Diversity Representation Based On Graph And Its Application

In recent years, biometric feature recognition has aroused great attention of researchers in many fields, where face recognition with its own advantages has become a research hot spot in biometric recognition. In face recognition, how to effectively extract the image feature has been one of the key problems to be solved. Among these methods, the subspace analysis methods have become a very active research direction currently. This paper will make an in-depth study of the diversity representation based on graph theory and measurement by using PCA and manifold learing techniques. The main contents and contributions of this paper are as follows:1. Supervised Diversity Embedding (SED) is proposed. SDE constructs an adjacency graph, namely diversity adjacency graph, over the training data to model the diversity of data. In order to obtain better classification performance, SDE assigns a larger weight to the variation of the values among nearby data from different classes, which efficiently improves the recognition accuracy. Extensive experiments indicate the efficiency of the proposed method.2. Two-dimensional Supervised Diversity Embedding (2DSED) is proposed by combining the two-dimensional feature extraction techniques and SDE.2DSDE directly calculates the diversity scatter matrix from the image matrices rather than image vectors, so it avoids transforming the image matrix into a vector, alleviates the impact of the problem of small samples and reduces the computational complexity. Based on this content, the2DSDE plus2DPCA is proposed to feather improve the performance of the2DSDE.3. Image Euclidean Distance based Two-dimensional Supervised Diversity Embedding (IED-2DSED) is proposed. IED-2DSED calculates the diversity scatter of the training data using the image euclidean distance rather than traditional euclidean distance. Image euclidean distance takes into account the spatial relationals among pixels in image, which will improve the robustness of the algorithm. Experiments on Yale. AR. UMIST and PIE face databases indicate the efficiency of2DSDE.