J-SOLO: Visible Aircraft Image Instance Segmentation Network based on Joint Attention and Boundary Loss

Aircraft image instance segmentation is an important task in aircraft type recognition and is of great significance in many application fields such as civil aviation and military reconnaissance. However, due to the special properties of aircraft images, such as illumination changes, complex target shapes, and background noise, the aircraft image instance segmentation task still faces many challenges. In this paper, we propose a Joint Attention and boundary-based Instance Segmentation Network (J-SOLO) for Visible Aircraft Images, which significantly improves the performance of instance segmentation through the introduction of a Joint Attention Module, Boundary Loss Function, and Second Generation Deformable Convolution. First, we introduce a joint attention module based on spatial and coordinate attention and embed it after the output of each multilevel feature fusion network layer to enhance the network’s sensitivity to positional features. Second, we design a boundary loss function and add it to the loss function of the mask branch to penalize the boundary error more effectively. Finally, we introduce a second-generation deformable convolution into the feature extraction network to replace the traditional convolution, which solves the problems of an inflexible object-aware field and insufficient generalization ability due to the fixed structure of the conventional convolution. Experimental results show that the proposed network performs well in instance segmentation on both the visible airplane image and the COCO dataset.