Instance Segmentation in Very High Resolution Remote Sensing Imagery Based on Hard-to-Segment Instance Learning and Boundary Shape Analysis

Although great success has been achieved in instance segmentation, accurate segmentation of instances remains difficult, especially at object edges. This problem is more prominent for instance segmentation in remote sensing imagery due to the diverse scales, variable illumination, smaller objects, and complex backgrounds. We find that most current instance segmentation networks do not consider the segmentation difficulty of different instances and different regions within the instance. In this paper, we study this problem and propose an ensemble method to segment instances from remote sensing images, considering the enhancement of hard-to-segment instances and instance edges. First, we apply a pixel-level Dice metric that reliably describes the segmentation quality of each instance to achieve online hard instance learning. Instances with low Dice values are studied with emphasis. Second, we generate a penalty map based on the analysis of boundary shapes to not only enhance the edges of objects but also discriminatively strengthen the edges of different shapes. That is, different areas of an object, such as internal areas, flat edges, and sharp edges, are distinguished and discriminatively weighed. Finally, the hard-to-segment instance learning and the shape-penalty map are integrated for precise instance segmentation. To evaluate the effectiveness and generalization ability of the proposed method, we train with the classic instance segmentation network Mask R-CNN and conduct experiments on two different types of remote sensing datasets: the iSAID-Reduce100 and the JKGW_WHU datasets, which have extremely different feature distributions and spatial resolutions. The comprehensive experimental results show that the proposed method improved the segmentation results by 2.78% and 1.77% in mask AP on the iSAID-Reduce100 and JKGW_WHU datasets, respectively. We also test other state-of-the-art (SOTA) methods that focus on inaccurate edges. Experiments demonstrate that our method outperforms these methods.