A multi-source feature extraction network for salient object detection.

Salient object detection (SOD) must capture the multi-level features from both global and local view. Furthermore, interiors and boundaries of salient objects must be processed simultaneously in order to generate a clear salient map with sharp boundaries. In addition, the object-part relationship should be taken into consideration to segment the salient object as a whole. To address above issues, we propose a novel multi-source feature extraction network (MFEN), which is capable of integrating salient features, boundary features and global feature, simultaneously. First of all, the multi-source global and local module (MGLM) is introduced to integrate multi-source features, composing of a series of hybrid dilation convolution modules with different dilated rate. Furthermore, the boundary detection module is introduced to predict the boundary map and boundary features, helping for locating the salient object and sharpen edge. In addition, the adjacent features from MGLM are fused progressively to generate the final salient map by feature fusion modules. Experimental results on five datasets demonstrate that our proposed MFEN outperforms recent 18 SOD methods. More importantly, the ablation study shows that the MGLM is an effective feature fusion module for multi-level and multi-source feature. [ABSTRACT FROM AUTHOR]