JPH1115973A - Image recognition device - Google Patents

Abstract

(57) [Summary] Regarding recognition processing that allows rotation of a target, increasing spatial resolution in order to enhance recognition performance requires an increase in apparatus cost or processing time due to a rapid increase in the amount of arithmetic processing. Is difficult to achieve. SOLUTION: Input each element of a target pattern,
A rotation angle estimator (ST53) that outputs an estimated rotation angle by a neural network in which a synapse load value is set in advance based on a reference pattern, and a phase difference between the estimated output value and a preset expected value is output. The determination unit (ST6) that corrects the target and determines whether the target and the reference are identical according to the similarity between the expected value and the corrected output value.
To ST8).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image recognition apparatus that specifies an object by performing arithmetic processing on image data of the object.

[0002]

2. Description of the Related Art CCD (Charge Coupled)
A so-called image recognition technology for recognizing a specific target pattern by performing arithmetic processing on image data captured by a device camera or the like is a FA (Factory Automation).
)) is widely used in the field and the like. In particular, as a flexible recognition technique, it is strongly required in many image recognition processes that correct recognition be performed irrespective of the position and inclination (rotation angle) of a target pattern. For example, "Mitsubishi Vision Sensor Unit Catalog, June 1995,
In the image recognition device (vision sensor unit) used in assembly factories etc. shown in "Mitsubishi Electric Co., Ltd.", the function of recognizing the position and inclination (rotation angle) of parts such as valves has already been realized. Has been

Conventionally, as a method of extracting a position, a difference between an object and other colors or patterns is used, and preprocessing such as projection of the amount is performed relatively easily from the entire screen. I was able to identify the approximate location of the object. However, in order to realize recognition allowing an arbitrary inclination (rotation angle), there has been a problem that the amount of arithmetic processing for allowing the rotation of the object after roughly specifying the position is significantly increased.

Next, the operation will be described. FIG. 8 is an explanatory diagram showing an image matching method of a conventional image recognition device. Prepare reference patterns corresponding to all rotation angles for the component pattern to be recognized, evaluate the similarity of the target pattern with all the reference patterns, and set the similarity of the most similar reference pattern in advance. When the value is equal to or smaller than the threshold value, the target pattern is assumed to be a pattern to be recognized. The angle of the reference pattern corresponds to the inclination of the component. That is, in the conventional pattern recognition method, recognition of target patterns with various rotation angles is substantially realized by preparing reference patterns for all rotated cases and executing recognition processing on each of the reference patterns. . As a result, there has been a problem that the calculation amount of the recognition processing increases in proportion to the number of rotation angles of the reference pattern prepared in advance. Note that the target pattern may be rotated with one reference pattern. However, in this case as well, the calculation amount of the recognition process increases in proportion to the number of rotation angles of the target pattern. If the resolution of the rotation angle is n, the amount of calculation for allowing rotation is generally proportional to the square of n. Therefore, if the angle resolution n is increased to ensure recognition accuracy, the amount of calculation will increase rapidly. In addition, there has been a problem that the processing time and the cost of the apparatus increase rapidly. Therefore, it does not cause a sudden increase in the amount of computation,
There is a strong demand for the development of image recognition technology that allows rotation.

FIG. 9 is a flowchart showing the operation of the conventional image recognition apparatus. The image recognition device handled here is characterized by the following conditions. Hereinafter, the target pattern is called a target. The target is characterized by an array data pattern on a two-dimensional plane. It is assumed that the center position of the target is specified to include a certain error by a preset preprocessing or the like. The data on the same radius circle with the center position of the given target as the origin is rearranged in the circumferential direction, converted into so-called polar coordinates, and the two-dimensional array data having dimensions of the radius and angle as an input pattern, A pattern unit (target pattern) handled in the recognition process is set (step ST1 to step ST4).

A similarity (a scalar amount indicating that the smaller the value is, the more similar the similarity is) between the prepared reference pattern for the target to be recognized and the input target pattern is calculated (step ST5). The basis of the recognition processing is to determine that the target is a target pattern to be recognized when the target is less than or equal to the set threshold, and to determine that the target is another pattern when the target exceeds the threshold (steps ST6 to ST6). ST8). By keeping the basic configuration unchanged, preparing reference patterns corresponding to all expected rotation angles, calculating the similarity for all of these reference patterns, and executing a process of comparing with a threshold value Substantially, identification of a specific pattern to be recognized is realized irrespective of the rotation angle of the target. Since the similarity between the reference pattern and the input target pattern is generally evaluated by the sum of the absolute values of the differences or the squares of the corresponding elements, the amount of calculation required for calculating the similarity is proportional to the number of elements. As the number of elements (spatial resolution) for realizing high recognition performance increases, the amount of computation increases proportionally. In addition, the amount of calculation for allowing an arbitrary rotation angle increases by the number of gradations in the angle direction, and therefore increases rapidly in proportion to the square of the spatial resolution related to the rotation angle.

[0007]

Since the conventional image recognition apparatus is configured as described above, in the recognition processing in which the rotation of the target is allowed, if the spatial resolution is enhanced in order to enhance the recognition performance, it is difficult to calculate. There has been a problem that the apparatus cost and the processing time increase with a rapid increase in the processing amount, and it is difficult to realize high recognition performance.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems. In order to realize pattern recognition which allows rotation of a target with a realistic apparatus cost and processing time, a rotation with a smaller amount of calculation is required. It is an object to obtain an image recognition device that can recognize an image.

[0009]

According to the first aspect of the present invention, there is provided an image recognition apparatus which inputs each element of a target pattern and outputs an estimated rotation angle by a neural network in which a synapse load value is set in advance based on a reference pattern. A rotation angle estimating unit, and a phase difference between an output value estimated by the rotation angle estimating unit and an expected value preset by the rotation angle estimating unit. A determination unit that determines the identity of the target object and the reference object according to the degree of similarity with the output value.

According to a second aspect of the present invention, there is provided an image recognition apparatus, comprising: a pre-rotation processing unit for shifting a target pattern in an angular direction according to a plurality of rotation angles to generate a plurality of rotation target patterns; And outputs an output value according to the rotation angle estimated for each of the rotation target patterns by a neural network in which a synapse load value is set based on a reference pattern of a reference object generated in advance. A rotation angle estimating unit, a post-processing unit that corrects the output value according to a phase difference between an output value corresponding to the estimated rotation angle and an expected value preset by the rotation angle estimating unit, and an expected value And a similarity between the output value corrected by the post-processing unit and
A determining unit that determines that the target object is the same as the reference object when the similarity is within a predetermined value.

According to a third aspect of the present invention, in the image recognition apparatus, the rotation angle estimating unit has a plurality of output neurons, and assigns an assigned angle to each output neuron.
The post-processing unit corrects the output values according to the phase difference between the output value corresponding to the assigned angle estimated by the rotation angle estimation unit and the expected value for each assigned angle preset by the rotation angle estimation unit. Then, the determination unit calculates the similarity between the expected value for each assigned angle and the output value corresponding to the assigned angle corrected by the subsequent processing unit, and when the similarity is within a predetermined value, the object is determined. It is determined to be the same as the reference object.

According to a fourth aspect of the present invention, in the post-processing unit and the determination unit, an output value corresponding to the rotation angle estimated by the rotation angle estimating unit is input to the post-processing unit and the determination unit, and the predetermined expected value is set. The similarity between the expected value and the output value is output by a neural network in which a synapse load value is set based on the value.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a flowchart showing the operation of the image recognition device according to the first embodiment of the present invention. The image recognition device handled here is characterized by the following conditions. Hereinafter, the target pattern is called a target. The target is characterized by an array data pattern on a two-dimensional plane. It is assumed that the center position of the target is specified to include a certain error by a preset preprocessing or the like. The data on the same radius circle with the center position of the given target as the origin is rearranged in the circumferential direction, converted into so-called polar coordinates, and the two-dimensional array data having dimensions of the radius and angle as an input pattern, The unit of pattern (target pattern) handled in the recognition process is set (step ST1 to step ST4: target pattern generation unit). The processing of these steps ST1 to ST4 is the same processing as the conventional technique.

The functional configuration of the image recognition apparatus according to the first embodiment is characterized in that the similarity calculation processing function (step ST5) in the conventional apparatus has three processing functions, namely, a pre-rotation processing function (steps ST51 and ST52: Pre-rotation processing section), rotation angle estimation processing function (steps ST53, ST53)
54: a rotation angle estimating unit), and a post-processing function (step ST55: post-processing unit) for performing a correction process for calculating a similarity based on a plurality of estimated angles. The pre-rotation processing unit performs a process of converting the given target pattern into a rotation target pattern corresponding to the specified rotation angle. The rotation angle estimation unit has a function of outputting the rotation angle of the input pattern, and can be realized by a neural network.

This neural network generally uses an error back propagation method (EBP: Error Back) so as to output a rotation angle according to a sample pattern to be recognized in advance.
A multi-layer perceptron type neural network learned by Propagation can be used. The post-processing unit performs a correction process for calculating a similarity based on a change amount of the estimated angle with respect to the rotation target pattern rotated according to the plurality of designated rotation angles. Further, the similarity between the expected value in the neural network and the output value corrected by the post-processing unit is calculated, and when the similarity is within a predetermined value, it is determined that the target object is the same as the reference object (step ST6 to Step ST
8: determination unit).

Next, the operation will be described. A processing outline and a functional configuration for each functional unit will be described with reference to the drawings. FIG. 2 is a conceptual diagram showing a pattern of the pre-rotation processing. The target pattern in FIG. 2 extracted from the image data of the target object and subjected to the polar coordinate conversion is n × 5 in the direction of the angle 0 and 5 in the direction of the radial radius r.
The case of five array data is illustrated. In this case, the resolution in the angular direction is 360 / n degrees. However, in the first embodiment, the resolutions in the angular direction and the radial direction are not particularly limited. A rotated target pattern obtained by rotating the target pattern is shown in the upper part of FIG.
Here, a state where the rotation processing is performed by 360 × 3 / n degrees is illustrated. The rotation processing of 360 × 3 / n degrees is expressed by shifting (rotating) the target pattern to the left by three elements in the angular direction. can do. That is, a rotation process corresponding to the number of shifts × 360 / n degrees can be realized by the shift amount in the angular direction. In the image recognition apparatus shown in FIG. 1, since the rotation angle is given to the pre-rotation processing unit, the shift amount corresponding to the given rotation angle θ (= n × θ / 36
The function of outputting a rotation target pattern shifted in the angular direction by 0) may be realized. In addition, as a means for designating the rotation angle θ, a configuration may be employed in which the shift amount is directly given instead of giving the angle θ.

FIG. 3 is a conceptual diagram showing a rotation angle estimating unit using a neural network. In this rotation angle estimation unit,
An example is shown in which a three-layer hierarchical neural network (multilayer perceptron) is used. FIG. 4 is a conceptual diagram showing the relationship between the rotation target pattern and the input neurons. This input neuron corresponds to each element of the input rotation target pattern, and as shown in FIG. 4, when the rotation target pattern is n × 5 array data, 5n input neurons are required. However, the number of input neurons is not specified, and by appropriate processing, for example, by averaging the values in the radial direction for each angle, n
For example, a method of compressing the input number may be used. The requirement in the first embodiment relating to the correspondence of each element of the target pattern to the input neuron is that the relative positional relationship between the input neurons always corresponds to the relative angle relationship between the elements of the corresponding rotated target pattern. It's just what we do.

Although the number of intermediate neurons is not particularly defined, this neural network uses a reference target pattern (reference pattern) to be recognized in advance, and
An error backpropagation method (EBP: ErrorBack Propaga) is used to output the rotation angle of the target pattern.
In order to realize high discrimination ability, it is necessary to set an appropriate number adapted to a target to be recognized. That is, if the number of intermediate neurons is too large, the false recognition rate increases, and if it is too small, the recognition rate decreases. Therefore, in order to realize high discrimination ability, it is necessary to adjust the number to an appropriate number according to the problem. Although the number of output neurons can be increased in accordance with the complexity of the target pattern to be recognized, the simplest case is shown with a single output neuron. The expected value of the output neuron during learning is sin as shown in the upper part of FIG.
Although a periodic function such as a function is desirable, the functional form of the expected value does not specify the first embodiment.
This neural network performs back-propagation learning in advance with n rotation patterns having a shift number of 0 to n-1 per reference target pattern to be learned, with each rotation angle as an expected value. However, in the first embodiment, a learning method and the like are not particularly defined.

FIG. 5 is an explanatory diagram showing the processing contents of the post-processing unit and the determination unit. The target pattern to be recognized is set in advance based on the estimated rotation angle output for each rotation target pattern by the rotation angle estimation unit. This shows a process of outputting the degree of similarity with. First, based on the global characteristics of the output pattern, an expected value pattern (s
(in function). This phase difference Δ indicates the rotation angle between the target pattern of the target object and the reference target pattern, and the phase difference Δ can be obtained by a method of obtaining the difference between the positions of the centers of gravity, the least square method, or the like. However, the means for obtaining the phase difference Δ is not specified in the first embodiment. And
After correcting the output value by the phase difference Δ, the square sum of the difference between the output value and the expected value for each rotation angle is calculated, and the value is output as a similarity (smaller the value, the more similar). ,
When the similarity is within a predetermined value, it is determined that the target object is the same as the reference object.

As described above, according to the first embodiment, the similarity can be obtained by the arithmetic processing by rotating the target pattern several times by the generalization ability of the rotation angle estimating unit realized by the neural network. Can,
The improvement of the recognition performance can be realized by improving the angle estimation ability of the neural network. That is, in the image recognition processing that allows the rotation of the target,
It is necessary to repeatedly execute the similar calculation only in the case of n rotations (when the angle gradation is n), and the amount of calculation increases in proportion to the square of n. However, according to the first embodiment, the calculation amount required for the rotation image recognition process is n
And an abrupt increase in the amount of arithmetic processing associated with an increase in the angle gradation n can be prevented. In addition, the rotation preprocessing unit can easily generate a rotation target pattern simply by shifting the target pattern,
There is an effect that it is possible to suppress an increase in processing in generating the rotation target pattern.

Embodiment 2 FIG. FIG. 6 is a conceptual diagram showing a rotation angle estimating unit using a neural network according to Embodiment 2 of the present invention. In the second embodiment, the discrimination performance is improved by using a plurality of output neurons of the neural network of the rotation angle estimation unit in the first embodiment. As shown in FIG. 6, the assigned angle is assigned to each output neuron, and the output neuron # 0 is assigned an angle of 0.
180 degrees, output neuron # 1 has an angle of 90-270
The output neuron # 2 is assigned an angle of 180 to 360 degrees, and the output neuron # 3 is assigned an angle of 270 to 90 degrees. Thereafter, in the post-processing unit, the output value is corrected according to the phase difference between the output value corresponding to the assigned angle estimated by the rotation angle estimation unit and the expected value for each assigned angle preset by the rotation angle estimation unit. Then, the determination unit calculates the similarity between the expected value for each assigned angle and the output value corresponding to the assigned angle corrected by the post-processing unit, and when the similarity is within a predetermined value, the object is regarded as the reference object. It is determined that they are the same.

As described above, according to the second embodiment, by providing a plurality of output neurons and dispersing the respective reaction angle regions, the degree of freedom of the non-linear conversion can be increased, and the discrimination ability can be improved. There is an effect that can be.

Embodiment 3 FIG. FIG. 7 is a conceptual diagram showing a post-processing unit and a determination unit using a neural network according to Embodiment 3 of the present invention. In the third embodiment,
In the first embodiment, as a configuration for realizing a post-processing and determination unit for obtaining a similarity from a plurality of estimated angle values, FIG.
This uses a neural network as shown in FIG. An output value corresponding to the rotation angle estimated by the rotation angle estimation unit is input to each input neuron, and learning is performed in advance so that the similarity and the rotation angle are output to the two output neurons. The number of input neurons of the post-processing neural network corresponds to the number of rotations in order to recognize the target pattern.

As described above, according to the third embodiment, there is an effect that the post-processing unit and the determination unit can be configured by a neural network.

[0025]

As described above, according to the first aspect of the present invention, each element of the target pattern is input, and the estimated rotation angle is output by the neural network in which the synapse load value is set based on the reference pattern. A rotation angle estimating unit, a phase difference between the estimated output value and a preset expected value, the output value of which is corrected, and a target object according to the similarity between the expected value and the corrected output value. Since it is configured to include a determination unit that determines the same as the reference object,
Due to the generalization ability of the neural network, pattern recognition that allows rotation with a relatively small amount of computation can be performed,
There is an effect that the processing speed can be improved and the apparatus cost can be reduced.

According to the second aspect of the present invention, a pre-rotation processing section for shifting a target pattern in an angular direction according to a plurality of rotation angles to generate a plurality of rotation target patterns, and each element of each rotation target pattern And enter
A rotation angle estimator that outputs an output value according to a rotation angle estimated for each of the rotation target patterns by a neural network in which a synapse load value is set based on the reference pattern, and a rotation angle estimator corresponding to the estimated rotation angle. A post-processing unit that corrects the output value according to a phase difference between the output value and a preset expected value, and calculates a similarity between the expected value and the corrected output value, and determines the similarity to a predetermined value. The rotation pre-processing unit can easily generate the rotation target pattern simply by shifting the target pattern, since the determination unit determines that the target object is the same as the reference object when the value is within the value. Thus, there is an effect that it is possible to suppress an increase in processing in generating the rotation target pattern.

According to the third aspect of the present invention, the rotation angle estimating section has a plurality of output neurons, and assigns the assigned angle to each output neuron. The post-processing section estimates the rotation angle by the rotation angle estimating section. The output value is corrected according to the phase difference between the output value corresponding to the assigned angle and the expected value set for each assigned angle set in advance by the rotation angle estimating unit. A similarity with an output value corresponding to the assigned angle corrected by the processing unit is calculated, and when the similarity is within a predetermined value, the object is determined to be the same as the reference object. By providing output neurons and dispersing the respective reaction angle regions, the degree of freedom of the non-linear conversion can be increased, and the discrimination ability can be enhanced.

According to the fourth aspect of the present invention, the post-processing unit and the determination unit input an output value corresponding to the rotation angle estimated by the rotation angle estimation unit, and perform a synapse based on a preset expected value. Since the similarity between the expected value and the output value is configured to be output by the neural network in which the weight value is set, there is an effect that the post-processing unit and the determination unit can be configured by the neural network.

[Brief description of the drawings]

FIG. 1 is a flowchart showing an operation of an image recognition device according to a first embodiment of the present invention.

FIG. 2 is a conceptual diagram showing a pattern of a pre-rotation process.

FIG. 3 is a conceptual diagram illustrating a rotation angle estimating unit using a neural network.

FIG. 4 is a conceptual diagram showing a relationship between a rotation target pattern and input neurons.

FIG. 5 is an explanatory diagram illustrating processing contents of a post-processing unit and a determination unit.

FIG. 6 is a conceptual diagram showing a rotation angle estimating unit using a neural network according to a second embodiment of the present invention.

FIG. 7 is a conceptual diagram showing a post-processing unit and a determination unit using a neural network according to a third embodiment of the present invention.

FIG. 8 is an explanatory diagram showing an image matching method of a conventional image recognition device.

FIG. 9 is a flowchart showing the operation of a conventional image recognition device.

[Explanation of symbols]

ST1 to ST4 Target pattern generator, ST6 to
ST8 determination unit, ST51, ST52 rotation preprocessing unit,
ST53 Rotation angle estimation unit, ST54, ST55 Post-processing unit.

Claims (4)

[Claims] 1. A target pattern generation unit that generates a target pattern from image data of a target object, and inputs each element of the target pattern, and sets a synapse load value based on a previously generated reference pattern of a reference object. A rotation angle estimator for outputting an estimated rotation angle by the neural network, and a phase difference between an output value estimated by the rotation angle estimator and an expected value set in advance by the rotation angle estimator. An image recognition apparatus comprising: a determination unit that determines whether the target object is the same as the reference object according to the similarity between the expected value and the corrected output value. 2. A target pattern generating section for generating a target pattern having a radius and an angle as dimensions obtained by performing a polar coordinate conversion with the center on the image data as an origin from image data of an object, and a plurality of target patterns set in advance. According to the rotation angle, the target pattern is shifted in the angular direction to generate a plurality of rotation target patterns, and a rotation pre-processing unit, and input each element of each of the rotation target patterns,
By the neural network in which the synapse weight value is set based on the reference pattern of the reference object generated in advance,
A rotation angle estimating unit that outputs an output value according to the rotation angle estimated for each of the rotation target patterns, and an output value corresponding to the rotation angle estimated by the rotation angle estimating unit and the rotation angle estimating unit. A post-processing unit that corrects the output value according to the phase difference from the set expected value, and calculates a similarity between the expected value and the output value corrected by the post-processing unit, and determines the similarity to a predetermined value. An image recognition device comprising: a determination unit that determines that the target object is the same as the reference object when the value is within the value. 3. The rotation angle estimating section has a plurality of output neurons, and assigns an assigned angle to each output neuron. The post-processing section responds to the assigned angle estimated by the rotation angle estimating section. The output values are corrected according to the phase difference between the output value and the expected value for each assigned angle preset by the rotation angle estimating unit, and the determination unit corrects the expected value for each assigned angle and the subsequent processing unit. Calculating a similarity with the output value according to the assigned angle, and determining that the target object is the same as the reference object when the similarity is within a predetermined value. 2. The image recognition device according to 2. 4. The post-processing unit and the determination unit input an output value according to the rotation angle estimated by the rotation angle estimation unit,
4. The image recognition apparatus according to claim 2, wherein a similarity between the expected value and the output value is output by a neural network in which a synapse load value is set based on a predetermined expected value. .