Salient object detection with detail-preserving pooling and feature channel refinement

This paper proposes a novel end-to-end approach for salient object detection task to enhance the performances. The traditional downscaling method such as max pooling layer is replaced by detail preserving pooling layer to capture more effective features. Moreover, the squeeze and excitation block is adopted to extract image features with the channel wise importance. Finally, densely connected architecture is introduced to maximizes feature reuse and reduce the computational cost in the process of generate saliency maps. The proposed method, on several public benchmarks acquires competitive or better performances than other similar approaches.