BEIG-SOLO: A Binary Edge Image Guidance based Aircraft Image Instance Segmentation Network

With the continuous advancement of aviation and sensor technology, aircraft image instance segmentation technology is becoming increasingly important in the military and civilian fields. Due to the characteristics of small intra-class similarity and large inter-class similarity of aircraft in aircraft images, as well as the influence of its shooting conditions and complex background, it is difficult for the existing aircraft image instance segmentation technology to segment aircraft images accurately and difficult separating targets from barrier and complex backgrounds. Therefore, how to accurately segment aircraft images based on their characteristics and improve the model’s robustness to occlusion, illumination, and viewing angle changes has become the focus of recent research. In response to the above challenges, we propose an aircraft image instance segmentation network based on binary edge image guidance (BEIG-SOLO). By adding a binary edge image guidance module and minimum cost allocation, the proposed network performs well in the aircraft image instance segmentation task, with advanced performance. First, we built a binary edge image guidance module (BEIGM), which combines the feature information of binary edge images and aircraft images, including a binary edge image information extraction module, top-down and bottom-up fusion module, and embedded into the output of the feature extraction network to enhance the network’s perception of the boundary feature information of the aircraft image. Secondly, we propose a positive and negative sample labelling strategy based on minimum cost allocation (MCA), which can calculate the sum of the classification and segmentation costs of all grids and mark the grid with the most minor combined cost as a positive sample to solve multiple Poor segmentation problem caused by overlapping aircraft instance objects. Finally, experimental results on the actual aircraft image and COCO dataset demonstrate our model’s advanced performance and robustness.