Ship target detection algorithm based on improved YOLOv8

Ship detection is essential for ensuring maritime safety, enabling intelligent shipping, and promoting marine conservation. This paper introduces an improved algorithm utilizing the YOLOv8 framework, emphasizing advancements in feature extraction, context capturing, and loss function optimization. Firstly, by incorporating a mixed pooling strategy in the SPPF module that combines max pooling and average pooling, the algorithm improves feature robustness and increases the receptive field, facilitating multi-scale detection of ship targets. Secondly, the implementation of a Global Attention Mechanism (GAM) allows for better capture of 18 global contextual information, thus enhancing both detection accuracy and computational efficiency. Lastly, the conventional intersection over Union metric in the loss function is replaced with Wise IoU v2, shifting focus from simpler to more challenging examples, which boosts classification performance significantly. Experimental results reveal a 6.5% increase in mAP@50 and achieving 92.3% mAP compared to YOLOv5, along with a detection speed improvement of 81 frames. When compared to YOLOv8, the proposed algorithm has been demonstrated to achieve a 3.5% improvement in mAP@50 and a 10-frame increase in detection speed.