FCCS-YOLO: Improved YOLOv8 with Contrastive Learning for Aircraft Detection in SAR Images

Compared to natural images, aircraft targets in Synthetic Aperture Radar (SAR) images are typically smaller and surrounded by complex backgrounds. Directly applying object detection algorithms based on optical images to SAR images often results in poor detection accuracy and inaccurate target localization. One of the reasons for this issue is that current networks do not fully utilize the semantic information between the target and background in the feature space. To address this problem, this paper proposes the FCCS-YOLO algorithm based on YOLOv8. Firstly, FCCS-YOLO optimizes the target detection layers of YOLOv8 to make it more suitable for small and medium-sized aircraft targets in SAR images. Secondly, the Conv-Passthrough-DSC(CPD) module is proposed as the primary downsampling structure to address the reduced receptive field caused by adjustments in the feature detection layers and to overcome the limitations of current mainstream downsampling structures. Thirdly, the Skew Intersection over Union (SIOU) loss function is introduced to further enhance the bounding box regression capability. Finally, a contrastive learning regularization method is proposed for aircraft detection in SAR images. This method not only addresses the impact of feature consistency on bounding box regression but also improves the network’s ability to perceive feature differences across categories. Experimental results show that the FCCS-YOLO model performs excellently across multiple evaluation metrics.