Image Recognition Based On Sparse Representation

With the rapid development of data collecting techonology and sensor techonology, dimensionality of natural images becomes higher and higher, thus, how to represent images effectively so as to facilitate subsequent processing, such as image analysis and recognition, has become one of the urgent problems to be solved in image processing, pattern recognition and machine learning, etc. Among the most methods, sparse representation has been an activce research field because of its good robustness and generalization ability and anti-jamming ability. However, it needs a large volume of training samples so as to obtain better sparse coefficients, which leads to complex computation, another disadvantage is that it neglects the inherent geometrical structure of image space, which will impair the recognition performance. This dissertation studies the image sparse representation and recognition by combining the dictionary learning and discriminant manifold learning. The main contents and contributions are as followes:1. Introduce the theoretical framework of compressed sensing and the principles of each component. Simple introduce the commonly used sparse representation method, measurement matrix construction method and reconstruction algorithm.2. A new image recognition algorithm based on the dictionary learning is presented. The proposed algorithm uses K-SVD to learn the dictionary database, which effectively reduces the number of atoms. Based on this dictionary, better sparse coefficients are obtained for image representation. And then, image recognition is realized via the residual errors.3. A new sparse coding algorithm is based on the discriminant manifold learning is presented. The proposed approach, namely Discriminat Graph Sparse Coding (DGSC), constructs two adjacency graphs to model the intrinsic geometry and discriminant structure of image space, and incorporates intrinsic geometry and discriminant structure into the sparse representation objective function to learning the dictionary and sparse coefficients. Experiments based on SVM indicate the effectiveness of our method.