Scholarlink

<p>The absence of visual details in infrared images of damaged ships often increases the difficulty of detection. To alleviate this issue, YOLO-IMS (Improved Marine Ship detector), an enhanced model based on YOLOv5s is proposed. Firstly, the efficient spatial multi-scale attention (ESMA) module is designed to improve the feature representation of infrared damaged ships. The module processes the input tensor through two parallel branches, performs feature extraction and weighting, and then aggregates the output through cross-spatial learning. An additional branch is added to capture long-range dependencies, enhancing spatial feature extraction capabilities. Secondly, the original CIoU Loss is substituted with the normalized gaussian wasserstein distance (NWD) Loss to accelerate model convergence during training. Finally, the spatial pyramid pooling fast plus (SPPFP) module is integrated to improve local feature fusion for infrared damaged ships without increasing the number of model parameters. Experimental results show that YOLO-IMS maintains the efficiency of YOLOv5s while achieving 77.0% mAP@0.5 on self-established infrared damaged ship dataset-an improvement of 10.5% over the baseline. The model also performs well on the standard IRay infrared human-vehicle dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

YOLO-IMS: an improved method for infrared damaged ship detection

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Wang Jianyuan,
  • Han Shu,
  • Chen Jinbao,
  • Zhang Yue,
  • Shi Donghao

摘要

The absence of visual details in infrared images of damaged ships often increases the difficulty of detection. To alleviate this issue, YOLO-IMS (Improved Marine Ship detector), an enhanced model based on YOLOv5s is proposed. Firstly, the efficient spatial multi-scale attention (ESMA) module is designed to improve the feature representation of infrared damaged ships. The module processes the input tensor through two parallel branches, performs feature extraction and weighting, and then aggregates the output through cross-spatial learning. An additional branch is added to capture long-range dependencies, enhancing spatial feature extraction capabilities. Secondly, the original CIoU Loss is substituted with the normalized gaussian wasserstein distance (NWD) Loss to accelerate model convergence during training. Finally, the spatial pyramid pooling fast plus (SPPFP) module is integrated to improve local feature fusion for infrared damaged ships without increasing the number of model parameters. Experimental results show that YOLO-IMS maintains the efficiency of YOLOv5s while achieving 77.0% mAP@0.5 on self-established infrared damaged ship dataset-an improvement of 10.5% over the baseline. The model also performs well on the standard IRay infrared human-vehicle dataset.