Research On Key Algorithms Of Face Alignment And Image Classification

With the epic development of modern information technology,biometric identification has become one of the hottest research directions in the field of pattern recognition.In many of these biometrics,human faces are closely related to our humanity and represent each of us in a certain sense.People identify friends and enemies by recognizing faces.Can they do this work through machines?Therefore,face recognition came into being,and because of its uniqueness,research on face recognition has never stopped in both industry and academia.Although face recognition research is growing vigorously,accurate and effective face recognition algorithms and systems are still rare.As we all know,in the natural environment,factors such as illumination,posture,and occlusion are unpredictable,and the face is a non-rigid shape,which has great influence on the accuracy of face recognition.From the 1960s to the 1970s,a large number of face recognition algorithms with important academic values were proposed,such as the Eigenface method proposed by Sirovich et al.,and the principal component analysis(PCA)proposed by Turk et al.,Fisherface and Linear discriminant analysis(LDA)proposed by Fisher et al.In the initial stage of face recognition research,these algorithms play a vital role in the future research.Since the 21st century,the rise of big data and artificial intelligence(AI)has brought new opportunities and challenges to the face recognition field.Face alignment,as the pre-stage of face recognition,is crucial for accurate recognition.The sparse representation theory is one of the key means to deal with the big data era.Based on the above ideas,this article has mainly done some work in these two aspects:(1)Based on the research of cascading regression model and random forest thoughts,a face alignment algorithm based on cascade regression model is proposed.Using the idea of random forest to build a two-level cascaded regression model,the local binary feature is used for face alignment,and the effect of the original algorithm is improved.(2)Through the study of sparse representation,we propose a new hash model,namely dictionary learning for hash representation(DLHR)?which further extends the framework of hash learning.Compared with the traditional hashing method,this method can achieve competitive classification and retrieval accuracy.This part of the work published in the CISP-BMEI 2017.(3)By summarizing sparse coding idea,this paper presents a large classification framework based on sparse representation:affine-constrained group sparse coding based on mixed norm(MNACGSC),and it provides a principled extension of the current ACGSC framework with multiple input samples and norms.A detailed explanation of the optimization scheme is given and some experimental evaluations of the image classification are made.This part of the work has been published on ICONIP 2017.This paper verifies the accuracy of face alignment on LFPW,HELEN and 300-W face databases.Image classification experiments were conducted in AR face database,E-Yale B face database and MNIST handwriting database.Experimental results show that the proposed model and algorithm not only improve the recognition rate,but also improve the efficiency.