Mirror Feature Fusion for Salient Object Detection

Abstract

The mainstream salient object detection method often fuse and refine at the feature level to achieve saliency object detection at present. However, different features play different roles in salient object detection due to the differences between the feature of each layer of the depth model, so simple fusion of the feature of each layer will affect the performance of the system. In view of the above observation, we propose a mirror fusion network (MFNet) in the paper, which is mainly composed of low-level denoising module (LDM), refinement module (RM), and mirror fusion module (MFM). With the assistance of three algorithm modules, not only the multi-level features can be fused effectively, but also the detailed object boundary can be obtained. We take advantage of a hybrid loss function to measure the system loss when the model is training. Experiments on five benchmark datasets show that the proposed method is superior to seven related deep salient object detection methods under three evaluation metrics.