A Salient Region Detection Model Using Semantic and Color Information

Abstract

Due to the lack of object semantic information, existing salient region detection models using visual stimulus cues and prior knowledge are diffcult to detect some complicated salient regions with heterogeneity, low contrast or large scale. We construct a semantic salient region detection network using the fully convolutional structure in this letter. After training effectively, the proposed network can effectively extract the semantic salient regions. Aiming to the defect of inaccurate boundary of the detected semantic salient regions, we further introduce the superpixel segmentation and color appearance model to generate the superpixel-level foreground probability. Finally, we propose a new optimization model to fuse the foreground and background probability information and semantic information to derive more accurate salient region map. The experiments comparing with 11 state-of-the-art models demonstrate the effectiveness and robustness of the proposed model.