DGrid: Dense Grid Network for Salient Object Detection

Abstract

AbstractThe performance of salient object detection has been significantly advanced by using fully convolutional networks (FCN). However, it still remains nontrivial to take full advantage of the multi-level convolutional features for salient object detection. In this paper, a dense grid network framework (denoted \(\mathbf {DGrid}\)) is proposed to solve the above problem, which mainly consists of the backbone module, extended module and fusion module. Specifically, \(\mathbf {DGrid}\) utilizes a multi-branch refinement mechanism for saliency detection. First, the backbone module is used to generate a coarse prediction map. Then, the extended module, which contains four branches, is used to improve the resolution and precision of the prediction map gradually from coarse to fine. Moreover, we proposed the densely connected strategy to fully fuse features at different levels. Finally, the fusion module densely fuses the highest level features of all branches to achieve the final saliency map. Experimental results on five widely used benchmark datasets demonstrate that \(\mathbf {DGrid}\) can improve the accuracy of detection by maintaining a high-resolution feature branch, and it outperforms state-of-the-art approaches without any post-processing.KeywordsSalient object detectionDense grid networkRefinementConvolutional features