PSCFNet: Prototype Learning and Spatial Consistent Feature Fusion for Efficient Salient Object Detection

Salient object detection (SOD) aims at distinguishing the most attractive regions from images. Existing encoder-decoder based SOD approaches have achieved significant progress by utilizing various multi-level feature fusion strategies. However, the problem of unclear object boundaries still remains unsolved due to discrepancies between multi-level features and spatial misalignment in feature fusion. In this study, we propose the PSCFNet to address these problems. The proposed network consists of a lightweight Prototype Guided Module (PGM) and a progressive Saliency Mining Module (SMM). Specifically, PGM establishes a semantic prototype to enhance the shallower features by utilizing the deepest features. Furthermore, SMM is proposed to aggregate the multi-level features by adopting a progressive feature alignment strategy. Experiments on five widely-used benchmark datasets demonstrate our proposed method achieves remarkable performance with 23.4% reduction in computational cost. We release the experimental code at https://anonymous.4open.science/r/PSCFNet-0430.