A Study Of Face Recognition Based On Local Feature Extraction

As a famous biological recognition skill,face recognition takes the advantages of its friendly acquisition way,the wide applications and the abundantly potential data,compared with other methods such as fingerprint identification,iris recognition and so on.As information age is coming,the technologies of computer and network are developing increasingly and information is becoming more and more important,higher requirements upon information processing related to face recognition are put forward.In addition,due to the development of the hardware and software technology,it is possible that the face detection,tracking and recognition in practical application system are real-time.The research of face recognition becomes white-hot.As we all know,the occlusion,illumination,noise together with the inner factors of facial expression and posture will cause great disturbance for the recognition task.It is a difficult but urgent task to build a powerful description of these factors.The local feature description is found to be easier to weaken or eliminate the constraints of the influence factors when compared with global character description.The two-dimensional discrete wavelet transform(2D-DWT),Gabor wavelet transform,local binary pattern(LBP)and convolutional neural network are commonly used operators with the ability of local details' description.The purpose of this thesis is to extract more effective local features.The research methods for face recognition are patch based description and fusion based description.The main contribution of this dissertation is summarized as follows:(1)Propose the face recognition method based on the non-uniform patch strategy and 2D-DWT(NUPDWT).The key point of the method is the non-uniform patch strategy which depends on the physical characteristics of the 2D-DWT's sub-bands and the integral projection technique.Based on the average image of all training samples,every face image can be partitioned with regional distinction.Compared with the manual partition and uniform partition strategies,the non-uniform patch strategy is more automatic and better for retaining the integrity of local information,it is more suitable to reflect the structure feature of the face image.Combined the patch strategy with the extracted feature of 2D-DWT,the classifier of the nearest neighbor and majority voting,the new method NUPDWT is introduced.Experiments are run on many face databases including the occlusion sub-database of AR.The obtained numerical results exhibit that NUPDWT outperforms the traditional 2D-DWT method and some state-of-the-art patch-based methods.Especially,it is quite suitable for the problem of occlusion.(2)Propose the face recognition method based on the self-adapting patch strategy(SAPFR).Along with the research of patch strategy,a new patch strategy is introduced which shows images distinction as well as regional distinction.The partition is according to each face image's structure instead of the average image.Compared with non-uniform patch strategy,it is more suitable to recognize images taken in the unconstrained environment.Based on the edge detected by 2D-DWT,the self-adapting patch strategy implements on the original image.Combined with the feature extraction method LBP and a new classifier based on the idea of sparse,the corresponding face recognition method SAPFR is proposed.The experiments are carried out on Georgia and LFW face databases.SAPFR is proved to be better than the other patch based or local feature based methods,which owes to the reasonable patch strategy and the effective combination with LBP and the classifier.(3)Propose the new LBP-Like Feature Based on Gabor Wavelets(GLLBP).Based on the acknowledge that the resulted sub-images of Gabor wavelet transform are related to each other,a novel feature concept is proposed through using the idea of LBP to do the analysis.GLLBP retains the advantages of LBP and Gabor wavelet transform in illumination and noise.Moreover,the information among the resulted sub-images is fully used.Furthermore,its extensions and corresponding algorithms are given.The experiments are carried out on the ORL,FERET,Georgia and LFW face databases.The numerical experimental results demonstrate that the proposed algorithm based on GLLBP feature possesses higher recognition rates than some other popular methods at present,which proves that the new feature extracted from Gabor resulted sub-images by LBP is more discriminative.(4)Propose three fusion units for convolutional neural network models,including nonlinear competitive unit(NCU),multi-feature fusion unit(MFFU)and multi-decision fusion unit(MDFU).NCU can strengthen feature propagation by comparing the elements from different network layers and selecting the larger signals element-wisely.MFFU and MDFU extract the feature or decision information from some network layers and fuse them as results of merger or addition,respectively.The experimental results about face verification task and visual classification task demonstrate that all these units can efficiently promote the performance of the base models.It is also proved that the fusion information is more discriminative which shows the advantage and potential of fusion strategy in networks.