Research On Single Sample Face Recognition Algorithm With Generic Dataset

Facial recognition is a biometric identification technology that identifies individuals based on their facial features.Related research involves multiple fields,including pattern recognition,image processing and analysis,computer vision,neural computing,and cognitive psychology.As early as the 1970 s,scholars began developing facial recognition algorithms.In the following decades,facial recognition technology made significant progress and is now widely applied in various areas of daily life.Although significant progress has been made in controlled facial recognition technology,facial recognition remains a challenging research problem under non-controlled conditions,due to factors such as illumination,facial expression,pose,and occlusion.In practical applications such as access control,criminal tracking,and video surveillance,only one training sample can be obtained,which is called the single sample per person(SSPP)problem.When dealing with the SSPP problem,as each class only has one training sample,the data dimension is much higher than the number of training samples.As a result,the performance of traditional classification methods may rapidly deteriorate.Meanwhile,if the classification method cannot observe the underlying variations of the gallery samples,such as pose,illumination,expression,and occlusion,the performance of the classification method will degrade when classifying the probe samples with variations.The above issues are studied in this paper.To obtain face variation information,standard practice is to collect an additional generic dataset,in which each class has multiple face images that cover the most predictable variation types.The work of this paper is based on the generic dataset.The research results are mainly reflected in the following aspects:(1)A shared generative adversarial network(Shared GAN)is proposed to generate samples with variations for the gallery samples.Due to the limited number of paired images in the generic dataset,the general image translation model cannot well generate samples with variations.Therefore,in order to improve the quality of the output image of the image translation model,Shared GAN combines the image translation model with the image generative model in the form of sharing decoding network in the generator.After the gallery sample set is expanded,we add it to a large public dataset together with the generic dataset.Then,a deep convolutional neural network model is trained on this new dataset for feature extraction.We also propose an extended softmax(E-Softmax)classification algorithm.Experimental results on three public datasets demonstrate the effectiveness of our method.On the FERET dataset,the recognition rate of the E-Softmax algorithm reaches 99.5%,which is0.6% higher than that of the commercial software Face++.(2)A low-rank regularized representation with block-sparse structure(LRR-BSS)is proposed.LRR-BSS divides each sample into multiple local patches,generates a subspace for each local patch,and then explores the relationship between subspaces.To accurately describe intraclass variations,we construct an intraclass variation dictionary from the generic dataset and introduce it into LRR-BSS.Hence,we further propose a low-rank regularized generic representation with block-sparse structure(LRGR-BSS).To reduce the size of the intraclass variation dictionary and eliminate the contour noise in the intraclass variation dictionary as much as possible,a robust intraclass variation dictionary learning algorithm is proposed.Experimental results on four public datasets demonstrate the effectiveness of LRR-BSS algorithm and LRGR-BSS algorithm.On the Extended Yale B dataset,the recognition rate of LRGR-BSS algorithm is 2.5% higher than Face++.(3)A uniform generic representation(UGR)algorithm is proposed.The global generic representation is robust to recognize non-discriminative regions such as the forehead and cheek,while the local generic representation can reduce the impact of face variations.UGR integrates the global and local generic representations into a unified framework and explores the consistency between them.UGR employs both the global and local features of the image,so the performance of UGR is better than both the global and local generic representations.The experimental results on four public datasets show that UGR is superior to the other representative comparison algorithms.(4)A patch based semi-supervised linear regression(PSLR)algorithm is proposed.Different from the previous algorithms based on the generic dataset,PSLR does not need to use the category information in the generic dataset.To learn face variation information,PSLR treats the label space of face images as a coordinate system,maps the training samples to unit coordinates,and maps the unlabeled samples to equidistant points.The theoretical analysis of how to use unlabeled samples is made in this paper.To further improve the performance of PSLR,a multi-stage PSLR(MPSLR)algorithm is proposed.MPSLR selects reliably labeled probe samples and adds them to the training process to improve the discrimination ability of the regression model.We conduct a detailed analysis in this paper on how to select reliable probe samples and avoid the negative impact of those selected samples with wrong estimated labels.The experimental results on five public datasets demonstrate the effectiveness of PSLR algorithm and MPSLR algorithm.On the Extended Yale B dataset,the recognition rate of MPSLR algorithm reaches 96.8%,3.8% higher than that of Face++.