Abstract: Aiming at the problems of blurred target features, complex background and missed detection of small targets in infrared ship images, a detection algorithm YOLO-IST (YOLO for infrared ship target) for ship targets in maritime traffic was proposed based on YOLOv8. Firstly, the C2f_DBB module and CPCA attention mechanism were introduced into the backbone network of the baseline model, and the recognition ability of the model to the target was improved by adding the feature extraction layer. Then, the C2f_Faster_EMA module was used to replace the C2f module in the neck network to improve the detection accuracy and speed of model. Finally, the multi-attention dynamic detection head, that is Dynamic Head, was used to optimize the model framework and enhance the detection effect of the model to small ship targets. The experimental results show that Recall, Precision, Map@50、Map@50-95 and F1score of YOLO-IST are 89.7%, 90.5%, 94.7%, 66.6% and 90.1%, respectively, which are improved by 4.5%, 3.8%, 4.4%, 4.7% and 4.2%, respectively, compared to the baseline model YOLOv8.The proposed model has a wide application prospect in maritime intelligent transportation.

Key words: traffic and transportation engineering; ship engineering; infrared target detection; YOLOv8; attention mechanism

摘要： 针对红外船舶图像目标特征模糊、背景复杂以及小目标漏检等问题，基于YOLOv8提出一种面向海上交通中船舶目标的检测算法YOLO-IST(YOLO for infrared ship target)。在基线模型的骨干网络中引入C2f_DBB模块和CPCA注意力机制，通过增加特征提取层来提升模型对目标的识别能力；利用C2f_Faster_EMA模块替换颈部网络中的C2f模块，以提升模型检测精度和速度；采用多重注意力的动态检测头Dynamic Head优化模型框架，增强模型对小船舶目标的检测效果。研究结果表明：YOLO-IST 的召回率Recall、精确率Precision、平均精度Map@50、平均精度Map@50-95和F1score分别达到89.7%、90.5%、94.7%、66.6%、90.1%，较基线模型YOLOv8分别提升了4.5%、3.8%、4.4%、4.7%、4.2%。该模型的提出在海上智能交通中具有较广泛的应用前景。

关键词: 交通运输工程；船舶工程；红外目标检测；YOLOv8；注意力机制