JPH07121713A - Pattern recognition method - Google Patents

Abstract

PURPOSE:To provide the recognition method capable of a recognition corresponding to the size change of an object in an image by extracting the direction component of an outline element from the outline of a detected image extracted from an image picked up with the image of the discriminating object, and preparing a normalized histogram from the distribution of relative frequency for each direction component. CONSTITUTION:When the work image of a reference work is inputted from an image input device 2, the outline of that work is extracted by an image processing part 3. The image processing part 3 displays the external size and hole or the like of the reference work as drawings shown by the outline by performing smoothing processing, differential processing and binarizing processing. This drawing is transformed into a chain code for each positioning candidate section such as a corner part or a hole by a chain code transformation part 4, and the normalized histogram is prepared from the chain code by a normalized histogram preparation part 5. Concerning the normalized histogram, normalization is performed by taking the direction exponent of the chain code on an abscissa and the relative frequency of the direction component of the chain code expressing the drawing on an ordinate.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition method for detecting an object by shape recognition from an image of the object, and is used, for example, in recognizing a positioning point of a work object by an industrial robot. Regarding a pattern recognition method.

[0002]

2. Description of the Related Art Conventionally, the methods adopted as the above pattern recognition methods can be roughly classified into two. One is a technique called blob analysis, which cuts out the target part from the image and finds the geometrical features of that part (blob) (area, center of gravity position, circumference, moment, number of holes, etc.) By comparing the feature amount with the feature amount of the model obtained in advance, it is possible to determine what the target blob is. Another method is a method called pattern matching or template matching, in which the images of the characteristic parts such as holes are stored as a model (template),
The object is recognized or the position of the object is detected by searching the input image for a part that matches the model.

[0003]

However, in the above-mentioned conventional method, when the size of the object in the image changes, that is, when the distance between the imaging device and the object changes. However, there is a problem that the object may not be recognized because the size of the object in the image changes. For example, when an image of an object is picked up and positioned by a visual sensor mounted on an industrial robot, the size of the image changes because the relative distance between the visual sensor and the object changes. SUMMARY OF THE INVENTION It is an object of the present invention to provide a pattern recognition method capable of recognizing a change in the size of an object in an image and also capable of coping with rotation or partial loss of the object in the image. .

[0004]

In order to achieve the above object, a first method adopted by the present invention is an identification target object in which a feature amount of a detection image extracted from an image including an identification target image is extracted in advance. In the pattern recognition method for recognizing the identification object by comparing with the feature quantity of the image, the contour line of the detected image is decomposed at a predetermined pitch, and the direction component of each decomposed contour line element is extracted to determine the direction. The appearance frequency of each component is calculated, and the appearance frequency is divided by the total number of contour elements to create a normalized histogram of contour elements, and a normalized histogram of contour elements of the detected image is obtained in advance. This is a pattern recognition method characterized by recognizing an identification target object by comparing with a normalized histogram of contour line elements. or,
The recognition of the identification target object can be performed using a neural network in which a normalized histogram of the identification target object image is learned as teacher data.

Further, the second method adopted by the present invention is to compare the feature amount of the detected image extracted from the image including the image to be identified with the feature amount of the image of the identification target object extracted in advance,
In the pattern recognition method for recognizing the identification object,
The contour line of the detected image is decomposed at a predetermined pitch, the direction component of each decomposed contour line element is extracted, the appearance frequency for each direction component is calculated, and the appearance frequency is divided by the total number of contour elements. Create a normalization histogram of the contour line elements and a rotation conversion histogram obtained by rotating and converting the normalization histogram at a predetermined angle. The rotation conversion histogram of the contour line elements of the detected image and the contour obtained in advance. This is a pattern recognition method characterized by recognizing an identification object by comparing a normalized histogram of line elements and a rotation conversion histogram. In addition, the recognition of the identification target object,
This can be performed by using a neural network trained with the rotation histogram of the identification object image as teacher data.

[0006]

In the image processing for pattern recognition according to the first method of the present invention, first, the direction component of the contour line element obtained by decomposing the contour line of the detected image extracted from the image of the identification object at a predetermined pitch is determined. To be extracted. The contour line element can be detected, for example, as a contour line pixel string of an image, and thus the contour line can be expressed as a chain code in which a direction component in which each pixel forms a contour line is extracted. By comparing this chain code with the previously extracted chain code of the identification object, it is possible to recognize whether or not the detected image is the identification object. In the present invention, in order to deal with the case where the size of the image of the identification target changes, a normalized histogram is created from the distribution of the appearance frequency for each direction component of the chain code. Since this normalized histogram corresponds to the contour line shape of the detected image, the shape feature can be recognized even if the size of the contour line of the detected image changes. Further, the recognition of the identification target object can be performed by using a neural network, and it is possible to deal with a partial defect of the contour line. Claims 1 and 2 correspond to this.
In addition, in the second method of the present invention, in addition to the processing by the first method, a rotation conversion histogram in which the created normalized histogram is rotated at a predetermined angle pitch is created.
By this processing, it is possible to cope with the case where the contour line in the detected image is rotationally changed by utilizing the excellent correspondence of the contour line to the graphic rotation of the chain code expression. This rotation conversion histogram can be used as teacher data for learning the neural network. Claims 3 and 4 correspond to this.

[0007]

Embodiments of the present invention will now be described with reference to the accompanying drawings to provide an understanding of the present invention. The following embodiments are examples of embodying the present invention and do not limit the technical scope of the present invention. Here, FIG. 1 is a flowchart showing the procedure of the pattern recognition method according to the present invention, FIG. 2 is a block diagram showing a configuration for executing the pattern recognition method according to the embodiment, and FIG. 3 is an image processing unit in the configuration of the embodiment. Fig. 4 is a block diagram showing the procedure for extracting the contour line of the object by Fig. 4, Fig. 4 is an explanatory diagram of the chain code expression for converting the contour line into the chain code, and Fig. 5 is a graph display showing an example of creating the chain code as a normalized histogram. is there. In the present embodiment, an example of pattern recognition for identifying a positioning point existing on a work from a work image picked up by a TV camera serving as a visual sensor of an industrial robot will be shown. The procedure of pattern recognition processing according to this embodiment will be described below with reference to the flowchart shown in FIG. or,
The processing operation in each processing step (S1, S2 ...)
Description will be made with reference to each configuration diagram and explanatory diagram.

The pattern recognizing apparatus 1 for executing the pattern recognizing method according to the present embodiment, by switching the selection switches 12b and 12c as shown in FIG.
A preparation state in which the pattern recognition unit 6 configured by a neural network learns the feature data of the part (identification target) to be a positioning point in advance, and a captured image of the work input from the image input device 2 configuring the visual sensor. The recognition state in which the feature amount of the detected image extracted from is identified by the neural network can be selected. First, the learning process for preparation for pattern recognition will be described. Selection switches 12a-12c
Is switched to the B side. The processing for this preparation will be described with reference to the flow shown in FIG.
When a reference work image is input from the image input device 2 (S1), the contour line of the work is extracted by the image processing unit 3 (S2). As shown in FIG. 3, the image processing unit 3 performs smoothing processing, differentiation processing, and binarization processing to display the outer shape, holes, etc. of the reference work as a line figure indicated by contour lines. The extracted line figure is converted into a chain code by the chain code conversion unit 4 for each positioning candidate site such as a corner portion and a hole (S3). The chain code divides the contour line at a predetermined pitch and extracts the direction component of each of the divided contour line elements. The pixels forming the contour line image are used as contour line elements, and the pixels between adjacent pixels are It can be shown as a series of directional indices linked according to the directional index. For example, the line figure as shown in FIG. 4B has a direction index 0 to 45 degrees divided as shown in FIG.
When expressed by 7, it can be converted into a chain code of 111177711133311. If this chain code is stored in combination with the starting point coordinate values of the line figure, the line figure can be stored with a storage capacity smaller than that of storing the coordinate values of all the pixels forming the line figure. .

Then, a normalized histogram is created by the normalized histogram creating unit 5 from the chain code (S4). The normalized histogram is, as shown in FIG.
The horizontal axis is the direction index of the chain code, and the vertical axis is the frequency of appearance of the chain code direction index that expresses the line figure, and is normalized. By this processing, it is possible to prevent the histogram of the chain code from changing when the size of the input image changes. Specifically, the frequency of appearance of each chain code direction index is divided by the number of pixels forming the contour line. By the processes of S1 to S4, the chain code of the contour line extracted from the input reference work image and its normalized histogram are created, so that each is stored in the chain code storage unit 7 and the normalized histogram storage unit 8. To be done. Next, the line graphics of the candidate parts to be the positioning points of the reference work are sequentially displayed on the display terminal 9, and the part to be the positioning points (identification purpose) is designated by the operator (S5). The normalized histogram of the part specified here is input to the pattern feature learning processing unit 10, and a rotation conversion histogram corresponding to the case where the input image is rotated is created (S6). In other words, by using the excellent feature of the rotation of the chain code representation of the line figure, if the data of the normalized histogram in which the line figure is rotated at a predetermined angle pitch is created, the target when the identification target is imaged is created. It is possible to deal with the case where there is rotation in the object position. In the rotation conversion histogram of the normalized histogram, in the case of the 45-degree division direction index shown in FIG. 4A, each direction index appearance frequency of the normalized histogram corresponds to the direction index corresponding to the counterclockwise rotation angle. Can be created by rearranging to.

Therefore, when the operator designates the position to be the positioning point from among the line graphics of the respective parts extracted as the candidate parts for the positioning point being sequentially displayed on the display terminal 9, it is created for each part. A rotation histogram is created by the pattern feature learning processing unit 10 for the normalized histogram. The pattern feature learning processing unit 10 trains the pattern recognition unit 6 configured by a neural network by a back propagation method using the normalized histogram and the rotation conversion histogram obtained by rotating the normalized histogram as teacher data ( S7). The weighting factor of each neuron obtained by this learning is stored in the pattern feature storage unit 11 (S8). By the above processing, the part to be the positioning point is set on the reference work, and the feature amount of this part is stored. When the positioning point is changed to another part, for example, when the work position by the robot is changed, when the designation by the operator is changed to a desired part, the feature amount is stored for the designated part by the above processing.

Next, the recognition process for the work to be measured will be described. Selection switches 12a, 12b, 12
When c is switched to the A side, processing for recognizing the designated portion stored by the above processing is performed from the input measurement work image. The processing procedure for this recognition will be described with reference to the flow of FIG. In FIG. 1B, the processing from image input (S9) to normalization histogram creation (S12) is performed in the same manner as the above-described processing of S1 to S4.
The created normalized histogram is input to the pattern recognition unit 6 (S13). Since the neural network constituting the pattern recognition unit 6 is loaded with the weighting factor of the neural network created as the feature amount of the positioning point stored in the pattern feature storage unit 11 (S1
4), the pattern recognition unit 6 can select the previously specified positioning point from the input image (S15).
In the above recognition processing, even when the distance between the camera and the measurement work changes and the size of the work in the input image changes, the chain code of the contour line is also converted into the normalized histogram. Accurate recognition operation is performed corresponding to the change. Further, even when the measured work position in the input image is rotated, since the weighting coefficient is stored in the rotation conversion histogram for each rotation angle, the work position can be rotated. Furthermore, since a neural network is used for shape recognition, it is possible to recognize even when a detection portion in the image is missing. In the embodiment described above, the method of recognizing the positioning point set on the work has been described, but the same processing can be applied to the case of recognizing a desired target object from a plurality of objects.

[0012]

As described above, the present invention extracts the direction component of the contour line element from the contour line of the detected image extracted from the image of the object to be identified, and normalizes it from the distribution of the appearance frequency for each direction component. Generalized histogram is created. Since this normalized histogram corresponds to the contour line shape of the detected image, the shape feature can be recognized even if the size of the contour line of the detected image changes. The recognition of the identification target object can be performed using a neural network. With this recognition method, it is possible to recognize even when the size of the identification symmetrical object in the image is changed, and it is possible to deal with a partial defect of the contour line or the like. Claims 1 and 2 correspond to this. Furthermore, in addition to the processing by the above method, by creating a rotation conversion histogram in which the created normalized histogram is rotated at a predetermined angle pitch, it is possible to cope with the case where the contour line in the detected image is rotated and changed. You can Since this rotation conversion histogram can be used as teacher data for learning the neural network,
The shape feature can be recognized even when there is a change in the size of the contour line of the detected image and rotation. Claims 3 and 4 correspond to this.

[Brief description of drawings]

FIG. 1 is a flow chart showing a processing procedure of a pattern recognition method according to the present invention with recognition preparation (a) and recognition (b).

FIG. 2 is a block diagram of an embodiment of a recognition apparatus that executes an embodiment pattern recognition method.

FIG. 3 is a flowchart showing a processing procedure in the image processing unit.

FIG. 4 is an explanatory diagram showing a direction index (a) for explaining a chain code and an example (b) of a coded figure.

FIG. 5 is a graph showing an example of creating a normalized histogram.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Pattern recognition device 2 ... Image input device 4 ... Chain code conversion unit 5 ... Normalized histogram creation unit 6 ... Pattern recognition unit 7 ... Chain code storage unit 8 ... Normalized histogram storage unit 10 ... Pattern feature learning processing unit 11 ... Pattern feature storage

Claims (4)

[Claims] 1. A pattern recognition method for recognizing an identification target object by collating a feature quantity of a detected image extracted from an image including an identification target image with a feature quantity of an image of an identification target object extracted in advance, The contour line of the detected image is decomposed at a predetermined pitch, and the direction component of each decomposed contour line element is extracted.
The appearance frequency of each direction component is calculated, and the appearance frequency is divided by the total number of contour elements to create a normalized histogram of contour elements, and a normalized histogram of contour elements of the detected image is obtained in advance. A pattern recognition method characterized by recognizing an identification target by comparing it with a normalized histogram of contour elements of the identification target. 2. The pattern recognition method according to claim 1, wherein the identification target object is recognized by using a neural network trained by using a normalized histogram of the identification target object image as teacher data. 3. A pattern recognition method for recognizing the identification target object by collating a feature quantity of a detected image extracted from an image including an identification target image with a feature quantity of an image of the identification target object extracted in advance. The contour line of the detected image is decomposed at a predetermined pitch, the direction component of each contour line element is extracted, the appearance frequency for each direction component is calculated, and the appearance frequency is divided by the total number of contour elements to calculate the contour line element In addition to creating a normalized histogram, a rotation conversion histogram obtained by rotating and converting the normalization histogram at a predetermined angle is created, and the rotation conversion histogram of the contour element of the detected image and the rotation conversion of the contour element obtained in advance A pattern recognition method characterized by recognizing an identification target object by comparing with a histogram. 4. The pattern recognition method according to claim 3, wherein the recognition of the identification target object is performed by using a neural network in which a rotation histogram of the identification target object image is learned as teacher data.