Finding Relationships between Human Affects and Colors Using SVD and pLSA

Abstract

In this paper, a new method is presented to automatically find relationships between human affects and colors. For this, the probabilistic latent semantic model analysis (pLSA) and singular value decomposition (SVD) is applied. The proposed method is composed of three modules: feature extraction, feature transform and pLSA training. We first segment the image using mean-shift clustering, then extract color compositions by analyzing the colors from one region and its adjacent regions. Next, for the occurrence matrix, the SVD and pLSA are used. Using SVD, the occurrence matrix is decomposed into rank and null space matrix, where the null space is discarded and only the space corresponding to the singular values is used for further processing. For the reconstructed matrix, the pLSA is applied to obtain the correlation between affective classes and color compositions. To assess the effectiveness of the proposed system, it was applied to index the images using human affects. Then the results showed the effectiveness of the proposed method.