A CONVLSTM-COMBINED HIERARCHICAL ATTENTION NETWORK FOR SALIENCY DETECTION

Attention mechanism has been applied in the salient object detection massively. In order to reasonably integrate feature maps at different blocks and focus on the impact between neighbouring pixel concurrently, we propose a novel attention mechanism network, i.e., the ConvLSTM-Combined Hierarchical Attention Network (CHAN), to differentiate the contribution of feature maps of different blocks and focus on the impact between neighbouring pixel at the same time. The spatial attention, combined with ConvLSTM, offers an efficient recurrent mechanism for sequential refinement of the location information of each feature maps. The block-wise attention generates response to different convolutional blocks, and fuses different features reasonably. Experiments show that our saliency model detects salient objects more accurately, thus favorably outperforms most state-of-the-art methods.