Fusion and Inpainting: A Novel Salient Object Detection Network

Aiming at obtaining an accurate salient image map, a common practice in salient object detection is to fuse multi-level features and incorporate auxiliary context information of the object. However, this often involves a large number of sampling operations, leading to the blurring of object features such as edges in complex regions. To address this issue, we propose incorporating an inpainting module after the multi-level feature fusion process to reconstruct these blurry areas. To this end, we present a novel salient object detection network structure that comprises a feature fusion module and an inpainting module. In the inpainting module, we use recursive gated convolution to capture both the global structure and local detail information of the salient object to repair the faded and unclear areas with high quality. We conduct extensive experiments on four benchmark datasets, demonstrating that the proposed network outperforms most state-of-the-art methods.