Research On Sparseland Model For Image Recognition

With the development of the mobile Internet and digital multimedia, im-ages play a significant role in communication between people, image recogni-tion will be more and more important in the future. Sparseland model, a cross subject of neurophysiology, computer science and artificial intelligence, has made great progress in image recognition and become one of the difficult and attractive fields recently.Firstly, in-depth research of the dictionary learning algorithm, sparse cod-ing algorithm and dimensionality reduction algorithm of sparse representation are conducted in this dissertation, and then a series of improved algorithms are proposed. Then, the effectiveness of the proposed algorithms is validated by substituting them in a sparseland model based image recognition system and performing experiments with it.The main contributions and innovations of this dissertation are as follows:1. According to the block sparsity and discriminative assumption of sparse-land model in image recognition, a Block Discriminative Dictionary Learn-ing (BDDL) algorithm is proposed. Firstly, the block sparsity is used to determine the non-zero entries of the sparse representation and simplifies the calculation of it. Secondly, a dictionary incoherence term is introduced into the objective function of BDDL to enlarge the intra-class distance and enhance the discriminative power of the dictionary. Thirdly, traditional sparse coding algorithms are replaced by structured sparse coding algo-rithms in the test phase. Finally, a gradient-based optimization strategy is also developed. The experiment on face dataset proves the effectiveness and robustness of BDDL2. On the basis of correlations between the signal in the same class, Structure-Constrained Low-Rank Dictionary Learning (SCLRDL) algorithm and Low-Rank and Partial Sparse Representation (LRPSR) algorithm are proposed. Firstly, SCLRDL constrains the non-zero entries of sparse representation by block sparsity, and introduces low rank regularization to model the cor-relations. Secondly, LRPSR concatenates the training samples and testing samples to prevent lowering the rank of insufficient testing samples. Fi-nally, The Augmented Lagrange Multipliers Method (ALM) based opti-mization strategy of SCLRDL and LRPSR is also derived. Experimental results of face recognition and object recognition show that SCLRDL and LRPSR can improve the accuracy of recognition.3. Graph-based Block Sparse Dimensionality Reduction algorithm (GBSDR) is proposed for the block sparsity of sparse representation. Firstly, The Block Sparse Representation Distance (BSRD) is designed which adopts different calculation of entries in the same class and entries in the sepa-rate class to reduce the errors caused by misalignment of non-zero entries within class and enhance the discriminative power of the algorithm. Sec-ondly, GBSDR constructs the adjacent matrix of samples based on BSRD and the dimensionality is reduced by preserving local property of the sig-nal. Finally, this thesis also gives the solution of GBSDR. Experimental results on synthetic data and face images prove that GBSDR can reduce the error rate of image recognition.4. To solve the "out-of-sample extension" problem of the non-linear dimen-sionality reduction algorithm, the Graph-based Block Sparse Linear Di-mensionality Reduction algorithm (GBSLDR) is proposed, and SCLRDL, LRPSR and GBSLDR are combined to form a Discriminative Block Sparse and Low Rank based image recognition (DBSLR) algorithm. Firstly, GB-SLDR is proposed by linearizing GBSDR to make it more practical. Then, DBSLR is proposed by substituting SCLRDL, LRPSR and GBSLDR in sparseland model based image recognition system. The effectiveness of the innovations in the thesis are validated by performing face recognition and object recognition experiments with DBSLR.