Research On The Detection Of Local Features Of Face In Complex Condition

With the development of artificial intelligence,more and more intelligent face detector products have entered people’s daily life,such as well-known face recognition attendance machines,face recognition access-control machines,intelligent face recognition authentication terminals.It can be said that it is on the course of the rapid development of face recognition products.Facial feature detection technology is one of the key technologies on the research and development of face recognition products.In most complex cases,the problem of lacking the information about the inherent characteristics of faces which are affected by many kinds of facial plastic and variable factors in the imaging process will directly affect the recognition performance of the product.Therefore the research of face feature detection under complex conditions has important practical significance.This thesis studies on the local facial feature detection technology under complex conditions,focusing on how to solve the problems that the target edge feature is missing and how to eliminate the influence on the feature extraction caused by the scale change and state change.The thesis makes a comprehensive review and systematic research on the above problems,mainly completing the following work:In the process of image edge detection,the existing smoothing denoising method generally has the problem of excessive denoising leading to the loss of important features.Therefore,the thesis puts forward a method of target gradient magnitude edge detection based on image fusion,The facial contour edge is taken as the object of study,Firstly,Mean-Shift algorithm is used to remove the noises of face images,and then the original image of the skin color is segmented by the color Gauss model,and then the method of fuzzy Gauss accelerates the emergence of skin regions and the generation of the mask which will make a fusion between the smooth image and the original image.The method retains the hair details of the image,enhances the edge details of facial contours and provides more effective information for face edge feature extraction.The thesis provides the method of facial feature detection through the Multi-scale Sampling Convolutional Neural Networks(MSCNN).As for the missing of partial important features when common Convolutional Neural Networks sampling in the fixed window,the thesis provides to construct the multi-scale sampling layer to replace the ordinary sampling layer in the network structure in order to realize the adaptive characteristic sampling in strain state.Firstly,the skin color is segmented by the Gauss mixture model in YCbCr color space so as to determine the potential target candidate window,and then the candidate image blocks which are in the different scales will be inputted to multi-scale sampling convolutional neural network.After a series of convolution and sampling operations,the system will form fixed characteristic expression,obtain multi-scale face information to improve the detection rate of multi-scale face image.Some important features will be missing when the face images change because of the change of gestures,facial expressions,occlusion and camouflage.While the current popular texture feature extraction operator has low accuracy and slow speed.Therefore,the thesis provides the algorithm of Parallel Multi-region Local Phase Quantization(PMLPQ)to solve it.In order to solve this problem,the researcher improves the method of using local phase quantization algorithm to extract the features,overcomes the problem that the change of states influences the features extraction,and finally gets the purpose of target feature stable acquisition.When the original face image is divided into several subregions of uniform size,the algorithm extract LPQ feature vector of each subregion in the method of parallel computing and then connects them together.Finally the Fisherfaces algorithm will reduce the dimensionality of the face feature extracted by the PMLPQ algorithm in order to achieve the stable condition of facial feature extraction.Finally,this thesis is integrated with image fusion method,face feature multi-scale sampling convolutional neural network method and face feature extraction method based on the PMLPQ algorithm and Fisherfaces.And it also designs and develops the platform of intelligent face recognition system.The test results show that the face feature detection method in this thesis is effective in practical application and it has very high robustness.