Joint training with the edge detection network for salient object detection

The U-shaped network has great advantages in object detection tasks. However, most of the previous salient object detection studies still suffered from inaccurate predictions affected by unclear object boundaries. Considering the complementarily of the information between salient object and salient edge, we designed a new kind of network to effectively perform the joint training with edge detection tasks in three steps. Firstly, we added a prediction branch on the bottom-up pathway for capturing the edge of salient objects. Secondly, salient object features, global context, integrated low-level details, and high-level semantic information are extracted by the method of progressive fusion. Finally, the feature of the salient edge is concatenated with that of the salient object on the last layer in the top-down pathway. Since the salient edge feature contains much information about edge and location, the feature fusion can locate salient objects more accurately. The results of experiments on five benchmark datasets demonstrate that the proposed approach achieves competitive performance.