CN101329722A

CN101329722A - Human face recognition method for performing recognition algorithm based on neural network

Info

Publication number: CN101329722A
Application number: CNA2007100423839A
Authority: CN
Inventors: 杜凤
Original assignee: SHANGHAI BACKCOM INTELLIGENT TECHNOLOGY Co Ltd
Current assignee: SHANGHAI BACKCOM INTELLIGENT TECHNOLOGY Co Ltd
Priority date: 2007-06-21
Filing date: 2007-06-21
Publication date: 2008-12-24

Abstract

Aiming at overcoming the situation that the existing face recognition is difficult to fulfill mainly due to the characteristic that a human face is used as biometrics, the invention provides a face recognition method of a multi-characteristics recognition algorithm based on components. The method achieves the process accuracy of the face recognition method which is based on neural network to carry out recognition algorithm.

Description

The face identification method of the algorithm of discerning based on neural network

Technical field

Patent of the present invention relates to a kind of face identification method of the algorithm of discerning based on neural network, has especially realized a kind of accuracy of process of face identification method of the algorithm of discerning based on neural network.

Background technology

At present, in the ordinary course of things known, the advantage of recognition of face is its naturality and the characteristics of not discovered by tested individuality.So-called naturality is meant that the biological characteristic that is utilized when this recognition method is carried out individual identification with human (even other biological) is identical.For example recognition of face, human also by observing relatively people's face differentiation and confirming identity, the identification that has naturality in addition also has speech recognition, bodily form identification etc., and fingerprint recognition, iris recognition etc. do not have naturality, because the mankind or other biological are individual by this type of biological characteristic difference.The characteristics of not discovered are also very important for a kind of recognition methods, and this can make this recognition methods not offensive, and are not easy to be cheated because be not easy to arouse people's attention.Recognition of face has the characteristics of this respect, it utilizes visible light to obtain human face image information fully, and be different from fingerprint recognition or iris recognition, need utilize electronic pressure transmitter to gather fingerprint, perhaps utilize infrared ray to gather iris image, these special acquisition modes are easy to be discovered by the people, thereby more likely by the camouflage deception.Though recognition of face has the incomparable advantage of a lot of other identifications, also there are many difficulties in itself.The difficulty of recognition of face mainly is that people's face brings as the characteristics of biological characteristic.

Summary of the invention

In order to overcome the situation that is difficult to realize that existing recognition of face mainly brings as the characteristics of biological characteristic because of people's face, the invention provides a kind of face identification method of the many feature recognition algorithms based on parts, this invention has realized a kind of accuracy of face identification method process of the algorithm of discerning based on neural network.

The technical solution adopted for the present invention to solve the technical problems is: utilize computer science to handle and realize.

The invention has the beneficial effects as follows, realized a kind of accuracy of face identification method process of the algorithm of discerning based on neural network.

Description of drawings

The present invention is further described below in conjunction with drawings and Examples.

Fig. 1 is a schematic diagram of the present invention

Embodiment

In order to overcome the situation that is difficult to realize that existing recognition of face mainly brings as the characteristics of biological characteristic because of people's face, the invention provides a kind of face identification method of the many feature recognition algorithms based on parts, this invention has realized a kind of accuracy of face identification method process of the many feature recognition algorithms based on parts.Its key step comprises: modelling, realistic model are set up and emulation experiment.

Its key step comprises:

1, people's face detects: judge whether there is people's face in the input picture, if having, provide the position of everyone face, size;

2, facial Feature Localization:, detect information such as the position of its major organs and shape to everyone face that finds;

3, people's face comparison:,, judge the identity information of this people's face with people's face contrast in the storehouse according to the result of facial Feature Localization;

Claims

Based on the face identification method of many feature recognition algorithms of parts in order to overcome the situation that is difficult to realize that existing recognition of face mainly brings as the characteristics of biological characteristic because of people's face, the invention provides a kind of face identification method of the many feature recognition algorithms based on parts, this invention has realized a kind of accuracy of face identification method process of the many feature recognition algorithms based on parts, it is characterized in that:

Its key step comprises:

1, people's face detects: judge whether there is people's face in the input picture, if having, provide the position of everyone face, size;

2, facial Feature Localization:, detect information such as the position of its major organs and shape to everyone face that finds;

3, people's face comparison:,, judge the identity information of this people's face with people's face contrast in the storehouse according to the result of facial Feature Localization;
2. described in 1, the feature that people's face detects is:

The principle that people's face detects is: identification people face is mainly according to people's feature on the face, that is to say according to those and between Different Individual, exist than big-difference for the then more stable tolerance of same individual. because people's face changes complexity, therefore feature statement and feature extraction are very difficult. before facial image being carried out feature extraction and classifying, generally need do geometrical normalization and gray scale normalization. geometrical normalization be meant according to people's face positioning result with image in people's face transform to same position and onesize, gray scale normalization is meant image is carried out processing such as illumination compensation that illumination compensation can overcome the influence of illumination variation to a certain extent and improve discrimination.
3. described in 1, the feature of facial Feature Localization is:

The principle of facial Feature Localization is: face recognition technology is that a kind of facial characteristics according to the people carries out the advanced biological identification technology that identity is differentiated automatically.System obtains the most important characteristics of portrait face by camera, protruding part as people's face, distance, position, angle and sizes in the face profile such as brow ridge, eyes, nose and mouth calculate their geometric feature then, again with template base in portrait compare and confirm.This technology and tradition use I.D., password and even iris recognition, fingerprint, the scanning of palm shape to compare, characteristics such as have the accuracy of identification height, intuitive is outstanding, basic data obtains easily and cost is low.

Design considerations

One of: the input coding problem

Obviously, the input of ANN must be certain expression of image, and therefore adopting what kind of mode that image is encoded just becomes a big key of design.I attempt image is carried out pre-service, decomposite zone or other local image features of edge, brightness, unanimity, then these feature fan-in networks.Find in the practice that the problem that every images has the characteristic parameter of varying number has appearred in this design, ANN requires the quantity of input block to fix.Through exploring repeatedly, I become 30 * 32 fixing brightness values to image encoding, the corresponding network input of each pixel.And is brightness that 0 ~ 255 brightness value is scaled in 0 to 1 the interval, so that network input and hidden unit and output unit are in same interval value.Here 30 * 32nd, originally 120 * 128 low resolution is summarized, each low-resolution pixel has the mean value computation of corresponding with it some high-resolution pixel brightness to obtain.Use such low resolution image, the quantity of input number and weights has been reduced to a scale that is easier to handle, thereby reduced the computing requirement, but also kept the resolution that enough is used for correct classification image simultaneously.

Two: the output encoder problem.ANN must export in four values one represent people's face in the input imagery towards.I have used the encode classification of these four kinds of situations of n 1 (1-of-n) method of getting, that is: a kind of in four each corresponding four kinds of possibilities of different output units gets output with the high likelihood predicted value as network.Select n 1 method of getting to represent objective function for network bigger degree of freedom is provided;

This external n gets in 1 coding, and the difference between the highest output and inferior high value output can be used as the degree of confidence to the network prediction.The desired value of these four output units is decided to be:＜0.9,0.1,0.1,0.1 〉; Encode face forward＜0.1,0.1,0.1,0.9 〉; Encode face＜0.1,0.1 to the right, 0.9,0.1 〉; Encode face＜0.1,0.9 left, 0.1,0.1〉the coding face is backward.Avoid using 0 and 1 as desired value be because the sigmoid unit can not produce such output for limited weights.If the attempt training network mates desired value 0 and 1, gradient decline can force weights to increase without limitation.

Three: network structure.The most general a kind of network structure of back-propagation algorithm is a hierarchical network, and each unit of one deck is connected to down each unit of one deck forward.Such structure has also been selected in my design, has used two-layerly, comprises a layering feedforward neural network of hiding a layer and an output layer.

Four: other parameters of learning algorithm.In this experiment, learning rate η is set to 0.3, and momentum α is set to 0.3.Give the lower value of these two parameters and can produce extensive precision about the same, but need the longer training time.If these two values are set De Taigao, training can not converge to a network with acceptable error.I have used gradient decline completely in whole test.The grid weights of output unit are initialized to little random value.Yet the weights of input block are initialized to 0, because visual being easier to of the weights of learning understood, and to the not significantly influence of extensive precision.The selection of the iterations of training can be by cutting apart available data for training set and verifying that independently set realizes.The gradient descending method is used to minimize the error that training set closes, and every the performance of 50 subgradient decline iteration according to checking set assessment primary network.The final network of selecting is to the highest network of checking set precision.

Five: the hiding expression of learning.2899 weights that obtain in the phase-split network.Consider four rectangles under adjacent people's face image among the figure, each rectangle has been described in the network weights in four output units (encoded left, preceding, right, on).Four little square expressions, four weights related with this output unit in each rectangle---Far Left is power w., the threshold value of its decision unit; Be to connect three weights of three hidden units then to this output.Weights are represented in square brightness, and brilliant white is represented bigger positive weights, and the negative weights that black dull expression is bigger are represented medium weights between the gray shade of centre.
4. described in 1, the feature of people's face comparison is

The principle of people's face comparison is: by collection in worksite comparison facial image, compare with face database after extracting face characteristic automatically, be used for carrying out personal identification or similar crowd searches.The result of analyst's face key feature point, if only utilize the peak valley of gray-scale map integral projection to judge the Y coordinate position of unique point, not enough robust for diversified human face structure and picture quality, the individual cases meeting takes place than mistake.Therefore this paper makes full use of the vision priori of human face structure, to key feature point the time, added trained priori engineer's scale relation, promptly from the forehead to eyes, eyes to the nostril, the Y coordinate proportionate relationship of nostril to face and face to lower jaw, with the embedded system is platform, develop face recognition algorithms, the people's face that forms the off line application is on this basis discerned product automatically.