Deep saliency detection via channel-wise hierarchical feature responses

Abstract Recently, deep learning-based saliency detection has achieved fantastic performance over conventional works. In this paper, we pay more attention to channel-wise feature responses and propose an end-to-end deep learning-based saliency detection method. Our model contains channel-wise coarse feature extraction, channel-wise hierarchical feature refinement (CHFR), and hierarchical feature maps fusion. The whole process is based on the squeeze-and-excitation-residual network (SE-ResNet) to explicitly and globally model the inter dependencies between the channels of its convolution features at slightly computational cost. Firstly, we make channel-wise feature extraction to produce coarse feature maps with much information loss. Then, CHFR is executed based on SE-ResNet modules to make hierarchical feature refinement. Finally, the hierarchical feature maps are fused to generate the final saliency map. The network applies softmax cross entropy loss in the training. Compared with other fifteen state-of-the-art approaches, the experimental results demonstrate the high computational efficiency and superior performance of the proposed approach according to comprehensive evaluations over six benchmark datasets.