JPH02239377A - Individual collating device - Google Patents

Abstract

PURPOSE:To improve collating accuracy with the use of a simple processing by executing the sum of product operation between an output at every picture element of an input picture and a weight concerning the output, and a threshold processing operation for the output of the sum of product operation by means of a neural network part. CONSTITUTION:A neural network circuit 6 executes the sum of product operation between the output at every picture element to have image-picked-up of an indefinite individual and the weight concerning the output, and a threshold processing for the result. The difference between an output Yj of the circuit 6 and a desirable value Yj0 corresponding to the output Yj is obtained, and the output Yj is fed back to the circuit 6. A deciding part 7 compares the output Yj of the circuit 6 with a prescribed value K concerning the registered individual, when only one Yj, whose value is >=K, exists, the individual is identified as the registered one. When the Yj is <K or >= the two individuals exist, the individual is decided not to conform to the registered one. Thus based on the highly reliable physical information of the individual, the collation accuracy is improved by the simple processing.

Description

[Detailed description of the invention] [Industrial application field]

The present invention relates to a personal verification device that matches unspecified individuals with registered individuals for use in crime prevention systems such as room entry/exit control.

[Conventional technology]

? The need for person verification arises, for example, when controlling access to confidential areas such as computer rooms and data rooms. This personal verification for room exit management may be performed by a person such as a security guard, or by an unmanned personal verification device. The main types of conventional personal identification devices are: (1) methods that use magnetic cards, IC cards, personal identification numbers, etc., and (2) methods that are based on physical information unique to individuals such as fingerprints, palm shapes, retinas, and handwriting.

[Problem to be solved by the invention]

The conventional techniques as described above have the following drawbacks or problems. ■With methods such as magnetic cards,
Theft or loss. It is useless if you don't have a cell phone or forget it. In other words, even if an individual is not registered, if he or she carries the magnetic card of a registered individual or knows the personal identification number, the individual will be determined to be a registered individual as a result of verification. Also,■
Although the method based on physical information does not have the problem as in (2), it requires complex processing for feature extraction, and the matching accuracy is hampered by positional deviation of the physical information, M sound, etc. The object of this invention is to solve the problems of the conventional technology, to be based on personal physical information, and to be easy to process.
Another object of the present invention is to provide a personal verification device with high verification accuracy.

[Means to solve the problem]

In order to solve this problem, the personal verification device according to the present invention is a device for matching an unspecified individual with a registered individual, and includes: an image input section that captures an image of the unspecified individual; a product-sum operation of an output and a weight related to this output;
a neural net unit that performs threshold processing on the result of the product-sum operation; a determination unit that outputs the result of the matching based on the output of the neural net unit and a predetermined value related to the registered individual; The weight value is updated by learning based on a comparison between the output of the neural net unit and a corresponding desired value.

[For use]

The neural net unit performs a product-sum calculation of the output for each pixel of the image input unit that images an unspecified individual and weights related to this output, and threshold processing for the result of this product-sum calculation. Based on the output and the predetermined value related to the registered individual, the determination section outputs a verification result. Moreover, the weight values are updated by learning based on a comparison between the output of the neural net section and the corresponding desired value.

【Example】

Embodiments of the personal verification device according to the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of this embodiment. In Fig. 1, 1 is an image input unit, for example, a TV camera, 2 is an A/D converter, and 3 is a preprocessing unit, which performs smoothing and image enhancement to correct blurred parts of the image, and then performs binary value processing. become 4 is a neural net section, which mainly consists of a frame memory 5 and a neural net circuit 6. The frame memory 5 stores image data for each pixel from the preprocessing unit 3, and the neural net circuit 6 realizes truly "soft" information processing using the human brain as a model, as will be described in detail later. Various calculations are performed based on image data. Reference numeral 7 denotes a determination unit, which outputs the output Yj of the neural net circuit 6.
is compared with a predetermined value K related to the registered individual, and Yj that is greater than or equal to K is determined.
When only one exists, the unspecified individual is identified as a registered individual according to K, and when Yj is less than K, or
If two or more of the above Yj exist, it is determined that the unspecified individual does not match the registered individual. Now, for the output Yj of the neural net circuit 6, on the other hand, the difference with the desired value Yjo (teacher signal) corresponding to the name, Yjo - Yj, is taken, and this is fed back to the neural net circuit 6, which will be described in detail later. The constant value related to the calculation of the neural net circuit 6 is updated so as to reduce the previous difference. This series of update processing corresponds to "learning". Note that Yjo is a value such that the output of the determination unit 7 indicates an accurate matching result. FIG. 2 is a schematic diagram showing the neural net circuit 6 in FIG. 1 and its operation. The neural net circuit 6 has a hierarchical structure consisting of an input layer H, an intermediate layer I, and an output NJ. Each layer consists of a plurality of units indicated by circles. Each unit in the input layer H receives corresponding pixel data from the frame memory 5, and each unit in the output layer J corresponds to the maximum number of people that can be registered. Each unit connects from the input layer H to the output layer J via the intermediate layer ■.
Connections within each layer and input layer H from the output layer J to the intermediate layer ■
There is no bond that goes towards . Input layer H. In each unit of the intermediate NI and output layer J, the following calculations are performed. Now, the input layer H. The output of each unit in the middle layer ■ and the output layer J is Yh, Yll Y
Let Wih be the strength (weight) of the connection between the corresponding units of the input layer H and the intermediate layer I, and Wji be the weight of the connection between the corresponding units of the intermediate NI and the output layer J. Then, Yi = 1 / (1 +exp (1ΣYh - Wih))Yj = 1 / (
1 +exp(1ΣYi−Wji)) Here, ΣYh
-Wih, ΣYi -Wji is a product-sum operation, and the operation for obtaining Yi and Yj is a kind of threshold processing using a sigmoid function. In addition, output Yj from all units of output iJ,
The difference from the corresponding desired value Yjo, Yjo
−Yj, the weight W
The values of ih and Wji are updated. Therefore, Yi,
Yj. Both Wih and Wji are functions of the number of weight updates, in other words, the number of learning times.

【Effect of the invention】

As explained above, in the present invention, the neural net section performs a product-sum operation of the output for each pixel of the image input section that images an unspecified individual, the weights related to this output, and the result of the product-sum operation. Based on this output and the predetermined value related to the registered individual, the judgment section outputs the verification result, and also the output of the neural net section and the corresponding desired value. The weight values are updated by learning based on comparisons with Therefore, the present invention has excellent effects compared to conventional techniques in that it is based on highly reliable personal physical information, is easy to process, and improves matching accuracy through learning.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of essential parts of an embodiment according to the present invention, and FIG. 2 is a schematic diagram showing a neural net circuit and its operation. Code explanation H2 input layer, ■: Middle layer, J: Output layer, 1: Image input section, 2: A/D converter, 3s preprocessing section, 4: Neural net section, 5: Frame memory, 6: Neural -Net circuit, 7: Judgment section. Figure 1

Claims (1)

[Claims] 1) In a device for comparing an unspecified individual with a registered individual, an image input unit that images the unspecified individual; a product-sum calculation of an output for each pixel of this image input unit and a weight related to this output; a neural net unit that performs threshold processing on the result of the product-sum calculation; and a determination unit that outputs the result of the matching based on the output of the neural net unit and a predetermined value related to the registered individual; A personal verification device, wherein the weight value is updated by learning based on a comparison between the output of the neural net unit and a corresponding desired value.