Sparse Representation Algorithm For Image Recognition Based On Discrimination Dictionary Learning

Image recognition has being a hot topic in the field of pattern recognition andcomputer vision, it has a wide range of application in military area, public security,industry and daily life. In recent years, the sparse coding method have been successfullyused in various image recognition applications, it has held more and more researchers’interest. This paper put forward three new sparse representation recognition algorithmsbased on the analysis and summary of the relevant research.Firstly, Considering the Gabor feature is robust to variations of illumination,expression and pose, Gabor Feature Set Discriminative Dictionary Learning algorithm forimage recognition is proposed. The proposed method first extracts image Gabor featureswith multi-scale and multi-orientation; then uses the augmented Gabor local feature matrixwhose dimension has been reduced to construct the initial feature dictionary, based on theFisher discrimination criterion, a structural dictionary, whose atoms have correspondenceto the class labels, is learned so that each sub-dictionary of the learned new dictionary hasgood representation power to the samples from the corresponding class; Consequently,each feature vector is coded on its associated new feature dictionary. Experimental resultsverify the effectiveness of the algorithm.Secondly, sparse representation algorithm for face recognition based on multi-scalelocal discriminative dictionary learning is proposed. First of all, each face image ispartitioned into multi-scale non-overlapping patches, building the multi-scale local imagesubsets and serve as original dictionaries respectively. And then, we make full use of thesample label information in each multi-scale original dictionary to learn discriminativeover-complete dictionary and optimal linear classifier jointly. After that, all the patches ofa test sample are coded based on sparse representation in corresponding dictionaryrespectively. Finally, a voting scheme is adopted to determine the types of the test sample.Experimental results demonstrate the rationality of the algorithm.Finally, considering the fact that global features and local features in a sample shouldcontribute differently to the pattern representation and classification, we present a novel face image recognition algorithm based on global and local features joint sparserepresentation. Firstly, global features are extracted from whole face images and localfeatures are extracted from some spatially partitioned image patches to construct thecorresponding feature dictionary. Then, each feature vector is coded on its associatedfeature dictionary. In addition, distinguish the contribution of different features byadaptive weighting in the coding domain. Consequently, the classification of a test sampleis based on the overall coding error for each class. Experimental results show that theproposed algorithm is feasible.