On Compressive Sensing-based Face Recognition Techniques

With the rapid development of network and information technology,especially the emergence of big data and mobile Internet,pattern recognition,as an important part of in-formation and artificial intelligence,has been widely used in machine vision,video surveil-lance,search engines,industrial automation and other fields.In these applications,a big problem one confronts is the huge amount of data to process.Compressive sensing or Com-pressed sampling is a hot topic of mathematics and information science in recent years.Compressive sensing has been considered as an effective way to deal with big data prob-lem.Taking face recognition as the main application,this thesis focused on developing new techniques using compressed sensing for classifier design,feature extraction,dimension reducing and image preprocessing.The main works and contributions of this dissertation are as follows:1.On robust classification algorithmsA novel classification algorithm based on subspace analysis was proposed.The clas-sifier for each class was designed as a projector so that the samples within the class are projected to a near zero vectors,while the samples not belonging to the class are projected far away from zero vectors.Experiments on open and closed universe test of ORL data set showed that the proposed subspace analysis-based classifier has a high recognition rate in closed universe test and it is more robust than other methods in open universe experiments.The proposed algorithm has low computational complexity which is suitable for distributed embedded system applications.2.Design of robust compressed sensing matrix based on prior informationBased on the prior information of signals,a new robust compression sensing design framework was proposed.Sparse projection matrix was obtained by minimizing the sparse representation error and the error between the true signal and its best linear projection with the clean measurement.An algorithm for the corresponding optimal design problem was derived.Experimental results on synthetic and real images showed the effectiveness of the new robust projection matrix CS system in pattern recognition and data compression.3.Dictionary learning based classification on salient contour featureAs dictionary learning classification is very costly with large amount of train features and the classical sparse representation classifier is sensitive to face alignment issues,a salient contour as the dictionary patch center was proposed for features selection.Mean-while the mirrored face was used to extended training samples.Experimental results show the great improvement in face recognition performance.4.Sparse representation based classification via optimal projection matrixAs a large amount of data leads to the problem of storage and transmission.A nov-el framework,called projection matrix optimization based compressive classification was proposed.The projection matrix was designed such that the coherence between different classes of faces was reduced and hence a higher recognition rate was expected.Experi-ments were carried out with five popularly utilized face databases(i.e.,ORL,Yale Extend,CMU PIE,and AR)and simulation results showed that the proposed system yields a great improvement performance in terms of the recognition rate and reconstruction error.5.ADLBP based face recognition systemA novel face recognition system was proposed for complicated environment.Face detection was first used.With the 3D reference model,then the frontalized face was used for the feature extraction.After LBP feature extraction and down sampling,the low rank sparse decomposition was used for image alignment.The equivalent dictionary produced by projection matrix optimization was used for robust non-negative sparse representation.The experimental results showed that it can significantly to improve the recognition perfor-mance to 94.74%in the close set test.In the open set test,with 80%recall rate,a more than 95%precision rate is achieved.