SAR Aircraft Target Detection Based on Improved YOLOv8

Aiming at the problems of weak generalization ability and poor detection effect caused by strong background interference and small size of aircraft target in SAR (Synthetic Aperture Radar) images, this paper proposes an aircraft target detection algorithm for SAR images based on improved YOLOv8. In view of the shortcomings of YOLOv8 network in small target detection, the Simplified-YOLOv8 network is proposed. Combined with the advantages of DCN (Deformable Convolutional Networks), a C2f_DCNv3 module is proposed to improve the feature extraction ability and anti-interference ability of the model. EMA (Efficient Multi-Scale Attention) is introduced into the neck network, so that the algorithm can retain more feature information in the stage of feature fusion. Inner-CIoU Loss is used as the loss function to improve the convergence speed and regression accuracy. The experimental results on the SAR-AIRcraft-1.0 dataset show that the proposed algorithm achieves a 69.4% reduction in parameter quantity and a 25.9% reduction in computation. The precision rate, recall rate and mAP reach 86.3%, 90.8% and 95.2% respectively. The proposed algorithm greatly reduces the model complexity while ensuring high detection effect. Compared with other popular algorithms, the proposed algorithm is more suitable for aircraft target detection in SAR images.