Contextual Attention Enhanced Network for Salient Object Detection in Optical Remote Sensing Images

Salient object Detection (SOD) have witnessed huge success in nature scene images (NSIs), however, they are unable to deal with SOD for optical remote sensing images (ORSIs). Although recent efforts on SOD for optical RSIs have achieved impressive results, we still face significant challenges due to cluttered backgrounds, multi-scale objects, and complex imaging conditions. To remedy the issue, we propose an original contextual attention enhanced network (CoAENet) for SOD in ORSIs, including contextual attention enhancer and detail refinement module. Specifically, contextual attention enhancer can mine the contextual attention information from the coarse-grained feature maps, which can improve the expressiveness of the fine-grained feature maps. The detail refinement module can learn the specific information of salient objects to refine the predicted results. Meanwhile, to improve the performance of the model from two perspectives of pixel and region, a hybrid loss is designed. Experimental results on two baseline datasets prove CoAENet successes to improve the detection results compared with existing methods. It reveals CoAENet provides an effective solution for SOD in ORSIs.