Real-time salient object detection based on accuracy background and salient path source selection

Abstract

Boundary and connectivity prior are common methods for detecting the image salient object. They often address two problems: 1) if the salient object touches the image boundary, the saliency of the object will fail, and 2) accurate pixel-wise or superpixel-wise computation needs high time expenditure. This study proposes a block-wise algorithm to reduce calculation time expenditure and suppress the salient objects touching the image boundary. The algorithm consists of four stages. In the first stage, each block is analyzed by an adaptive micro and macro prediction technique to generate a saliency prediction map. The second stage selects background and salient sources from the saliency prediction map. Background sources are extracted from the image boundary with low saliency value. Salient sources are accurately positioned in the region of salient objects. In the third stage, the background and salient sources are used to generate the background path and salient path based on minimum barrier distance. The block-wise initial saliency map is obtained by fusing the background and salient paths. In the fourth stage, major-color modeling technology and visual focus priors are used to complete the refinement of the saliency map to improve the block effect. In the experimental result, the proposed method produced the best test results among other algorithms in three dataset tests and achieved 284 frames per second (FPS) speed performance on the MSRA-10 K dataset. Our method shows at least 29.09% speed improvement and executes in real-time on a lightweight embedded platform.