FL-YOLOv8: a SAR ship detection model based on improved YOLOv8

Synthetic Aperture Radar (SAR) imagery finds extensive applications in both military and civilian domains due to its inherent advantages, such as all-weather capability, high resolution, and complete coverage. However, SAR images encounter several limitations, including unclear edge profile information, multi-scale representation, high sparsity, and a high percentage of small target ships. Consequently, these factors contribute to relatively low accuracy, poor model positioning capabilities, and difficulty in feature extraction in target detection. To overcome this limitation, the present study introduces a novel SAR ship detection method, FL-YOLOV8. It enhances the 160X160 detection feature map by incorporating FHP(four-head prediction) to identify targets larger than 4X4 and replaces the original detection head with LSCDH(lightweight shared convolutional detection head). First, due to the relatively large downsampling multiples in yolov8, it becomes challenging to capture the feature information of small targets. By incorporating a feature head, it becomes feasible to integrate shallow and deep feature information. Second, LSCDH enhances the feature representation of the model, accommodates inputs at various scales, and minimizes both the number of parameters and computational effort. Furthermore, comprehensive experiments conducted on the benchmark dataset HRSID demonstrate the superior performance of FL-YOLOV8 in ship detection, achieving an accuracy of 92.8%.