Saliency detection via PCA of image patches and ICA-R

Abstract

A variety of computational models based on different views for visual saliency detection have emerged in computer vision. In this paper, we combine RGB space and Lab space to represent original features of each image patch in the vector form, and employ principal component analysis (PCA) to select effective features. The saliency cue is obtained by integrating the contrast and distribution of image patches in the reduced dimensional space and supposed to be the reference for saliency foreground. Then we regard saliency map computation as a source signal separation problem in the framework of independent component analysis with reference (ICA-R). Our approach utilizes the common ICA-R to separate the saliency object from the background of the input image with the saliency cue. The main benefits consist of two aspects. On one hand, the combination of Lab and RGB color spaces in both the saliency reference computation and saliency result computation is able to produce better saliency maps than those using only one color space. On the other hand, the input image is assumed to be a mixture of saliency foreground and background, which enables the saliency detection can be embedded into the source separation problem. This source separation problem is solved by ICA-R with the reference for saliency foreground. Extensive experiments on public saliency benchmarks demonstrate that compared with the state-of-the-art algorithms, the proposed method is more effective.