Discriminant Projection For The SRC

Feature extraction and classification are two basic problems in pattern recognition and subspace learning is a kind of classical feature extraction method which has been widely applied to image recognition, pattern recognition and other fields. The goal of feature extraction is to find a projection matrix which makes the low-dimensional description meet the following analysis, such as classification. K-Nearest Neighbors(KNN) is a classical classifier. The disadvantage of KNN is that it is very sensitive to occlusion and noise, which results in the poor robustness. Compared with KNN, SRC(Sparse Representation-Based Classifier) can achieve better classification performance for image noise and occlusion. And it has been widely used in image recognition, computer vision, pattern recognition and so on. However, the existing feature extraction methods mostly have no direct link to SRC which leads to the bad performance of SRC. Based on the SRC, this article aims to find a more suitable discriminative projection method for the SRC. The main contents are as follows:1. In view of the fact that the sparse representation classifier steered discriminative projection(SRC-DP) only considers the reconstruction residual of the sample data, and ignores the local discriminative geometric structure of the data, this paper puts forward a sparse classification-based linear discriminative projection method. The purpose of this algorithm is to find a projection matrix which can make the training samples meet the following conditions after projection. The reconstruction residual is as small as possible when reconstructing the training samples with the training samples from the same class. At the same time, find out the atoms from different classes which are similar to the training samples and make the distance between the atoms and the training samples as large as possible. It describes the local discriminative geometric structure of the data in a good way. Experiments on five databases demonstrate the effectiveness of the proposed algorithm.2. In view of the fact that sparse description cannot guarantee the non-zero coefficient corresponding atoms belong to the same class, this paper proposes a sparse classification-based local discriminative projection algorithm to get the extracting discriminant features in a better way. This algorithm uses the atoms from the same class as a dictionary, which can make the reconstruction residual as small as possible. This can effectively describe the local intrinsic geometric structure of data. Meanwhile, it combines the reconstruction information with the local discriminative geometric structure, which avoids the shortcomings of just considering one of them. The experiments demonstrate that this algorithm has certain superiority.