Semantic feature based multi-spectral saliency detection

Abstract

Saliency detection aims to locate the distinctive regions in images and can be extensively applied to many applications. Up to now, most of effort has put into visible images and the related methods usually encounter difficulty for images with complex background. In this paper, we propose a semantic feature based multi-spectral saliency detection method using the complementarity of infrared and visible images. We use the thermal infrared image to relieve the difficulty of visible images with complex background, while still utilizing the rich texture and color information in visible images. Specifically, we firstly uses the Convolutional Neural Network to extract high-level feature from superpixels obtained by segmenting visible and infrared images, and then the initial saliency maps of both spectrums are computed, respectively. After that, two initial saliency maps are fused via a Total Variation (TV) minimization model and finally the fused result is linearly combined with the enhanced foreground salient object map to obtain the final saliency detection result. Experiment results reveal that the proposed method outperforms the baseline methods.