YOLOv8-EVR: An Enhanced Algorithm Based on YOLOv8 for SAR Dataset

The images in SAR dataset exhibit high intraclass similarity and inter-class variability, which often leads to visual confusion during manual inspection, resulting in low efficiency and a high error rate. Existing methods are prone to missed detections and false detections when detecting multi-scale objects in complex backgrounds while some methods require a lot of computation. To enhance the accuracy and efficiency of target recognition, a model named YOLOv8-EVR is proposed which is based on the YOLOv8 framework. In this model, the VanillaBlock and the RepViTBlock are successively employed in Head section to improve small-target detection and reduce model complexity, and the effective Squeeze and Extraction (eSE) is used to extract as much channel information as possible in a simpler way. On sar-aircraft-1.0 dataset, the precision rate, recall rate and mAP@50 reach 90.4%, 82.3% and 90.6% respectively. Compared with the YOLOv8 model, the accuracy is improved by 3.3% and the mAP@50 is improved by 1.3%.