Multi-scale Dynamic Convolution and Attention-Guided Network for Salient Object Detection in Optical Remote Sensing Images

Abstract Salient object detection in optical remote sensing images focuses on extracting the most visually distinctive and attention-grabbing regions within such images. To address the issues of high computational cost and complex background interference of this task, a light-weight multi-scale dynamic convolution and attention-guided network is proposed. Firstly, the lightweight EfficientNet-b0 network is used as the backbone to extract different levels of information from the image. The sparse self-attention guidance module is employed to derive deep semantic features. At the same time, dynamic convolution weights are produced to direct the multi-scale feature fusion module in capturing and integrating contextual information. Secondly, a feature detail enhancement module is introduced to enrich the object's detailed information and extract edge features for supervision. A semantic guidance module is applied at the end to merge hierarchical features, resulting in the final saliency maps. The proposed method is shown to substantially reduce computational costs, as evidenced by comparative evaluations on the ORSSD and EORSSD datasets, deal with complex background interference, and obtain higher detection accuracy than existing methods.