JPH0652317A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To reduce auxiliary work conventionally required for standard dictionary preparation or the like so as to recognize an object at the pattern recognizing device. CONSTITUTION:A pair of feature levels are calculated for each standard objective 1 previously selected for each kind of objects through an image input part 2, A/D conversion/preprocessing part 3 and feature detection part 4, and stored in a feature level storage part 7. Next, a pair of feature levels are extracted from the storage part 7 for each standard object and inputted to a neuro computer 5, and the computer learns the input so as to minimize the square error of the identification output and a desired identification output as teacher data. Afterwards, the feature level of the unknown object 1 is given to the computer 5 having completed the learning, and the identified result of the object is provided from the identification output through a judge part 6.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition device for recognizing what an object is or for inspecting the quality of the object.

[0002]

2. Description of the Related Art In a conventional pattern recognition apparatus of this type, an object is observed by an image input apparatus, the input image is processed, and feature quantities such as area, width, and height are extracted, and each object is detected. When creating a standard dictionary that gives the reference value of the feature quantity, the dictionary was created by performing statistical processing by changing the position / orientation multiple times in order to consider the variation due to the position / orientation of the object.

Further, the determination is made by comparing the feature amount of the input image with the feature amount of the dictionary and determining whether or not it is within the range of the upper and lower limit values. That is, the feature quantity of the input image is (x ₁ , x ₂ , ..., X _i , ..., X _n ) from the numbers i = 1 to n, and the i-th feature quantity dictionary of the object A is (Amin _i , Amax _i ), the object of the input image is determined to be A when Amin _i ≦ x _i ≦ Amax _i holds for all (x _i ) of i = 1, ..., N. When this relation is not established, similarly, it is compared with the feature quantity dictionary (Bmin _i , Bmax _i ) of the object B, and Bmin _i ≦ x _i ≦ for all (x _i ) of i = 1, ..., N. When Bmax _i is satisfied, the object of the input image is determined to be B. In this way, it was sequentially compared with the feature quantity dictionary to confirm what the object was.

Further, when inspecting the quality of an object, it is premised that what the object is is known, and it is inspected whether it is abnormal by comparing the feature amount dictionary with the feature amount of the input image. Was.

[0005]

As described above, in the conventional pattern recognition apparatus, it is necessary to input the image a plurality of times in order to consider the variation in the position / orientation of the object in order to create the standard dictionary. Engineering was troublesome. Furthermore, when there are variations in the target object, it was necessary to change the target object and obtain a large amount of data. Moreover, it was troublesome to deal with the case where the target changed due to the line change.

An object of the present invention is to eliminate the above problems, and it is an object of the present invention to provide a pattern recognition apparatus which can be recognized only by learning once or a small number of data.

[0007]

In order to solve the above-mentioned problems, the pattern recognition apparatus according to claim 1 observes an object (1 or the like) through image input means (image input unit 2 or the like), Processing the input image to determine what the object is,
Alternatively, in a pattern recognition device that outputs a signal for identifying whether or not the object is abnormal, means for processing the input image to extract a feature amount (A / D conversion / preprocessing unit 3, feature extraction unit 4) and a neuro computer (5, etc.) that inputs the extracted feature quantity, performs arithmetic processing, and outputs the above-mentioned identification signal.

According to a second aspect of the pattern recognition apparatus of the first aspect, in the pattern recognition apparatus of the first aspect, the learning of the neurocomputer is performed from one target object which is a standard selected for each type of the object. A set of feature quantities extracted through the quantity extracting means is input to the neurocomputer, and at this time, the identification signal actually output from the neurocomputer and the desired identification signal corresponding to the type of the target object. At least the weighting of the neural network in the neurocomputer is changed so as to minimize the squared error between and.

[0009]

[Action]

(1) Regarding the invention according to claim 1, means for observing an object with an image input device, processing the input image to extract a feature amount, and inputting the feature extraction amount to perform arithmetic processing to obtain the object. The pattern recognition device is composed of a neuro-computer section which outputs an identification signal indicating whether or not it is abnormal. Note that the neuro computer section is preliminarily trained with a small amount of feature amount data of the object. In this way, it is possible to recognize with a small number of learning data by utilizing the redundancy of the neuro computer operation.

(2) Concerning the invention according to claim 2, the learning of the neurocomputer unit is extracted from one target object, which is a standard selected for each type of object, through the feature amount extraction means. A set of feature quantities is input to this neurocomputer unit, and at this time, the identification signal actually output from this neurocomputer unit and the desired identification signal (output teacher signal) corresponding to the type of the target object are output. The weighting of the neural network in this neurocomputer is changed so as to minimize the power error. In this way, when there is no large variation in position / orientation and there is little variation in the object, it is possible to recognize the object only by learning the feature amount once for each type.

[0011]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a pattern recognition apparatus of the present invention, 1 is an object, 2 is an image input unit such as a television camera, 3 is an analog / digital conversion and preprocessing unit, and 4 is a feature extraction unit. 5 is a neuro computer, 6 is a determination unit, and 7 is a feature amount storage unit.

Next, the learning operation (of the neuro computer 5) of the apparatus shown in FIG. 1 will be described. After observing the standard object 1 with the image input unit 2 and converting the analog grayscale image signal 2a output from this input unit 2 by image scanning into the digital image signal 3a with the A / D conversion / preprocessing unit 3 ,
The feature extraction unit 4 measures the feature amount of the standard object from the digital image signal 3a. This feature quantity is stored in the feature quantity storage unit 7.
Remember. Next, the characteristic amounts of the standard objects 1 of other types are also sequentially measured and sequentially stored in the characteristic amount storage unit 7. After measuring the feature amounts of all types of standard objects (recognition targets) to be learned, one set of feature amounts for the standard objects in the storage unit 7 is input to the neurocomputer 5 for each type of object. , The weight of the neural network in the computer 5 so as to minimize the square error between the identification signal as the output of the computer 5 at this time and the desired identification signal (output teacher signal) corresponding to the type of the object. Appendix 4 and the threshold value are changed.

Here, the threshold value is a parameter representing the input / output characteristic of each neuron forming the neural network in the computer 5, and the weight is a parameter representing the strength of connection between neurons. In this way, all standard recognition objects are learned. Examples of the feature amount include area, width, height, perimeter, radius, number of holes, area of holes, number of corners, and the like.

Next, the recognition operation (of the neuro computer 5) of the apparatus of FIG. 1 will be described. The unknown object 1 is observed by the image input unit 2, converted into the digital signal 3a by the A / D conversion / preprocessing unit 3, and then the feature quantity a _{i is} obtained by the feature extraction unit 4.
(However, i = 1 to n) is measured. The feature amount is input to the learned neuro computer 5, and the identification signal output from the computer 5 is applied to the determination unit 6 for determination.

FIG. 2 shows an example of the structure of a neural network having a three-layer structure consisting of an input layer 11, a hidden layer 12 and an output layer 13 in the neurocomputer 5. Here, the circles in the figure are neurons that compose each layer.

[0016]

According to the present invention, the pattern recognition device is provided with a neuro computer, and a set of feature quantities extracted from one standard object selected for each type of object is input in advance to this neuro computer. Then, the neurocomputer performs learning so that the neurocomputer outputs a desired identification signal indicating the type of the object, and then the learned neurocomputer inputs the feature amount extracted from the unknown object. Then, since the type of the target object is discriminated from the output identification signal of the computer, the target object can be learned with a small amount of data, and the engineering labor is simplified. Then, at the time of recognition, it became stronger against fluctuation by lowering the threshold value. At the time of inspection, it is possible to check minute differences by raising the threshold value and tightening it.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a pattern recognition device as an embodiment of the present invention.

FIG. 2 is a diagram showing a configuration example of a main part of a neurocomputer.

[Explanation of symbols]

 1 Object 2 Image Input Section 3 A / D Conversion / Preprocessing Section 4 Feature Extraction Section 5 Neurocomputer 6 Judgment Section 7 Feature Quantity Storage Section 11 Input Layer 12 Hidden Layer 13 Output Layer

Claims (2)

[Claims] 1. An object is observed through an image input means, and the input image is processed to output a signal for identifying what the object is or whether the object is abnormal. A pattern recognition apparatus, comprising: a means for processing the input image to extract a feature amount; and a neurocomputer for inputting the extracted feature amount and performing arithmetic processing to output the identification signal. Recognition device. 2. The pattern recognition apparatus according to claim 1, wherein the learning of the neurocomputer is extracted from one target object, which is a standard selected for each type of object, through the feature amount extraction means. Another set of feature quantities is input to the neurocomputer, and at this time, the square error between the identification signal actually output from the neurocomputer and the desired identification signal corresponding to the type of the target object is minimized. As described above, the pattern recognition apparatus is characterized in that at least the weighting of the neural network in the neurocomputer is changed.