Scholarlink

<p>To address the inherent trade-off between detection accuracy and real-time performance in the intelligent nondestructive testing of small rail surface defects under complex scenarios, this paper proposes MRM-YOLO, a lightweight, high-precision detection algorithm based on the YOLOv5s architecture. First, a Mixed-domain Modulation (MM) mechanism is introduced into the backbone network to effectively suppress non-homologous background noise through adaptive frequency-domain filtering while modulating spatial features. Subsequently, a Morphological Re-parameterization (MR) module is designed in the feature fusion network to directionally enhance minute defect features via asymmetric convolutions, achieving near-zero extra inference latency through structural re-parameterization. Additionally, a novel AF-MPDIoU loss function is integrated to accelerate regression convergence for targets with extreme aspect ratios via an adaptive gradient re-weighting strategy. Experimental results on a self-constructed rail defect dataset demonstrate that MRM-YOLO achieves a mean average precision (mAP@0.5) of 93.4% and an inference speed of 154.8 FPS. Compared to the YOLOv5s baseline and existing methods, the proposed algorithm effectively balances detection accuracy and computational efficiency, successfully breaking the strict trade-off bottleneck between extreme lightweight inference and high-recall safety requirements, thereby providing a highly reliable and real-time solution for practical railway inspection environments.</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

MRM-YOLO: An Intelligent Identification and Nondestructive Testing Method for Small Defects on Rail Surfaces

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Jialong Fu,
  • Jianxin Zhou

摘要

To address the inherent trade-off between detection accuracy and real-time performance in the intelligent nondestructive testing of small rail surface defects under complex scenarios, this paper proposes MRM-YOLO, a lightweight, high-precision detection algorithm based on the YOLOv5s architecture. First, a Mixed-domain Modulation (MM) mechanism is introduced into the backbone network to effectively suppress non-homologous background noise through adaptive frequency-domain filtering while modulating spatial features. Subsequently, a Morphological Re-parameterization (MR) module is designed in the feature fusion network to directionally enhance minute defect features via asymmetric convolutions, achieving near-zero extra inference latency through structural re-parameterization. Additionally, a novel AF-MPDIoU loss function is integrated to accelerate regression convergence for targets with extreme aspect ratios via an adaptive gradient re-weighting strategy. Experimental results on a self-constructed rail defect dataset demonstrate that MRM-YOLO achieves a mean average precision (mAP@0.5) of 93.4% and an inference speed of 154.8 FPS. Compared to the YOLOv5s baseline and existing methods, the proposed algorithm effectively balances detection accuracy and computational efficiency, successfully breaking the strict trade-off bottleneck between extreme lightweight inference and high-recall safety requirements, thereby providing a highly reliable and real-time solution for practical railway inspection environments.