Learning Fine-Grained Information with Capsule-Wise Attention for Salient Object Detection

With the popularity of convolutional neural networks being used for salient object detection (SOD), the performance has been significantly improved. However, how to integrate crucial features for modeling salient objects needs further exploration. In this work, we propose an effective feature selection scheme to solve this task. Firstly, we provide a Simplified Atrous Spatial Pyramid Pooling (SASPP) module to lightweight the multi-scale features. Dealing with the SASSP features, we design a pixel-level local feature selection scheme named Multi-Scale Capsule-wise Attention (MSCA). It aggregates features from multi-scales by dynamic routing and helps the network to generate fine-grained prediction maps. In addition, we exploit holistic features by the Spatial-wise Attention and Channel-wise Attention (SA/CA) mechanisms, which adaptively extracts spatial or channel information. We also propose a Multi-crossed Layer Connections (MLC) structure in the upsampling stage, to fuse features from not only different levels but also different scales. The salient object prediction is performed in a coarse-to-fine manner. By conducting comprehensive experiments on five benchmark datasets, our method achieves the best performance when compared to existing state-of-the-art approaches.