KR100567765B1 - System and Method for face recognition using light and preprocess - Google Patents

Abstract

The present invention discloses a face recognition method in a face recognition system.

The face recognition method through the analysis and preprocessing of the lighting environment of the present invention comprises a first step of determining an illumination state for the face image photographed using an evolutionary neural network; A second step of performing preprocessing on the face image using an evolution algorithm according to the determination result; And performing a face recognition through a third step of registering and recognizing the pre-processed face image, thereby maintaining stable recognition performance. Furthermore, face recognition can be applied in real time by performing evolutionary neural network learning in advance.

Face Recognition, Evolutionary Neural Networks, Evolutionary Algorithms, Lighting Templates

Description

System and Method for face recognition using light and preprocess             

1 is a flowchart illustrating a face recognition method according to the present invention in a face recognition system.

2 shows a direct encoding according to the invention,

3 shows an illumination template according to the invention,

4 is a view showing an embodiment of an image generated as a virtual image according to the present invention;

5 is a view for explaining the configuration of the entire system according to an embodiment of the present invention.

The present invention relates to a face recognition method, and more particularly, to determine a lighting state of a face to be recognized in a system based on image analysis using an evolutionary neural network, and to find a corresponding preprocessing using an evolutionary algorithm. The present invention relates to a face recognition method that maintains the face recognition performance stably by enabling an appropriate response.

In general, face recognition technology is applied to verify identity such as voice recognition, fingerprint recognition, and eye recognition, and it can be implemented at a relatively low price compared to other identification verification technologies, and recognizes special behaviors for recognition. There is no need to take it, which can prevent the objection of the object. In particular, in recent years, as the multimedia PC (PC) is popularized, the image acquisition device is basically installed, so that a face recognition technology application is possible without a separate hardware configuration.

Conventionally, a method of recognizing a face from an input image uses many of the following methods.

First, the principal component analysis methodology using EigenFace extracts the principal component of the face, i.e. the eigenface, through several faces, and extracts the weighted vector through the dot product of the extracted eigenface and the newly input face. Recognition may be performed by comparing the extracted weight vector with similarity with a previously registered weight vector.

Second, the methodology using the elastic graph matching method uses the elastic graph matching algorithm for the face detection and recognition process based on the preprocessed image. The algorithm extracts feature points from the input image, connects the extracted feature points to each other using an elastic graph, and extracts feature vectors through the Gabor function to perform recognition.

Thirdly, the methodology using local feature analysis (LFA) is derived from the eigenface method, and it is proposed to solve the problems caused by the change of facial expression, lighting, and posture, which are problems in the eigenface method.

Fourth, the methodology using correlation forms a mosaic within a certain range of the face region obtained after acquiring a face region based on a unique face from a face image input from a camera.

Finally, methodologies using neural networks are generally recognizable by learning about abstract data without any other statistical or mathematical development. However, the neural network has a disadvantage in that the convergence speed is very slow for learning data having a certain size or more, and in the case of a commonly used MLP (Multi layer perceptron), the entire pattern must be re-learned to increase the learning pattern. It is also possible to fall into local minimums and not find the optimal solution.

All of the conventional face recognition techniques have an issue of not being successful in face recognition when the lighting environment is severely changed due to the appearance-based model.

Accordingly, an object of the present invention for solving the above problems is to determine the lighting state of the face to be recognized by using an evolutionary neural network, and to find a corresponding preprocessing using an evolutionary algorithm, so that an appropriate response to the actual environment is possible. It is to stably maintain face recognition performance.

Face recognition method through the illumination environment analysis and pre-processing of the present invention for achieving the above object, in the lighting environment monitoring module, a virtual face image made by matching a face image captured by a CCD camera or the like and a predefined illumination template A first step of determining an illumination state of the photographed face image by learning an evolutionary neural network using; According to the determination result, the filter bank evolution module finds a combination of image filters for minimizing the influence of the determined lighting by using an evolution algorithm, and then cancels the influence of lighting on the face image using the image filter. A second step of performing preprocessing to generate the minimized image; And a third step of registering the pre-processed face image in a database in the registration and recognition module and performing face recognition by comparing the face image with previously registered face images.

Hereinafter, with reference to the accompanying drawings will be described in detail a preferred embodiment of the present invention.

1 is a flowchart illustrating a face recognition method according to the present invention in a face recognition system.

The face image photographed by the CCD camera or the like may be recognized as a different face according to the position of light or illumination at the time of imaging, so the influence of illumination on the face image should be minimized.

To this end, the face recognition system first determines an illumination state of the face image as to how much the current face image is affected by illumination (step 101).

The face recognition system directly encodes a node connection structure as shown in FIG. 2 using an evolutionary neural network based on an error backpropagation neural network to determine an illumination state.

In general, the back propagation neural network takes a long time to learn and falls into a local minimum, but the evolutionary neural network of the present invention can avoid the disadvantages of the back propagation neural network by applying the advantages of global search.

The learning of the evolutionary neural network is performed by using the virtual face image as shown in FIG. 4 by matching the photographed face image with the illumination template as shown in FIG. 3. At this time, the illumination template is made based on the value after taking the overall average for the face image under normal normal lighting.

In order to understand the lighting state using the evolutionary neural network, the template of FIG. 3 is generated by applying sine and cosine weights to the addition and multiplication miscellaneous images for the blank image. The template on the bright side is a negative ghost, and the template on the dark side is a positive ghost. When this template is applied to a general front face in a normal lighting state, a virtual image as shown in FIG. 4 is generated, and 100 or more such images are generated for each template and used as learning data of an evolutionary neural network. Passing the image through the evolved neural network thus learned makes it possible to determine the lighting state of the image based on the pre-learned result.

4 is added to the face in the normal illumination state using the template of FIG. 3, a virtual image as shown in FIG. 4 appears. The values in the template include both positive and negative numbers, and if they are positive, they are converted to a dark image and to negative values.

When the face recognition system completes the determination of the lighting state of the face image, the face recognition system uses an evolutionary algorithm to find a combination of image filters for performing different preprocessing according to various lighting states, thereby canceling the influence of the lighting on the face image (step) 102).

To this end, the present invention uses the weight value as a chromosome of the evolutionary algorithm using Equation 1.

l = [log2 ([Umin, Umax] / π + 1)] + 1

i _valreal = (i _valbin ) 10 × π + Umin

In the evolutionary algorithm, the goodness-of-fit calculation for the filter combination is represented by Equation 2 in each defined lighting environment.

Here, T is the result of passing the filter for the registered image,

          OrgT is the registered video with the same ID as T,

          NotT is a registered video with a different ID from T,

           k is the total number of registered images,

           m represents the number of images registered per person.

One filter chromosome exists per predefined environment and the format is shown in Table 1.

One  Whether to use a homomorphic filter 2  Use of Illumination Filter 3  Whether to use Histogram filter 4-10  Homomorphic filter parameters.

One 2 3 4 5 6 7 8 9 10     use / not use homomorphic

The number of defined environments is defined as nine as shown in FIG. 3, and each filter combination is derived after adding artificially defined lighting for face data having the same lighting state as its own.

The combination of filters used will cancel out most of the miscellaneous artifacts in real life if you reverse them. The three filters used in Table 1 each have a role to offset their effects on bright or dark lighting, deflection lighting, and so on.

The evolutionary algorithm is used to find an optimal method of applying them. After applying a filter to the images to which the template of FIG. 3 is applied, the difference is derived by comparing with the original image of the corresponding image, and the most difference is obtained through evolution. You will find fewer filter combinations.

This combination can be seen from the chromosomes for the evolutionary algorithms 1 through 10 shown in Table 1.

For example, the chromosome results that appeared after sufficient evolution

One 0 One 0 0 0 One 0 One 0

 In this case, the homomorphic filter and the histogram filter are used in sequence, and the parameter of the homomorphic filter is applied to 1010 (2).

When the process of finding the filter combination using the evolutionary algorithm is completed, the face recognition system uses the combined filter to cancel out the effect of the lighting, thereby generating an image with the least impact of the lighting.

The face recognition system registers the generated face image in the database and compares the face image with the registered faces and performs face recognition (step 103).

In other words, the process of finding the time-consuming evolutionary neural network and filter combination is performed in advance, and the face registration and recognition is performed in real time.

At this time, the recognition of the face is performed by using Gabor transformation after extracting the face region from the registered face image.

5 is a view for explaining a face recognition system using the above-described lighting environment configuration. Referring to FIG. 5, the entire system may include an illumination environment monitoring module and a filter bank evolution module. The lighting environment monitoring module uses a configured environment, and the filter bank evolution module performs a function of configuring an environment. The filter bank evolution module constructs a chromosome for the environment using an evolution algorithm. Accordingly, the entire system registers and recognizes an illumination environment monitoring module for determining an illumination state of the captured face, a filter bank evolution module for preprocessing the face image according to the determined result, and registering and recognizing the preprocessed face image. It can be implemented by including a registration and recognition module. Therefore, the system proposed in the present invention may be referred to as a dual system structure that performs the process of finding the evolutionary neural network learning and filter combination in advance and performs registration and recognition in real time.

As described above, the face recognition method through the analysis and preprocessing of the lighting environment of the present invention estimates the lighting state of the face image and generates an image that minimizes the influence of the light through preprocessing, and then uses the face recognition to detect the face. By doing so, stable recognition performance can be maintained. Furthermore, face recognition can be applied in real time by performing evolutionary neural network learning in advance.

Claims (5)

The lighting environment monitoring module is configured to determine an illumination state of the captured face image by learning an evolutionary neural network using a virtual face image made by matching a face image captured by a CCD camera or the like with a predefined lighting template. Stage 1; According to the determination result, the filter bank evolution module finds a combination of image filters for minimizing the influence of the determined lighting by using an evolution algorithm, and then cancels the influence of lighting on the face image using the image filter. A second step of performing preprocessing to generate the minimized image; And A third step of registering the preprocessed face image in a database in a registration and recognition module and performing face recognition by comparing the face image with previously registered face images; Face recognition method through the lighting environment analysis and pre-processing comprising a. delete The method of claim 1, wherein the lighting template A face recognition method based on an illumination environment analysis and pre-processing, characterized in that the overall average is taken for a face image under normal lighting and then made in advance based on the value. delete An illumination environment monitoring module for determining an illumination state of the captured face image by learning an evolutionary neural network using a virtual face image made by matching a face image captured by a CCD camera and a predefined lighting template; After finding a combination of image filters for minimizing the influence of the lighting according to the determination result of the lighting environment monitoring module, and using the image filter to cancel the effect of the light in the face image to generate an image with a minimal impact of the illumination A filter bank evolution module for performing the; And A registration and recognition module that registers a face image preprocessed by the filter bank evolution module in a database and compares the face image with a previously registered face image to perform face recognition; Face recognition system through the lighting environment analysis and pre-processing, characterized in that comprises a.

Images