Research On Face Recognition Based On Binary Hash Feature Learning In Unrestricted Environments

As one of the most promising key technologies in biometric recognition,face recognition technology has received continuous attention from researchers in the fields of artificial intelligence and computer vision.At present,face recognition technology has achieved desirable performance in restricted environment,and many technologies have been used in practical application.However,there are still many unresolved issues in unrestricted environment,such as single extreme intra-class variation,lack of training samples,inconsistent image modalities,and complex and diverse intra-class variations,which cause the performance of existing face recognition technologies to be far from practical requirements.To improve the discriminative ability,robustness and accuracy of the face recognition technology,this dissertation mainly focuses on the single extreme intra-class variation,the extreme shortage of training samples in face recognition with single sample per person(SSPP),the cross-modal problem in heterogeneous face recognition,and the large-scale face retrieval problem in complex scenes.The main research contents and contributions of this dissertation are as follows:(1)Aiming at the problem of single extreme intra-class variation in unrestricted environment,a local feature hashing with binary auto-encoder algorithm is proposed.Existing shallow real-valued feature learning algorithms usually lack robustness,and shallow binary hash feature learning algorithms have insufficient discriminative ability and generate accumulated quantization errors in the optimization process.In order to solve these problems,a Local Feature Hashing with Binary Auto-encoder(LFH-BAE)algorithm is proposed.The algorithm introduces a binary auto-encoder into the hash learning process to extract structural information with reconstruction capability from the initial features,so that the robustness and discriminative ability of the algorithm can be strengthened.To break the combinatorial complexity of the original nesting problem of the binary auto-encoder,an auxiliary coordinate is introduced to decompose the original problem into several separate and tractable sub-problems.An iterative optimization algorithm based on the Augmented Lagrangian method is proposed to optimize these sub-problems.It not only considers the discrete nature of binary codes,but also takes the bit-balance constraint and bit-independence constraint into account,effectively avoiding the accumulated quantization errors and improving the quality of the binary feature.Extensive experimental results on the public datasets FERET,CAS-PEAL-R1,LFW and PaSC demonstrate that the proposed algorithm can effectively deal with the single extreme intra-class variation problem,and achieves superiority recognition performance to existing algorithms.For example,compared with similar algorithms,the minimum increase in the verification rate(at FAR=0.01)of the LFHBAE algorithm on the PaSC’s All image set and Frontal image set is 2.1%and 3.2%,respectively.(2)Aiming at the extreme shortage of training samples in SSPP face recognition in unrestricted environment,a multi-scale fusion collaborative representation based on semantic-aware local hashing feature model is proposed.Existing SSPP face recognition algorithms have the following shortcomings:insufficient discriminative ability,lack of descriptive power and single feature scale.This dissertation proposes a Multi-scale Fusion Collaborative Representation based on Semantic-aware Local Hashing Feature(MsFCR-SLHF)model,which remedies the defects in the perspectives of feature extraction and classifier design.In the feature extraction stage,firstly,an asymmetric graph regularized binary autoencoder is introduced to extract structure information and semantic information from the general training data,improving the discriminative ability of the model.Secondly,by performing multi-scale clustering and pooling on the learned binary codes,multiscale SLHF features that contain contour information and detail information are obtained.Thus,single feature scale problem can be successfully solved.In the feature matching stage,a multi-scale fusion collaborative representation classifier is designed.The classifier performs weighted fusion of SLHF features at different scales,and requires local SLHF features of the same scale to collaborate with each other,which enhances the model’s descriptive power and discriminative ability.Extensive experimental results on the popular SSPP datasets AR,Extended YaleB,CMU-PIE,and LFW show that the proposed model effectively compensates for the deficiencies of the existing algorithms and achieves excellent performance in SSPP face recognition tasks.For example,compared with the existing optimal algorithm,the MsFCR-SLHF model improves the recognition accuracy by 13.01%on the LFW.(3)Aiming at the cross-modal problem in heterogeneous face recognition in unrestricted environment,a Pseudo Supervised Hash Coding based on Integrated Angular Reconstruction algorithm is proposed.Existing heterogeneous face recognition algorithms have large semantic "gaps"between modalities,insufficient preservation of similarity information within modalities,and inducing accumulated quantization errors in the optimization process.A Pseudo Supervised Hash Coding based on Integrated Angular Reconstruction(PSHC-IAR)algorithm is proposed to address these issues.The algorithm introduces the integrated angle reconstruction term,non-negative spectrum clustering term,pseudo-label consistency term,and integrated quantization error term into the same objective function.By joint learning of common binary codes,pseudo-label,and coupled modal hashing maps,the accuracy of heterogeneous face recognition is increased.On the one hand,the PSHC-IAR algorithm uses integrated angle reconstruction embeddings to allocate the same binary code to the initial features of the same identity in each modal,which solves the problem of insufficient preservation of similarity information in the modal.Meanwhile,the same binary code is assigned to heterogeneous initial features derived from the same identity between modalities,which bridges the semantic "gap"between modalities.On the other hand,the PSHC-IAR algorithm combines nonnegative spectrum clustering analysis and pseudo-label consistency to extract potential relevant information from heterogeneous data and use it as pseudo-label to guide binary code learning,which further reduces semantic "gap" between modalities.In addition,an efficient iterative optimization algorithm,named as Augmented Lagrangian method based on principle coordinate updating,is proposed to optimize the objective function.It not only guarantees the compatibility among the common binary code,pseudo-label and coupled hashing maps,but also effectively avoids accumulated quantization errors.Extensive experimental results on the popular heterogeneous datasets HFB,CASIA NIR-VIS 2.0,CUFSF and Multi-PIE demonstrate that the proposed algorithm effectively solves the problems of existing algorithms and improves the accuracy of heterogeneous face recognition.For example,compared with existing optimal algorithm,the PSHC-IAR algorithm increases the recognition rate by 10.7%on the CUFSF.(4)Aiming at the problem of large-scale face retrieval in complex scenarios,a deep supervised hashing based on adaptive anchor graph updating model is designed.Existing deep hashing algorithms have the following three issues:low utilization of semantic information,poor scalability,and absence of useful constraints.A Deep Supervised Hashing based on Adaptive Anchor Graph Updating(DSH-AAGU)model is designed to address these issues.The model uses a pairwise supervised matrix and an adaptive anchor graph to integrate deep feature learning and hash learning into the same end-to-end framework,and jointly learns deep hash functions and binary hash codes.The DSH-AAGU model uses pairwise supervised matrix to guide deep hash function learning and hash code learning asymmetrically,extracting pairwise similarity information.Meanwhile,it uses an adaptive anchor graph to guide hash code learning and retains semantic relevant information of the anchor graph.The mutual cooperation of the two matrix not only effectively solves the problem of insufficient utilization of semantic information,but also ensures the scalability of the model.In addition,the DSH-AAGU model also incorporate the bit-balance term and the bit-independence term into the objective function,and employs an iterative optimization algorithm based on self-adjusting discrete proximal linearized minimization to optimize it.This optimization algorithm not only improves the quality of binary hash codes and deep hash functions,but also promote the training efficiency of the model.Extensive experimental results on two popular face datasets YouTube and UMDFaces as well as one general image set CIFAR-10 show that the proposed model not only has good scalability,but also can make full use of the semantic information of the database.The DSH-AAGU model outperforms those existing deep hashing algorithms in both training efficiency and retrieval performance.