), F _xy (x
2, y1) for each of the two partial areas, a third summer and a fourth summer, a subtractor for subtracting the output data of the fourth summer from the output data of the third summer, and a subtractor. Is output to the corresponding area sum data F _xy (x1, y1, x
2, y2), and a mask corresponding stage value output device that outputs the value of the mask stage corresponding to the corresponding region. According to a sixth aspect of the present invention, in the configuration of the third aspect, the convolution value output means includes a plurality of corresponding area sum data (A) obtained from the corresponding area sum data acquisition means. F _xy (x1, y1,
x2, y2),... F _xy (xm, ym, xn, y
n)) and a number of multipliers each multiplying the value (V (1), V (2),... V (n)) of each stage of the mask corresponding to the plurality of corresponding regions, , And a convolutional value output device for outputting the output data of the adder as a convolutional value of the convolution target pixel.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 6 is a diagram for explaining a mask simplification process which is a part of the configuration of the convolution processing method according to the embodiment of the present invention, that is, a diagram showing a method for digitally counting a mask. Among them, FIG.
(A) is a case in which a calculation is performed using a mask obtained by sampling a theoretical mask at the same sampling interval as image data in a general case, and the digital digitization of a mask in a conventional convolution process will be described. are doing. In the present embodiment, as shown in FIG. 6B, the mask is set for an arbitrary area that is a multiple of the sampling interval of the image data. At this time, the number of mask steps is the same as or smaller than that of the conventional method.

In the two-dimensional case, the mask value having a symmetrical structure with respect to the center of the kernel is transformed into a pyramid shape in which the respective steps, which are square shapes, are stacked and placed. This is because the amount of calculation can be reduced by making the area to be masked multiplied by a square in order to utilize the accumulated image data.

Next, the accumulated image data will be described.
The cumulative image data is data obtained by accumulating data such as the brightness value of each pixel existing inside a rectangular area defined by the origin of the entire image data (hereinafter, simply referred to as the image origin) and the accumulation target pixel. This can be expressed by Equation 2 below. In Expression 2, F _xy (x, y) is the accumulated image data.

(Equation 2)

FIG. 7 shows an example of the accumulated image data. When accumulation processing is performed on each pixel of the image data shown in FIG. 7A, data shown in FIG. 7B is obtained. When the data shown in FIG. 7 (2) is further accumulated once, the data shown in FIG. 7 (3) is obtained.

In such an accumulation process, the size of the memory for storing the accumulated image data must be able to store the total of the entire data when the entire data has the maximum value. For example, if 512 × 512 data is used, the maximum value that can be stored must represent 255 × 512 × 512, so 26 bits are required for each pixel, and therefore, in order to store the entire data, Requires 1M bytes.

For example, the size of the mask is 5 × 5, the number of steps of the mask is 5, and the mask value is M (1, 1).
To M (5,5). The data of the target pixel of the convolution process and the data of the peripheral pixels are converted from I (1,1,) to I
If (5, 5) is set, the data of the convolution target pixel will be I (3, 3) because the kernel center is matched. The result value G of the convolution process at this time is calculated as in Expression 3.

(Equation 3)

At this time, if the mask is simplified as described above, the mask can be expressed in three stages, and in this case, only three mask values are required. That is, the center of the mask: M (3,3) = V (1) The next stage: M (2,3), M (2,3), M (2,3)
4), M (3,2), M (3,4), M (4,2), M
(4,3), M (4,4) = V (2) Last stage: All mask values are V (3)

Therefore, the equation (3) can be expressed as follows: G = I (3,3) × V (1) + [I (2,2) + I (2,3) + I (2,4) + I (3,3) 2) + I (3,4)
+ I (4,2) + I (4,3) + I (4,4)] × V (2) + [I (1,1) + I (1,2) + I (1,3)
+ I (1,4) + I (1,5) + I (2,1) + I (2,5) + I (3,1) + I (3,5) + I (4,1) +
I (4,5) + I (5,1) + I (5,2) + I (5,3) + I (5,4) + I (5,5)] × V (3), and This can be written as in Equation 4.

(Equation 4)

Therefore, the convolution time can be greatly reduced by simply calculating the sum of the data of the corresponding areas from the above-mentioned accumulated image data and using the difference value between the stages of the already provided mask. In other words, if the mask value is modified so as to be expressed in several stages, and the value of each stage is divided so as to be partially calculated in the next stage, the calculation process is greatly reduced. This point is shown in FIGS.

FIG. 8 is a diagram showing image data to be convolved. FIG. 9 shows cumulative image data of the image data shown in FIG. FIG. 10 shows the mask values, and FIG.
FIG. 1 is a diagram showing that the mask of the mask value shown in FIG. 9 is separated into two stages. FIG. 12 is a diagram showing a conventional convolution operation process, and FIG. 13 is a diagram showing a process of the convolution operation of the present embodiment in which a mask is divided into stages. FIG. 14 is a diagram showing a process of summing multiplied values as shown in FIG. 12 in order to obtain a result value of a target pixel in a conventional convolution operation process, and FIG. FIG. 14 is a diagram showing a process of summing multiplied values as shown in FIG. 13 to obtain a result value of a target pixel in FIG.

A comparison of the amount of calculation for two-dimensional data between the conventional method and the method of the present embodiment is as follows. The size of the target image data is M × N,
Assuming that the size of the mask is m × n, the conventional method requires multiplication and addition of the number of times corresponding to the size of the mask for each pixel to be convolved, so that M × N × (m × n multiplication and m × n times).

In the method of the present embodiment, when the number of mask steps is the maximum, three additions and one multiplication are required for each of M × N × 2 target pixels by the number of times corresponding to the number of mask steps. Therefore, [two additions + (three additions and one multiplication) × the number of mask stages] is required. Therefore, if
If the size of the mask is 13 × 13, there is a difference in the calculation amount of about 24 times the multiplication and 7 times the addition, so it can be seen that according to the present embodiment, the time required for the calculation can be greatly reduced. Therefore, by using the method of the present embodiment, the operation time is greatly reduced as compared with the case of using the existing method, and the method can be effectively used in a field requiring high-speed processing, and a large-sized mask can be used. , A filter, etc., can improve the accuracy of the result.

FIG. 16 is a block diagram showing an embodiment of the convolution processing apparatus according to the present invention. put it here,
Reference numeral 100 denotes an image input by an image pickup device, which is converted into digital image data, and data of a current pixel which is an accumulation target pixel in the converted image data and a current horizontal line which is a horizontal line to which the current pixel belongs. Adds the accumulated data from the line origin to the previous pixel that is the previous accumulation target pixel, and adds the result value and the same horizontal direction as the current pixel in the previous line that is one horizontal line before the current pixel. Is cumulative image data generating means for generating cumulative image data by adding data of a rectangular area formed by the image origin and the current pixel by adding the cumulative image data of the pixel at the position of.

The cumulative image data generating means 100 converts an image input from an image pickup device such as a camera into a digital value and outputs it as image data f (x, y), and an image data output device 111. A previous accumulated image that temporarily stores the data of the current pixel that is the accumulation target pixel of (x, y) and the accumulated value of each data of the horizontal line up to the previous accumulation target pixel calculated in the previous accumulation process. First adder 11 for adding the output data of data storage 112
3, a buffer 114 in which the sum by the first adder 113 is stored, and a pixel at the same horizontal position as the current pixel in the previous line which is one horizontal line before the horizontal line to which the current pixel belongs. A previous line cumulative image data storage unit 117 storing the cumulative image data,
By adding the data output from the buffer 114 and the previous line accumulated image data output from the previous line accumulated image data storage unit 117, each data in the square area defined by the image origin and the current pixel is accumulated. It is composed of a second summer 115 and an accumulated image data storage 116 for storing the sum of the second summer 115 as accumulated image data.

Reference numeral 200 denotes the cumulative image data F generated by the cumulative image data generating means 100.
_xy (x, y) from a predetermined number of partial area data F _xy (x
2, y2), F _xy (x1, y1), F _xy (x1, y
2), F _xy (x2, y1) is extracted, and this is selectively added and subtracted to obtain a corresponding area (x1, y1, x2, y2).
Is a corresponding area total data obtaining means that obtains the total data F _xy (x1, y1, x2, y2) and outputs the value of the mask stage corresponding to the corresponding area (x1, y1, x2, y2). . The corresponding area total data obtaining means 200 obtains four necessary partial area data F _xy (x2, y2) and F _xy from the cumulative image data (F _xy (x, y)) generated by the cumulative image data generating means 100. (X1, y1),
_{F xy (x1, y2),} F x y (x2, y1) first to fourth partial area extractor for extracting 211,212,21
3,214 and the four extracted partial area data F
_xy (x2, y2), F _xy (x1, y1), F
_xy (x1, y2) and _Fxy (x2, y1) are respectively added to the two partial areas by a third adder 215 and a fourth adder 216, and a fourth adder is obtained from the output data of the third adder 215. Subtractor 21 for subtracting the output data of the adder 216
7 and the output of the subtractor 217,
_xy (x1, y1, x2, y2) and a corresponding area sum data output unit 218, and a corresponding area (x1, y1,
(x2, y2), and a mask-corresponding stage value output unit 219 for outputting the value of the mask stage corresponding to the mask stage.

Reference numeral 300 denotes a corresponding area (x1, y) obtained from the corresponding area total data acquisition means 200.
1, x2, y2) corresponding area sum data F _xy (x1,
y1, x2, y2) and a corresponding mask corresponding stage value, sum all the multiplication result values of the corresponding area, and output the result value to the convolution value of the target pixel. The convolution value output means 300 outputs a plurality of corresponding area sum data F _xy (x1, y1, x2, y2) obtained from the corresponding area sum data acquisition means 200.
F _xy (xm, ym, xn, yn) and mask corresponding step values (V (1),
V (2),... V (n)), a multiplier 321 for summing the output data of each of the multipliers 301, and an output data of the summer 321. A convolution value output unit 331 that outputs a convolution value of the target pixel.

First, an image data output unit 111 in the cumulative image data generating means 100 converts an image input from an image pickup device such as a camera into a digital value, and outputs image data f (x,
y), the first summer 113 outputs the accumulated value of the image data from the line origin of the current horizontal line obtained from the previous accumulated image data storage 112 to the previous pixel, and the image data output unit 111. The image data of the current pixel is summed. The result value is temporarily stored in the buffer 114. The second adder 115 adds the image data output from the buffer 114 and the previous line accumulated image data output from the previous line accumulated image data storage 117 to form a quadrangular region formed by the current pixel and the image origin. Are accumulated. Then, the result value is stored in the cumulative image data storage 116,
As a result, accumulated image data is generated.

The accumulated image data thus generated is
The first to fourth partial area extractors 211 to 214 that are input to the corresponding area total data acquisition means 200 and are included in the corresponding area total data acquisition means 200
The necessary four partial area data are extracted from the accumulated image data generated from 0. Then, the third adder 215 and the fourth adder 216 add the extracted four partial area data to each of the two partial areas. then,
The output data of the fourth adder 216 is subtracted from the output data of the third adder 215 by the subtracter 217. Subtractor 217
Is output through the corresponding area sum data output unit 218. Then, the corresponding area total data acquisition output stage 200
If, as described above, the corresponding area total data is obtained,
The value of the mask stage corresponding to the corresponding area is also output through the mask corresponding stage value output unit 219.

The output of the corresponding area sum data and the value of the corresponding stage of the mask are performed by the convolution value output means 30.
The result is multiplied by the multiplier 301 within 0, and the result is summed by the adder 3
At 21, the output is summed with the outputs of the other multipliers, and the resulting value is output as a convolution value of the target pixel through a convolution value output unit 331.

In the above description, the process of calculating the convolution value for one target pixel has been described. In practice, however, the convolution process described above is repeated many times to obtain the convolution value of one image data as a whole. Note that, depending on the hardware configuration, the process of generating the accumulated image data includes:
It can be performed at the same time as the acquisition of digital image data, and after the cumulative image data larger than the size of the mask is calculated,
The process of obtaining the corresponding region sum data and the process of outputting the convolution value may be performed in separate processes.

FIG. 17 is a diagram showing an example of application of the configuration of the present embodiment to Laplacian Gaussian processing. Among them, FIG. 17A shows image data before processing. FIG. 17b shows the convolution value obtained by the conventional convolution processing method,
FIG. 17C shows the convolution values obtained by the convolution processing method according to the present embodiment, and FIG. 18 is a comparison diagram of the convolution result values. In the example shown in FIG. 18, the average value is about 131,
It is a processing result of an image having a bright spot having a difference of about 4 at the center. As can be seen from FIG. 18, it can be seen that the processing method of the present embodiment is not significantly different from the conventional processing method.

[0033]

As described above, according to the present invention, in the convolution processing, there is an effect that high-speed convolution processing can be performed regardless of the size of the mask.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating an example of a one-dimensional convolution process.

FIG. 2 is a diagram illustrating that data of a convolution target pixel and data of a pixel located around the target pixel are stored using a delay unit when acquiring image data in the conventional method.

FIG. 3 shows a target pixel obtained by adding data of surrounding pixels through summing circuits 60, 61 and 64 from among the five data obtained according to FIG. 2, and excluding the four data summed up by the summing circuits 60 and 61; Is a hardware diagram illustrating a process of multiplying the data of the data by a shifter 66 and then subtracting the two values by a subtractor 68 to obtain a Laplacian.

4 delays the digital image signal obtained from FIG. 3 to a predetermined level using delay units 78 and 79 to obtain necessary data, and uses primary circuits using summing circuits 70 and 74 and shifters 72 and 77. FIG. 3 is a diagram illustrating a process of obtaining an original Gaussian in terms of hardware.

FIG. 5 is a hardware diagram showing a process for obtaining 7n × 7n Gaussian using the circuit shown in FIG. 4;

FIG. 6 is a view for explaining a mask simplification process which is a part of the configuration of a convolution processing method according to an embodiment of the present invention;
That is, it is a diagram showing a method of digitally counting a mask.

FIG. 7 is a diagram illustrating an example of accumulated image data.

FIG. 8 is a diagram illustrating image data on which convolution is performed.

9 is cumulative image data of the image data shown in FIG.

FIG. 10 shows an example of a mask value.

11 is a diagram showing that the mask of the mask value shown in FIG. 9 is separated into two stages.

FIG. 12 shows a conventional convolution operation process.

FIG. 13 is a diagram illustrating a process of performing a convolution operation according to the present embodiment in which a mask is divided into stages.

FIG. 14 is a diagram showing a process of summing multiplied values as shown in FIG. 12 to obtain a result value of a target pixel in a conventional convolution operation process.

FIG. 15 is a diagram showing a process of summing the multiplied values as shown in FIG. 13 to obtain a result value of the target pixel in the embodiment.

FIG. 16 is a block diagram showing an embodiment of the convolution processing device of the present invention.

FIG. 17 is a diagram illustrating an example of application of the configuration of the present embodiment to Laplacian Gaussian processing.

FIG. 18 is a comparison diagram of convolution result values.

[Explanation of symbols]

 Reference Signs List 100 accumulated image data generation means 200 corresponding area sum data acquisition means 300 convolution value output means

Claims (6)

[Claims] 1. A convolution processing method for performing convolution processing between image data and a mask, comprising: a mask setting process of setting a mask by dividing the image data into an arbitrary area which is a multiple of a sampling interval of the image data; A process of generating cumulative image data for storing the cumulative image data in a memory; and a corresponding area for obtaining, from the cumulative image data, sum data of a corresponding area that is an area of the cumulative image data corresponding to the area of each stage of the mask. The sum data acquisition process, multiplies the acquired sum data of each corresponding region by the value of each stage of the mask corresponding to the corresponding region, and then sums all the multiplication results and outputs the result as the convolution value of the convolution target pixel A convolution processing method comprising a convolution calculation process. 2. The size of the memory for storing the accumulated image data is such that when the total accumulated image data has a maximum value, the sum of those data can be stored. The convolution processing method according to claim 1. 3. A convolution processing device for performing a convolution process between image data and a mask, wherein the convolution processing device converts an image input from an image pickup device into digital image data, and is an accumulation target pixel in the converted digital image data. The data of the current pixel and the accumulated data from the line origin of the current horizontal line, which is the horizontal line to which the current pixel belongs, to the previous pixel, which is the previous accumulation target pixel, are added. By adding the current pixel in the previous line, which is the previous horizontal line, and the accumulated image data of the pixel in the same horizontal direction as the current pixel, the accumulation of each data of the square area formed by the image origin and the current pixel is performed. A cumulative image data generating means for generating image data; and a cumulative image data generated by the cumulative image data generating means. Means for extracting a predetermined number of partial area data, selectively adding and subtracting the data to obtain corresponding area total data, and also outputting a value of a mask stage corresponding to the corresponding area. And multiplying each corresponding area total data obtained by the corresponding area total data obtaining means by the value of the corresponding mask stage,
A convolution value output means for summing all of the multiplication results and outputting the result value as a convolution value of a convolution target pixel. 4. An image data output device for converting an image input from an image pickup device into a digital value and outputting the converted image data as image data f (x, y), wherein the image data f (x, y) y), the data of the current pixel which is the accumulation target pixel and the output data of the previous accumulation image data storage for temporarily storing the accumulation value of the data of the horizontal line up to the previous pixel calculated in the previous accumulation. A first adder to be added, a buffer in which the sum of the first adder is stored, and a horizontal line at the same horizontal position as the current pixel in a previous line that is one horizontal line before the horizontal line to which the current pixel belongs. A previous line accumulated image data storage device that stores accumulated image data of pixels, image data output from a buffer, and output from a previous line accumulated image data storage device A second adder that adds together the previous line accumulated image data and accumulates each data of a quadrangular region formed by the image origin and the current pixel; and accumulated image data that stores the sum of the results of the second adder as accumulated image data. 4. The convolution processing device according to claim 3, comprising a storage device. 5. A method according to claim 1, wherein the corresponding area sum data obtaining means obtains necessary four partial area data F _xy (x2, x2) from the cumulative image data (F _xy (x, y)) generated by the cumulative image data generating means. y2), F _xy (x1, y1),
_{F xy (x1, y2),} F xy and first to fourth partial area extractor for extracting (x2, y1), extracted four partial region data _{F xy (x2, y2),} F x y (x1 , Y
1), a third adder and a fourth adder for adding F _xy (x1, y2) and F _xy (x2, y1) to each of the two partial areas, and a fourth adder from the output data of the third adder. A subtracter for subtracting the output data of the adder; and an output of the subtracter, the corresponding area combined data F _xy (x1, y
4. The apparatus according to claim 3, further comprising: a corresponding area sum data output device for outputting as (1, x2, y2); and a mask corresponding stage value output device for outputting a value of a mask stage corresponding to the corresponding region. Convolution processing device. 6. The convolution value output means includes: a plurality of corresponding area sum data (F _xy (x1, y1, x2, y) obtained from the corresponding area sum data acquisition means.
2),... F _xy (xm, ym, xn, yn))
, V (1), V (2),..., V (n), respectively, and a plurality of multipliers for multiplying the respective values (V (1), V (2),... V (n)) of the mask corresponding to the plurality of corresponding regions; 4. The convolution process according to claim 3, further comprising: a summation unit that sums the respective output data of the multipliers; and a convolution output unit that outputs the output data of the summation as a convolution value of a convolution target pixel. apparatus.