Bi-directional Features Reuse Network for Salient Object Detection.

Recently, unidirectional convolutional neural networks have been widely used for salient object detection. However, most methods cannot solve common problems (i.e., the loss of valid information, tiny predicted feature, and isolated features in one block), which lead to inefficient feature reuse and blurred salient object edges. To address these problems, we propose a novel bi-directional features reuse network (BDFRN) for salient object detection, which consists of two subnets: forward-skip subnet and reverse-connect subnet. The forward-skip subnet employs an encoder-decoder structure to remedy the loss of salient details, and progressively refine the size of the predicted feature; meanwhile, the reverse-connect subnet can transmit the location features from top blocks to bottom blocks, such that these features can be reused and communicated between different blocks. Extensive experiments are conducted to demonstrate the performance of the proposed method, as compared with baseline methods.