Multi-scale structure aware network for salient object detection

Abstract

The methods based on convolutional neural network (CNN) have greatly boosted the performance of salient object detection. However, detecting salient objects of multiple scales remains a challenge. To identify salient objects with different scales, we propose a novel multi-scale structure aware network, which can effectively integrate the semantic and detail information of multi-level features to enhance the complementarity between the features at different scales. Specifically, we first propose a global context enhancement module (GCEM) to enhance the relationship between highlevel features in CNN through the atrous spatial pyramid pooling module. Then, we utilize the channel and spatial attention mechanisms to suppress the background noises in above enhanced features. Next, we design feature integration module (FIM) to selectively fuse the low-level detail features, high-level semantic features, and global context information generated by GCEM. Finally, the proposed GCEM and FIM module are progressively cascaded three times to generate final saliency maps. Experimental results on three benchmark datasets demonstrate that our approach achieves superior performance over state-of-the-art methods.