An improved lightweight ship detection model based on YOLOv8

In recent years, the progress in the field of deep learning has been nothing short of extraordinary, prompting an escalating trend in research dedicated to the identification of ships in satellite imagery. This technology is essential for various maritime applications, including safety at sea, rescue operations, and the management of shipping fleets. Consequently, there is a significant demand for models that can deliver high precision, operate in real-time, and are adaptable to different scenarios. This study introduces a refined version of the algorithm, drawing inspiration from YOLOv8—YOLOv8-SD. We have enhanced the Backbone, Neck, and Head parts of the YOLOv8 model, achieving lightweight while enhancing the model's detection precision. The specific improvements will be explained in the methods section. The dataset utilized in this paper is the ship part of WHU-RSONE-OBB. After 300 training epochs, compared with YOLOv8n, GFLOPs decreased from 8.7 to 6.9, and mAP50 increased from 89.7 to 94.8.