Research On Feature Extraction Method For Facial Image Based On Joint Encoding And Convolutional Network

Face recognition is an active research direction in the field of image processing and computer vision,through several decades of development,remarkable progress has been made.However,face recognition is still a challenging research topic because facial images are usually affected by gestures,illumination,occlusion and other variations in the shooting process,which reduces intra-class similarity and increases inter-class similarity simultaneously.Classical face recognition system mainly contains four parts: detection,alignment,feature extraction and classification,in which feature extraction plays the most important role in the whole recognition system.With face recognition as the application,and improving the discriminability and robustness of facial image features as research objective,this thesis proposes three feature extraction methods in three aspects: handcrafted designing,shallow learning and deep learning.The main content and novelty of this thesis are summarized as follows:(1)In the aspect of designing manual feature.We propose a handcrafted feature extraction approach called joint encoding of multi-direction line binary patterns(JEMDLBP).JEMDLBP consists of three components: direction line binary patterns(DLBP),dominate patterns statistic,multi-direction joint encoding.Unlike local binary patterns(LBP)that performs encoding in local square space,DLBP performs enconding in local line space.Dominate patterns statistic could adaptively select dominate DLBP patterns in different databases and different directions.Multi-direction joint encoding can capture correlation information between the DLBP patterns of various directions,thus the discriminative power is enhanced.Experimental results on four representative face recognition databases and one face expression recognition database show that the proposed JEMDLBP feature is superior to traditional LBP based features.(2)In the aspect of shallow learning feature.We propose a new model named discriminative probabilistic latent semantic analysis(DpLSA),which solves the traditional face feature extraction problem by transforming it into the document topic analysis problem,and provides a brand-new idea for face feature extraction.To tackle the shortage that the feature P(z|d)(topic-image distribution)learned by topic model p LSA based face recognition algorithm is meaningless,the proposed DpLSA exploits the structure information of each category of training images and presents a new word-topic distribution initialization,therefore a one-to-one correspondence between latent topics and facial image categories is created.As a result,the problem about how to select appropriate topic number in classical pLSA model is solved.The sensitivity of EM algorithm to model initialization is solved.And training speed of topic model is accelerated.Moreover,the learned feature P(z|d)of DpLSA is semantic and discriminative to perform recognition task directly.The experiment results on several facial image databases indicate that DpLSA has good robustness to illumination,occlusion,as well as pose.(3)In the aspect of deep learning feature.We propose a feature learning method called cascaded K-means convolutional feature learning(CKCFL).This method uses the idea of feature extraction from Convolutional Neural Network and the architecture of light deep neural network,it provides a good tradeoff between performance and efficiency.CKCFL conducts intensive study on the three core modules of deep neural network: filter learning,non-linear processing and pooling.In the filter learning layer,the well-known K-means algorithm is used to accelerate filters learning speed.In the non-linear transformation layer,the convolution features are transformed nonlinearly using the activation function,our empirical experience suggests that the recognition rate of Tanh function is better than that of other activation functions.In the pooling layer,spatial pyramid second-order pooling is adopted.On the one hand,it characterizes the correlation between different dimension features,overcoming the shortage of traditional first-order pooling that only extracts information of the features from the same dimension.On the other hand,it can simultaneously extract local and holistic features.The influence of different parameters and components of CKCFL on the recognition rate are discussed in detail.Experimental results on restricted and unrestricted facial image databases such as AR,Extended Yale B,FERET,LFW demonstrate the discriminative capability and robustness of CKCFL feature.