Simple yet effective color principal and discriminant feature extraction for representing and recognizing color images

In this paper, we investigate the problem of extracting two-dimensional color principal and discriminant features for understanding color images. Specifically, two simple yet effective color image feature extraction criteria, called Color Principal Component Analysis (ColorPCA) and Color Linear Discriminant Analysis (ColorLDA), are proposed for color image analysis. The presented criteria can preserve color and topology information of pixels in images, and extract features directly from color images in an efficient manner by eigen-decomposing a single eigen-problem. In modeling the criteria, color image scatter matrices are defined. Like PCA, LDA and their two-dimensional (2D) extensions, our methods only need to choose the number of projection vectors. More importantly, the matrices to be eigen-decomposed in our criteria have the same size as 2DPCA and 2DLDA that are very efficient. To achieve an orthogonal projection matrix, trace ratio ColorLDA is also presented. We also present the alternative versions of our approaches for feature learning through mining row or column information of the images. Extensive simulations on benchmark datasets are conducted to evaluate our algorithms. The investigated cases demonstrate the effectiveness and efficiency of our techniques, compared with other most related state-of-the-art 1D and 2D criteria.