Integrated Sparse Description Of Discriminant Projection And Image Recognition

Image recognition is a hot research topic in the field of pattern recognition and machine learning, which refers to using the computer to analyze and process images, and then identify different patterns of target and object. With the rapid development of information acquisition technology and computer technology, the actual image dimension is increasingly higher. So how to learn the features to effectively describe the data is the key and difficult point for the present study. Feature extraction is an effective solution, and its purpose is to dig the hidden intrinsic low dimensional geometric description of high-dimensional data. Feature extraction can also reduce the computational complexity of the algorithm. This article is based on sparse description, and the sparse description suitable for discriminant feature extraction is studied, the main contents and contributions are as follows:First, Linear Discriminant Analysis(LDA) only considers the global geometric structure of data, which leads to problems such as inaccurate internal geometry description and not very good classification performance. With the advantages of sparse description(better depicting the local geometric structure of data), a Sparse Description-based Linear Discriminant Projection(SD-LDP) is put forward. The purpose of this algorithm is to find the projection matrix, which makes the low dimensional description not only satisfy the LDA criterion but also suitable for Sparse description Classifier(SRC). Compared with LDA and sparse description based SRC- DP algorithm, this algorithm can better depict geometric structure of data and its performance is relatively stable. The experimental results on several database have verified the effectiveness of the proposed algorithm.Second, the SRC- DP algorithm is sensitive to noise and corruption, to address this problem, the robust sparse discriminant projection algorithm is proposed. The algorithm first uses the RPCA to preprocess the training samples to get the clean images, achieving the goal of de-noising; then considering the sparse distribution of coefficients, pointing to the class center sparse discriminant projection algorithm is proposed. The algorithm makes the low dimensional description a better sparse description, which means the coefficients of the non-zero value distribute in the location corresponding toatoms of the same class as far as possible while coefficients corresponding to atoms of different classes are zeros or very small. Due to the above two points, the algorithm better improves the recognition performance and generalization ability. The experimental results have proved the rationality of the proposed algorithm.