HFNet: Hierarchical feedback network with multilevel atrous spatial pyramid pooling for RGB-D saliency detection

Multiscale features have received considerable attention for improving saliency detection. However, existing methods only perform decoding at multiple scales without exploring feature refinement. We propose a hierarchical feedback network (HFNet) with multilevel atrous spatial pyramid pooling (MASPP) to refine multiscale features for RGB-D saliency detection. The improved MASPP for adaptive refinement is applied to hierarchical network modules to obtain multiscale information. Then, the detailed multiscale information is used for decoding based on a channel attention mechanism with joint information guidance, and the output is fed to the next stage through reverse attention. This iterative refinement reuses feature information and predicts more precise saliency maps with detailed information. Experimental results on seven benchmark datasets show the effectiveness of the proposed HFNet, and ablation studies confirm the effectiveness and superiority of applying its different strategies. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

In this study, a hierarchical feedback network (HFNet) is proposed to refine multiscale features for RGB-D saliency detection. Experimental results demonstrate the effectiveness of HFNet, and ablation studies confirm the superiority of its different strategies.