Progressive Multi-scale mutual Feedback network for salient object detection

The aggregation of multi-level features extracted from convolutional neural networks has been a pivotal factor contributing to the remarkable progress achieved by most existing salient object detection models. However, the discrepancy between features of different levels has not been comprehensively studied. How to design more effective fusion strategies to reduce this discrepancy has become an important problem in SOD. In this paper, we propose a progressive multi-scale mutual feedback network (PMFNet) to solve the above problems, which is composed of the Context-aware Residual Pyramid Module (CRPM), the Cross Attention Module (CAM), and the Progressive Interweaved Feedback Decoder (PIFD). CPRM integrates various feature information using different receptive fields, enabling the model to acquire both multi-scale information and preserve fine local details of the features. CAM can distinguish between high-level and low-level features and incorporate feature crossing to mitigate the adverse impact of redundant features. Besides, to capture finer details, PIFD can iteratively refine multi-level features using feedback mechanisms. The experimental results on 5 benchmark datasets demonstrate that the proposed method outperforms 12 state-of-the-art methods, showing its superior performance.