Intensifying graph diffusion-based salient object detection with sparse graph weighting

Salient object detection based on the diffusion process on the graph has achieved considerable performance. It mainly depends on the affinity matrix construction considering the local structure. This paper aims to depict the local and global structures from image features, intensifying the graph-based diffusion model by simultaneously integrating the sparse graph matrix and affinity graph matrix. The contribution work computes the affinity graph matrix and delivers an affinity matrix by incorporating the sparse representation and diffusion process. It estimates a sparse graph matrix by integrating sparse representation and laplacian smoothness. To this end, a two-stage graph-based diffusion model has been constructed by embedding the manifold smoothness and manifold reconstruction. The first stage follows the boundary-prior to generate a coarse saliency map. After, the second stage combines the saliency map and Harris convex hull to obtain the foreground seeds. Extensive experiments on six benchmark datasets have demonstrated the superiority of the proposed method compared to other state-of-the-art methods.