LRDE:Long-Range Dependency and Salient Feature Enhancement for Co-SOD
Longsheng Wei, Siyuan Guo, Guanyu Zong, Jiu Huang · 2022 41st Chinese Control Conference (CCC) · 2022
Different from other salient object detection (SOD) for the task of extracting salient objects in a image, collaborative saliency object detection (Co-SOD) targets a set of related image groups, and uses algorithms to find and locate the common collaborative salient objects in related image groups. The current SOD has achieved remarkable results, but most of the previous researches focus on how to obtain the consistency between images, while the contextual semantic information of images and the features of salient objects are less utilized. To solve these problems, we propose a novel network to capture long-range dependencies on images and enhance (LRDE) the representation of salient objects. Specifically, In LRDE, we capture the long-range dependencies by the Non-local module to obtain contextual semantic information. Furthermore, in order to highlight the features of salient objects and enhance the representation of salient objects, we propose the Coordinate Attention module, which embeds the location-attention channel attention and encodes the feature maps as orientation-aware and position-aware attention maps, and It is applied complementary to the input feature map, thereby enhancing the representation of salient objects. Experiments show that the method outperforms twelve improved methods on four publicly available datasets.