JP2000057337A - Image processor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To implement an optional filter kernel on the hardware and to fast produce an orientation map of a low-noise image. SOLUTION: This image processor includes a controller 1 which outputs a prescribed filter kernel shown in the discrete value that can be shown in a floating point, an image sensor device 2 which performs a convolutional operation based on the input image data obtained by plural photosensors and the filter kernel, a memory device 3 which stores the result of the convolutional operation and a main controller 4 which performs a prescribed weighting process to the convolutional operation result and also performs a total control operation. In such a constitution, the operations of the filter kernel which can be shown in the continuous floating points are simulated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a preprocessing for reducing the dimension of data of an input image, and an image processing for efficiently obtaining image data that can be used for postprocessing of a subsequent image based on a calculation result of the preprocessing. It concerns the device.

[0002]

2. Description of the Related Art As a conventional image processing apparatus for generating an orientation map used in the course of image processing, there is one proposed by Knutson and Granland described in the following document 1. Reference 1: H. Knutsson and G .; H. Gr
anlund. Texture analysis u
sing two-dimensionalquadr
attune Filters, IEEE Comput
er Society Works on Co
mputer Architecture for P
atern and Image Database
Management, pages 206-213, 198
3.

In the conventional image processing apparatus proposed by Knutson and Granland disclosed in the above-mentioned document 1, a quadrature pair filter (Quadra) is used.
cure pair filters) are used.
A quadrature pair filter refers to two filters that have the same frequency response, but differ in phase from each other by π / 2 radians. Such a pair of filters can be mathematically obtained by Hilbert transforming two different filters. The image processing device proposed by Knutson and Granland measures the orientation of the image area by examining the output value of the quadrature pair filter at the appropriate location in the image.

In the conventional image processing apparatus proposed by Knutson and Granland, it is difficult to detect the orientation in an arbitrary image area. To circumvent this difficulty, Freeman extended the above method proposed by Knutson and Granland,
An image processing device in which a quadrature pair filter can detect some orientations in an arbitrary given image region has been proposed (see Document 2). Reference 2: Freeman, W.M. T. , Steerab
le Filtersand Local Analy
sis of Image Structure (adjustable filter and local analysis of image structure), Massa
customs Institute of Tec
hnology, Doctoral Dissertat
ion, June 1992.

[0005] The Freeman image processing apparatus disclosed in Document 2 can detect, for example, vertical and horizontal orientations at intersections of intersections in an image area. Although the method adopted by Freeman can be widely and deeply mathematically processed, such as obtaining the optimal state of a certain filter for a certain problem, it can be applied to the hardware environment (essentially any filter kernel). Implementing a quadrature pair filter (value) was difficult.

[0006]

Since the conventional image processing apparatus is configured as described above, there is a problem that it is difficult to realize a necessary filter kernel by hardware. For example, conventional image detection semiconductor chips developed to date include (n / 8, n / 7, n
/ 6, n / 5, n / 4, n / 3, n / 2, n, 0) to realize a filter kernel value. Here, n takes a value of 1, 2, 3, 4, 5, 6, 7, and 8. In addition, for example, values such as -n / 8 and -n / 7 in which a negative sign is added to the above values can be realized by hardware, but an arbitrary filter kernel value can be realized by hardware. It was difficult to realize. That is, Freeman's filter kernel values are real values shown in an n × n array, and it is generally common to approximate these real values with filter kernel values that can be processed by a conventional image sensor device. There was a problem that it was difficult. That is, since the conventional image sensor device can only implement the values of a small subset of rational numbers, it is generally difficult to approximate a real value such as a Freeman filter kernel value. There is a problem that it is impossible to numerically implement it in a sensor device.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and implements an arbitrary filter kernel represented by a real value such as a Freeman filter kernel value on hardware. An object of the present invention is to provide an image processing apparatus capable of generating an orientation map of a low-noise image at high speed in order to improve the quality of an approximate value of the orientation map.

[0008]

In the image processing apparatus according to the present invention, the filter kernel supply means outputs a predetermined filter kernel represented by a discrete value which can be displayed by a floating point, and the image sensor means comprises a plurality of image sensor means. A light detection sensor, and performs a convolution operation using input image data obtained by detecting an image with the plurality of light detection sensors and the filter kernel output from the filter kernel supply means, and Means for storing an operation result of the convolution operation output from the image sensor means; control means for controlling operations of the filter kernel supply means, the image sensor means, and the memory means; The operation result stored in the filter kernel is read, and the operation result is stored in a location set in relation to each of the filter kernels. Run the weighting process is characterized in that to simulate the operation of the weighting process sum filter that can be displayed on the Floating Point continuously seeking the resulting value in the kernel. Then, an arbitrary filter kernel can be easily achieved on the image sensor device, and a filter kernel corresponding to an optimum filter kernel theoretically derived based on the heuristic rule is provided.

In the image processing apparatus according to the present invention, the image sensor means includes a plurality of light detection sensors arranged in a two-dimensional array, and uses a predetermined filter kernel supplied from the filter kernel supply means. Calculating the orientation of each pixel in the input image data obtained by the plurality of light detection sensors, storing the data of the calculation result regarding the orientation, and controlling the memory based on the heuristic rule. Determining an orientation of each pixel in the input image data corresponding to an individual point on an array formed by the light detection sensor, using data of a calculation result stored in the memory means. It is assumed that. That is, evaluate the orientation of each single point by comparing the results of several filter kernel outputs in a simple manner, and use each filter kernel to determine any particular orientation in the vicinity of a given pixel. However, by calculating its intensity and using this information, heuristics can infer which orientation is the strongest at a given point.

In the image processing apparatus according to the present invention, the control means calculates the orientation angle using the orientation of each pixel in the determined input image data, and calculates the angle of each orientation using the non-linear mapping. And the obtained orientation vector is averaged over a small sub-region of the two-dimensional data as the input image data, and averaged by a second nonlinear mapping. The orientation vector is inversely transformed into an averaged orientation angle expression, the orientation angle is averaged, and the input image data is averaged. As a result, noise in the image is reduced, and the quality of the orientation map is improved.

[0011] In the image processing apparatus according to the present invention, the image sensor means includes a plurality of variable sensitivity light detection sensors arranged in an array, and input image data obtained by detecting an image with the plurality of light detection sensors. And performing filtering using the discrete value filter kernel output from the filter kernel supply unit, the memory unit stores the result of the filtering operation, and the control unit stores the filtering operation result stored in the memory unit. The method is characterized in that a predetermined weight is applied to the operation result, the sum of the weighted values is calculated, and the operation of the continuous value filter kernel is synthesized.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. The image processing apparatus according to the first embodiment of the present invention synthesizes a continuous filter kernel that can be represented by a floating point, and its operation is based on the principle of linear superposition. The filter kernel is
It is a set of rational numbers of real numbers. The image filtering process is described by a plurality of terms in a mathematical expression as a convolution operation. In the convolution operation performed by the image processing apparatus according to the first embodiment, the input image area and the output image area surrounding the output pixel position are obtained for each output pixel by the image sensor apparatus 2 which takes in the input image and the filter kernel value. To form a weighted sum of the pixel values of The weight of this sum is given by the appropriate filter kernel coefficients, and the convolution operation has a linear characteristic. Because the convolution operation is mathematically linear, a filter kernel generated with continuous values is represented as a weighted sum of discrete value filters.

A discrete filter kernel of continuous values [-
Given 5.3 1; 0 4.8], the above-described filter kernel as a weighted sum of filter kernels having discrete values simply gives the following equation (1): [−5.3 1; 0 4.8] = − 5.3 [10; 00] + [01; 00] +4.8 [00; 01]. . . . . . . . . (1) can be expressed as Here, [ab; cd] is 2
× 2 array means the same in the following description.

Prior to obtaining a discrete-valued filter kernel using a possible floating-point representation, the procedure described above is applied to filter kernels of arbitrary size and set values, as long as the filter kernel coefficients can be represented. Expand,
It is generally possible to set that value. This procedure is called "filter disassembly." Equation (1) above
The filter of the form [10; 00] on the right side of is termed the "basic filter", and the coefficients before these terms are termed "basic filter weights".

The convolution operation executed by the image processing apparatus according to the first embodiment is a linear operation in a strict mathematical sense, and the output of the linear operation can be expressed as follows. Using image data array a, filter kernel F = F
(X, a) is defined as the convolution of the filter kernel array x, the result of which is the new output array F, and similarly, the filter kernel G = F
If (y, a), the filter kernel F (ix + ji, a) = iF (x, a) + jF
(Y, a) = iF + jG. Here, i and j are scalar constants, and x and y are the filter kernel arrays described above.

Therefore, if there is an arbitrary filter kernel z that can be represented by the equation z = ix + zy, a memory device 3 described later is provided in the image processing device, and an intermediate result of a necessary new addition operation is stored in this memory device. 3 and the non-implementable filter kernel z
Translate into two implementable filter kernels x and y. Thereby, x and y become filter kernel arrays that can be implemented on the semiconductor chip of the image sensor device 2. Finally, the equation F (ix + j
y, a) = iF + jG, coefficients i and j in filter z
Is exactly equivalent to the weight of the basic filter in the decomposition of

FIG. 1 is a block diagram showing an image processing apparatus according to a first embodiment of the present invention. In the figure, reference numeral 1 denotes a controller (filter / kernel supply means);
Output a predetermined pre-programmed filter kernel. Reference numeral 2 denotes an image sensor device (image sensor means) which is arranged in a two-dimensional array and has a plurality of light detection sensors having functions such as variable sensitivity, and uses a filter kernel output from the controller 1 to A predetermined image processing such as a convolution operation is performed on the input image data. Reference numeral 3 denotes a memory device (memory means) which stores an intermediate result of the image processing operation calculated by the image sensor device 2. Reference numeral 4 denotes a main controller (control means) for outputting final image data using the intermediate result stored in the memory device 3 and for controlling the operation of the entire image processing apparatus.

Next, the operation will be described. A plurality of light detection sensors provided in the image sensor device 2 enter a visual image emitted from a target object (not shown), and the input image data generated by the optical sensor. The convolution process is performed using the filter kernel F. After that, the processing result data output from the image sensor device 2 is stored in the memory device 3. These series of processes are performed on all the filter kernels G, H,... Determined by decomposing the filter kernel F which is originally a continuous value and supplied from the controller 1. . . Repeated for I. When the convolution operation results for all the filter kernels are stored in the memory device 3, the main controller 4 outputs a filter corresponding to the filter kernel F and a filter for generating a newly weighted output image array F '. Multiply the basic weight associated with kernel F by

Then, this multiplication processing is performed by a set of filter kernels {G, H,. . . , I} for all filter kernels. All operation results obtained by the multiplication process are {F ′, G ′,. . . , I ′}, and stored in the memory device 3 again. Finally, the main controller 4 reads the weighted output data newly stored in the memory device 3 and adds the corresponding elements in each element in the array of the weighted output data.
Through this process, the image processing apparatus according to the first embodiment performs the filtering using the arbitrary filter kernel F by using the filter kernels {F, G,. . . , I} and the basic filter weights obtained in the filter decomposition process.

FIG. 2 is an explanatory diagram showing the relationship between an input image, a filter kernel, and an output image in the image processing apparatus according to the first embodiment. In the convolution process, the filter
The kernel is multiplied one-to-one with the corresponding input image.
For example, the pixels A to K of the input image and the filter
Each pixel of A to K in the kernel is multiplied. After the multiplication, the nine multiplication results of A to K of the filter kernel are added together, and the addition result Σ is the memory device 3 corresponding to the position Ｅ in the output image corresponding to the position E in the input image. Is stored within.

As described above, according to the first embodiment, the image sensor device 2 captures the input image obtained by the image sensor device 2 and the filter kernel value output from the controller 1, and The controller 4 uses the memory device 3 to calculate, for each pixel, a weight given by an appropriate filter kernel coefficient of the pixel value of the input image area surrounding the output pixel position and the sum thereof, and to weight the discrete value filter. Outputs a filter kernel generated by continuous values expressed as a sum, so that the orientation to the image sensor device
A filter kernel for implementing the map algorithm can be easily achieved on the image sensor device, and a filter kernel corresponding to the optimal filter kernel derived in a heuristic sense is implemented in hardware. it can.

Embodiment 2 FIG. Since the configuration of the image processing apparatus according to the second embodiment is the same as that of the image processing apparatus according to the first embodiment, the components such as the controller 1, the image sensor device 2, the memory device 3, and the main controller 4 have the same reference numerals. This will be described with reference to FIG. In the image processing apparatus according to the second embodiment, an orientation map is generated by applying several predetermined filter kernels to an input image. An orientation map is an array obtained from an input image, and indicates an orientation in the input image. An angle of a reference vector in the input image is called an orientation. In the image processing apparatus according to the second embodiment, each filter used for detecting a specific orientation in an input image
Create a kernel to determine which pixel corresponds to which orientation based on simple heuristic rules. For example, in the example shown in FIG. 3, a total of eight filter kernels (a) to (h) are shown. Each filter kernel is responsive to pixels located along one direction of the principal axis every 9 pixels vertically and horizontally.

Next, the operation will be described. First of all,
The input image data obtained by the image sensor device 2 is filtered using a total of eight kernels (a) to (h) shown in FIG. The output value of the filtering operation result from the image sensor device 2 is stored in the memory device 3. Next, the main controller 4 compares the result of the filtering operation for each pixel. The results of these operations are evaluated using a heuristic rule, and the average output of all filter kernels at a specific point is averaged between the maximum (S _max ) and minimum (S _min ) of the filter kernel outputs. Compare using values.

FIG. 4 is an explanatory diagram showing the result of the comparison in numerical values. For example, the maximum value (S _max ) of the filter kernel output is 4 and the minimum value (S _min ) is 2.
As can be inferred from this procedure, a given pixel should be assigned to the maximum output filter kernel orientation or the minimum output filter kernel orientation by comparing the overall average filter output with the minimum and maximum average filter outputs. Or not.

A heuristic rule that can be used for the above operation is given by, for example, the following inequality (2). 4C + (s _max + s _min )> 3/8 × {s _i . . . . . . . . . . . (2) In this example, the quantity (s _max + s _min ) represents the average of the minimum filter kernel output at the same location and the maximum ( _max ) filter kernel output at a particular location in the output data array. Here, C indicates the value of the unfiltered pixel. These quantities (s _max + s
_In order to calculate _min ), the image processing apparatus according to the second embodiment performs the following operation.

First, in the image sensor device 2, eight predetermined filter kernels (a) to (h) shown in FIG.
Is used to filter images incident on a plurality of image sensors provided in the image sensor device 2, that is, an array of image sensors (not shown). The calculation results of the filter kernels (a) to (h) by the image sensor device 2 are stored in the array as s ₁ ,. . . s _i,. . . And stores named s _8. Each of these output arrays is continued pixel by pixel and the corresponding pixel elements of each of the eight output arrays are compared. The maximum and minimum values of the eight elements are stored along with the index i of the particular filter having these maximum and minimum values. These two quantities represent the left side of inequality (2) above. The right side is formed by taking the sum of the corresponding outputs for each of the eight filters.

The right side of the above inequality (2) represents the average value of all filter outputs at the same point in the output data array. By evaluating the inequality, that is, determining whether or not the inequality (2) is satisfied by the main controller 4, the direction of the orientation of the image in the area within the filter kernel under consideration is determined. . In particular, if the main controller 4 determines that this inequality does not hold, then the orientation of the image in the area under consideration will be based on those axes in the filter kernel with the minimum filter output at that point. Corresponds to the orientation.

Conversely, if the main controller 4 determines that this inequality holds, then the orientation of the image in the area of the point under consideration will result in the orientation of those axes in the filter having the largest output value at the same point. Corresponds to the orientation. When determining the maximum output and the minimum output, the magnitude including the sign of the filter output is considered. The numerical example shown in FIG. 4 shows this concept more clearly.

FIG. 5 is an explanatory diagram showing a specific example of the orientation map. In the figure, (a) shows an input image input by the image sensor device 2. Then, this input image is binarized (Binariza)
5), 12 as shown in FIG.
A binary image of 7 × 127 pixels is obtained. Based on the binarized image, the orientation of each filter kernel is calculated, and an entire orientation map of the input image as shown in FIG. 5C is obtained. The orientation map shown in FIG.
This is the resolution of a two-dimensional vector.

FIG. 6 is an explanatory diagram showing a process of binarization by the image sensor device 2. In FIG. 6, (a) of FIG. 6 shows a numerical value of an image of each pixel.
This corresponds to the numerical value given to each filter kernel shown in FIG. FIG. 6C shows an equation for obtaining an orientation map. FIG.
FIG. 3 is an explanatory diagram showing filtering using a filter kernel with respect to certain input image data obtained by the image sensor device 2. In the figure, each filter kernel has 9 × 9 pixels.

Inequality (2) by the main controller 4
In the evaluation operation (1), as shown in FIG. 8, a numerical value is given to the orientation at a specific point in the output data array. In the drawing, (a) shows a case where the orientation angle of a certain pixel is 63.4 degrees, and (b) shows a case where the orientation angle of a certain pixel is -45 degrees. Since the inequality holds because which filter is stronger at a particular point, a new value corresponding to the angle of orientation of those axes in the corresponding filter kernel is assigned to each point, hence Each point in the output data array, one of eight, corresponding to the orientation of the axis of the filter kernel under consideration, at this stage has an angle value associated with the new value. become.

The averaging process by the main controller 4 takes in the output angles obtained by the process using the heuristic rule and averages these angles. That is, the conversion f (x) = (cos2x, sin2x) is executed, and each angle is converted into a normalized vector. Due to the discontinuity at π that exists in the concentric table system, this conversion operation or one equivalent functionally to this conversion is required.

The angles corresponding to 0 and π radians in these processes represent the same direction, and when they are averaged, they correspond to one of the two original angles. Averaging 0 and π radians gives a value of π / 2 radians, which is undesirable because it is orthogonal to the original direction. By executing the above conversion before the averaging process by the main controller 4, the orientation O converted to the vector O = (1,0) and the orientation π converted to the vector P = (1,0) Is found. Next, 1/2
Evaluate the average of these vectors defined by × (O + P). This averaging is a normal arithmetic averaging and can generally be extended to any number of vectors to be averaged.

The average value in the above example can be evaluated as Θ = (1, 0). To convert back to the original angle representation, perform an inverse transformation to obtain 、 = × arctan (0/1) = × arctan (0 ⁺ ) = 0. Here, 0 ⁺ of the title is meant to be approached from the point of origin the right to 0. Function f (x) = arctan
(X) is a normal main branch, that is, arctan (x) ∈
(-Π / 2, π / 2) is not defined on FIG.
(A), a branch that becomes discontinuous at the origin 0: y = arctan (x) ∈ (0, π). FIG. 9B is an explanatory diagram showing an example of a standard branch approaching zero.

The orientation of the filter kernel will be described more clearly with reference to FIG. FIG.
Consider two nearly vertical orientations shown at (a) of 0, θ = π / 2 and δ = π / 2−ε. Where ε
Has a smaller angle than π / 2. By the above-described conversion, as shown in FIG. 10B, Θ = (− 1, 0)
And Δ = (− cos2ε, sin2ε). Therefore, it can be seen that the average vector 等 し い is equal to ×× ((− 1−cos2ε), sin2ε), as shown in FIG. In order to inversely convert the vector expression into the angle expression, the main controller 4 performs an inverse transformation of f (x). As shown in (d) and (e) of FIG.
The inverse of the vector Α is represented by the symbol α, where α = γ / 2.
Here, γ = arctan (α ₂ / α ₁ ). Further, α ₂ and α ₁ correspond to the y and x components of the vector Α, respectively.

Finally, in the limit where ε approaches 0, ie, as δ and the (intuitive) average approach π / 2, the arctan (x) branch defined above is used. Is that α approaches π / 2. Also, when standard branching is used, α approaches 0.

Therefore, in the final stage of the orientation process, the following process is performed. The orientation assigned to each point corresponds to the maximum response of one of the eight filter kernels described above, and the array Z
Is stored within. Thus, each element of array Z is a discrete angle value between 0 and π. The main controller 4 executes an averaging process on the array Z.

As an averaging process, the main controller 4 first converts the array Z into an arbitrary predetermined N × M size equal to a specific sub-region in the original array Z as shown in FIG. Divide into grids. Here, the grid means an array Z having n × m components, for example, n / m.
Individual elements obtained by sub-dividing into an N × M array having a size of N × m / M. The new N × M sub-elements are
Are themselves new arrays, these new arrays Z
_The main controller 4 executes the above-described averaging process for each of _ij . That is, the array Z _ij is n / N ×
One of the arrays of size m / M is shown. surely,
The array Z _ij contains more than one element and the averaging process is extended for all individual elements of Z _ij . In this way, after completion of the averaging process by the main controller 4, the array Z (having an original size of n × m) is
N × M consisting of averaged orientation angles
Is replaced by a new array Z ′ of size (N <n, M <m).

In the original array Z, each element is one of a set of discrete values from 0 to π, that is, as shown in FIG. 3, a set of eight discrete values (π / 2, arct
an2, π / 4, arccot2,0, π-arcco
t2, 3π / 4, π-arctan2), but the advantage of the newly formed array Z ′ is clearly the advantage of the array Z _ij corresponding to each region in the original image input by the image sensor device 2. The elements of array Z 'obtained through averaging will have consecutive floating point angle values on the same interval (0, π). As described above, since the newly created array Z ′ has continuous values, it is possible to more accurately specify the direction of the structure in the analyzed image.

Therefore, in the image processing apparatus according to the second embodiment, each position of the finally obtained image output array corresponds to the floating point corresponding to the orientation of a specific area in the original image input by the image sensor apparatus 2. An orientation map can be generated to include the angle values of

As described above, according to the second embodiment, the result of the output of the filter kernel is compared to evaluate the orientation of each single point, and the main controller 4 determines the orientation near the given pixel. Calculates the intensity of the image and uses the result of the calculation to heuristically infer which orientation is the strongest at a given point. An orientation map can be generated, and noise in the input image can be reduced to improve the quality of the approximation of the orientation map.

[0042]

As described above, according to the present invention, the filter kernel supply means outputs a predetermined filter kernel represented by a discrete value which can be displayed in a floating point, and the image sensor means outputs the input image. Using the data and the filter kernel, a convolution operation is performed, the memory means stores the operation result of the convolution operation, and the control means reads the operation result stored in the memory means, The two-dimensional filter kernel is configured to simulate the operation of the filter kernel that can be displayed in a continuous floating point by executing a predetermined weighting process and an addition operation thereof set in association with each other. There is an effect that an image processing device that realizes the above calculation can be implemented on a semiconductor chip.

According to the present invention, the image sensor means calculates the orientation of each pixel in the input image data using the filter kernel, the memory means stores the data of the operation result regarding the orientation, and the control means Is configured to determine the orientation of each pixel in the input image data based on the heuristic rule using the data of the above-described operation result, so that the image processing that realizes the operation of the two-dimensional filter kernel The effect is that the device can be implemented on a semiconductor chip.

According to the present invention, the control means calculates the orientation angle by using the orientation of each pixel in the determined input image data, and normalizes the angle of each orientation by using non-linear mapping. And the obtained orientation vector is averaged over a small sub-region of the two-dimensional data that is the input image data, and the averaged orientation vector is obtained by a second nonlinear mapping. Is inversely transformed into an averaged orientation angle expression, the orientation angle is averaged, and the input image data is averaged, so the noise in the input image is reduced and the approximate value of the orientation map is obtained. The effect that the quality of the can be improved A.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an image processing apparatus according to a first embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a relationship between an input image, a filter kernel, and an output image.

FIG. 3 is an explanatory diagram showing an example of a filter kernel.

FIG. 4 is an explanatory diagram expressing a comparison result of a filter kernel output by a numerical value.

FIG. 5 is an explanatory diagram illustrating an example of an orientation map in the image processing device according to the second embodiment of the present invention;

FIG. 6 is an explanatory diagram showing a process of binarization by an image sensor device in the image processing device according to the second embodiment of the present invention;

FIG. 7 is an explanatory diagram showing an example of filtering an input image using a filter kernel in the image processing apparatus according to the second embodiment of the present invention;

FIG. 8 is an explanatory diagram showing a specific example of an orientation in the image processing device according to the second embodiment of the present invention;

FIG. 9 is an explanatory diagram illustrating an example of branching in the image processing device according to the second embodiment of the present invention;

FIG. 10 is an explanatory diagram relating to an orientation of a filter kernel in the image processing apparatus according to the second embodiment of the present invention;

FIG. 11 is an explanatory diagram showing an example in which an array in an averaging process is divided into grids of a predetermined size in the image processing device according to the second embodiment of the present invention;

[Explanation of symbols]

1 controller (filter / kernel supply means), 2
Image sensor device (image sensor means), 3 memory device (memory means), 4 main controller (control means).

 ──────────────────────────────────────────────────続 き Continued on the front page F term (reference) 5B057 BA02 BA13 CA08 CA12 CA16 CB08 CB12 CB16 CC01 CD20 5C077 LL18 LL19 PP01 PQ12 PQ18 PQ22 SS01

Claims (4)

[Claims] 1. A system comprising: a filter kernel supply unit for outputting a predetermined filter kernel represented by a discrete value that can be displayed as a floating point; and a plurality of light detection sensors, wherein an image is detected by the plurality of light detection sensors. Image sensor means for performing a convolution operation using the input image data obtained by the above and the filter kernel output from the filter kernel supply means, and the operation result of the convolution operation output from the image sensor means. Memory means for
Controlling the operation of the filter kernel supply unit, the image sensor unit, and the memory unit, reading an operation result stored in the memory unit, and relating the operation result to each of the filter kernels Control means for executing a predetermined weighting process set as described above, calculating a sum of values obtained by the weighting process, and simulating an operation of a filter kernel which can be displayed in continuous floating point. Processing equipment. 2. An image sensor means comprising a plurality of light detection sensors arranged in a two-dimensional array, wherein the plurality of light detection sensors are provided by using a predetermined filter kernel supplied from a filter kernel supply means. Calculate the orientation of each pixel in the input image data obtained by the sensor, the memory means stores data of the operation result regarding the orientation, and the control means stores the data in the memory means based on the heuristic rule. 2. The orientation of each pixel in the input image data corresponding to an individual point on an array constituted by the light detection sensor is determined by using data of the stored operation result. Image processing device. 3. The control means calculates an orientation angle using the orientation of each pixel in the determined input image data, and normalizes the angle of each orientation using a non-linear mapping. And the obtained orientation vector is averaged over a small sub-region of the two-dimensional data that is the input image data, and the averaged orientation vector is averaged by a second nonlinear mapping. 3. The image processing apparatus according to claim 2, wherein the orientation of the input image data is averaged by inversely converting the orientation angle into an orientation angle, and averaging the orientation angles. 4. An image sensor means comprising a plurality of variable-sensitivity light detection sensors arranged in an array, and input image data obtained by detecting an image with said plurality of light detection sensors and said filter kernel supply means. Performing filtering using the discrete value filter kernel output from the memory unit, the memory means stores the result of the filtering operation, and the control means assigns a predetermined weight to the operation result stored in the memory means. 2. The image processing apparatus according to claim 1, wherein the sum of the weighted values is calculated to synthesize the operation of the continuous value filter kernel.