Salient object detection via multi-spectral co-connectivity and collaborative graph ranking

Salient object detection (SOD) has become an active research direction with extensive applications in computer vision tasks. Although integrating RGB and infrared thermal (RGB-T) data has proven to be effective in adverse environments, it is difficult for RGB-T SOD methods to highlight the salient objects completely when objects cross the image boundary. To address the aforementioned problem, this paper proposes an effective RGB-T SOD algorithm based on multi-spectral co-connectivity (MSCC) and collaborative graph ranking. Specifically, we introduce the multi-spectral weighted color distance to construct an improved undirected weighted graph and compute the MSCC-based saliency map. Simultaneously, the MSCC-based background probability map is also calculated and employed in the following processing of real background seeds selection. Then, we utilize collaborative graph learning (CGL) and calculate the CGL-based saliency map in a two-stage ranking framework. Finally, we integrate these two saliency maps through multiplying or averaging to enhance the final saliency result. The experimental comparison results of 5 quantitative evaluation indicators between the proposed algorithm and 9 state-of-The-Art methods on RGB-Thermal datasets VT821 and VT1000 datasets demonstrate the robustness and superiority of the proposed work.