Multi-level Features Selection Network Based on Multi-attention for Salient Object Detection.

Both the attention mechanism and Salient Object Detection aim to locate the most obvious regions in an image. However, common algorithms focus on micro attention but neglect the similarity in the macro perspective. Besides, they also ignore the differences among multi-scale information. To tackle these problems, we propose MFS-Net that progressive select features to predict salient regions. First, we design the Pyramid Attention module that integrates channel and spatial attention to extract semantic information for multi-scale high-level features and design the Self-Interaction Attention module to extract detailed information for multi-scale low-level features. Besides, to refine the saliency edge, we propose the Semantic-Detail Attention module which exploits high-level features to guide low-level features in a macro-attention manner. Finally, we selectively integrate global context information by the Interaction-Fusion Attention module, aiming to learn the relationship among different salient regions and alleviate the dilution effect of features. Experimental results on six benchmark datasets demonstrate that the proposed method performs well compared with 20 state-of-the-art methods.