Improved Deformable Convolution Method for Aircraft Object Detection in Flight Based on Feature Separation in Remote Sensing Images

Aircraft object detection in remote sensing images is a challenging task, especially for small objects, complex backgrounds, and aircraft objects in flight. Due to the lack of contextual relationship between the aircraft object and the surrounding background during flight and the small number of pixels in the object itself, the use of regular rectangular convolution for feature extraction results in many background pixels being sampled. This article proposes YOLO-FRS for aircraft object in flight detection, which uses a new module based on deformable convolution (DCN), the feature response separation deformable convolution (FRS-DCN) module. The FRS-DCN module adds semantic segmentation supervision on the basis of stacked DCNs, so that the background and object in the input of DCNs in this module are separated as much as possible. We design a soft label annotation and loss calculation method for semantic segmentation supervision. In addition, we propose a flight-state aircraft object dataset that includes multiple backgrounds and cloud interference. YOLO-FRS was tested on the proposed dataset, and the results showed that the FRS-DCN module improved the performance of aircraft object detection. Compared with multiple mainstream deep convolution models, the YOLO-FRS model also exhibits competitive performance.