Research On Sketch Face Synthesis And Recognition

Face recognition as one of the important technologies of rapid identity identification,has been widely used in the fields of social security and information security.The traditional face recognition is the matching of the face photos taken by video device to the photos in the database.But in criminal cases,there are many situations in which the face photos are unavailability.Then,It need to depend on the suspect sketch face which is drawn by forensic sketch artist according to the information provided by eyewitnesses and look up in police photo library.Because of the large difference between sketch and photo,the traditional face recognition method cannot get the ideal result.The artificial search is time-consuming and laborious,and the result is not good.So the research on sketch face synthesis and recognition is very important.This technology is not only widely used in law enforcement,but also in the field of digital entertainment and cartoon production.This paper mainly studies the sketch face synthesis and the sketch face recognition.Two sketch face synthetic methods are proposed.One is the cascade sketch face synthesis method based on Dual-scale Markov Network.Firstly,large-scale Markov Network is adopted to get the initial sketch image.Then the initial sketch with the test and training images are entered to the small-scale Markov Network to synthesize the final sketch.This method also has reasonably extended the search area of the neighbor patches and proposed LSNM neighbor search method.It successfully solves the problems of synthesis method based on single scale sketch such as the Mosaic effect,law enforcement,and gets the higher quality sketch images.The second is the sketch face synthesis method based on Gradient Fully Convolutional Network.This method applies the face sketch synthesis to the field of deep learning and uses the Fully Convolutional Network for sketch face synthesis.What's more,the gradient prior information is added in the network to obtain the complete structure details of the synthesized sketch image in the eyes,nose,mouth and other parts.To the recognition of sketch face,this paper mainly extracts the features with the low sensitivity of texture and use a variety of information fusion strategy.There are three sketch face recognition method are proposed.One is the sketch face recognition method based on Multi-HOG feature fusion.The global HOG feature and local HOG feature are extracted from the sketch and photo respectively,then the PCA and S-NLDA methods are used to reduce the feature dimension.Feature level fusion is used to obtain the multi-HOG feature and identify the sketch face.This method combines the global feature and local feature,which is a good representation of the whole face contour,and fully utilizes the local detail characteristics of the face.The second is the Multi-PHOG and sensitivity fusion based sketch face recognition method.The blocked mode based on the scale pyramid is used to instead of the fixed method to extract HOG feature and get the PHOG feature.Through the sensitivity method,we calculate the weights of different features.Finally,the global PHOG feature and local PHOG feature are fused with weights and complete the recognition.These processes can improve the applicability of the proposed method.The third is the sketch face recognition method based on the fusion of structural information and feature information.The method uses for reference of face recognition cognitive theory which mainly relies on the structural information and feature information for face recognition.We extract the HOG based structure information and sparse kernel prototype representation based feature information for sketch face recognition.This method considers the structural correlation and independent distribution of the human face,so improves the accuracy of the sketch face recognition.In addition,the paper also adds two auxiliary attribute information,including race information and gender information to the sketch face recognition process.These information can filter face images and improve the recognition rate.