A Dual-Branch Awareness Network for Small Object Segmentation in Large-Scale Remote Sensing Scenes

The more detailed and accurate earth observation has been made driven by the advancement of satellites and sensors optical photography technology, which poses both a challenge and an opportunity to small object segmentation task. However, the inherent difficulty and inadequate consideration still make small object segmentation task inevitably encounter a performance gain bottleneck. In this paper, we consider the longstanding but underestimated challenges in this task and give a point-to-point solution to response them. Specifically, we introduce a discriminative structure, i.e., a dual-branch awareness network for small object segmentation, named DASNet. In this structure, we propose the small object activation branch and the fuzzy refinement branch to avoid the negative influence of redundant background and ensure the small object segmentation accuracy, respectively. These two branches work collaboratively to mimic the process of human visual perception on small object. Finally, we propose a hierarchical unbiased loss to eliminate the bias against small objects in the regression process. Extensive experiments demonstrated that DASNet is competitive against some advanced methods for small object segmentation.