Abstract: In infrared imaging, moving ship targets often lack clear intuitive information, such as size, shape, and texture, resulting in low detection accuracy and false detections of small targets in complex sea and sky backgrounds. Therefore, this study proposes an infrared ship detection method by improving YOLOv8. First, a hybrid convolution module was designed, and dynamic offset technology was used to adjust the sampling position to address the diversity of target shapes and better adapt to target characteristics at different scales by providing multiple convolution kernel parameters. Second, the SimAM parameter-free attention mechanism was introduced after SPPF to focus on key areas of the image while reducing computational complexity. The non-maximum suppression algorithm was then optimized to reduce missed detections caused by the suppression of small target detection boxes by large target detection boxes. Finally, the loss function was optimized by replacing the original loss function with EIoU to improve target localization accuracy and overall detection performance. Experimental results show that the improved algorithm achieved a 94.67% mAP on the infrared ship dataset, representing an increase of 3.83% compared with the original mAP. Compared with other classic algorithms, each evaluation index was improved to varying degrees, which demonstrates that the improved algorithm achieves better detection performance and can satisfy the real-time detection task of infrared targets.