Abstract | ||
---|---|---|
The boundary quality is a key factor determining the success of accurate salient object detection (SOD). A number of edge-guided SOD methods have been proposed to improve the boundary quality, but shown unsatisfactory performance due to the lack of a comprehensive consideration of multi-level feature fusion and multi-type feature aggregation. To resolve this issue, we propose a novel Bidirectional Collaboration Network (BCNet), which integrates effective multi-level feature fusion and multi-type feature aggregation into a unified edge-guided SOD framework. Specifically, we first utilize multiple Consistency Saliency Maximization (CSM) modules to propagate the highest level semantic representations in a top-down progressive pathway to generate both global edge representations and a series of region representations. Multiple Bounded Feature Fusion (BFF) modules are then utilized to refine the region features with the edge features. The CSM and BFF modules enable robust multi-level feature fusion and multi-type feature aggregation with only little extra computation, which allows a high computational efficiency. Finally, BCNet is jointly trained with edge and region losses in an end-to-end manner. Extensive comparisons are conducted with 17 state-of-the-art methods on five challenging benchmark datasets. Thanks to the use of CSM and BFF modules, our BCNet outperforms existing deep learning based SOD methods, including the latest edge-guided ones, in terms of both detection accuracy and processing speed. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1016/j.neucom.2021.01.034 | Neurocomputing |
Keywords | DocType | Volume |
Salient object detection,Edge detection,Feature fusion,Deep supervision | Journal | 437 |
ISSN | Citations | PageRank |
0925-2312 | 1 | 0.35 |
References | Authors | |
0 | 5 |