基于边缘重校准的多尺度特征织物疵点检测

Multi-Scale Feature Fabric Defect Detection Based on Edge Recalibration

中文关键词: 织物疵点检测  YOLO11  边缘重校准  多尺度特征融合  深度学习

英文关键词: fabric defect detection  YOLO11  edge recalibration  multi-scale feature fusion  deep learning

中文摘要:

      针对织物疵点检测中疵点尺度多样、边缘模糊及背景纹理复杂等问题，导致传统检测方法精度不足的挑战，文章在 YOLO11n 网络基础上提出改进模型 MERS-YOLO11。该模型引入 CSP-Multi Scale Edge Information Enhancement 模块增强边缘特征提取能力；设计 Recalibration FPN 结构实现高低层特征双向融合；扩展 P2 (1/4) 检测头提升微小疵点检测能力；采用 EMA-Slide Loss 损失函数解决样本不平衡问题。实验结果表明，改进模型 mAP@0.5 达 54.1%，较基准模型 YOLO11n 提升 5.7%，参数量仅为 3.74 M。消融实验和可视化结果证实了该模型在微小疵点检测和复杂织物背景下的优越性能。

英文摘要:

      Aiming at the challenges of insufficient precision in traditional detection methods due to diverse defect scales, blurred edges, and complex background textures in fabric defect detection, this paper proposes an improved model MERS-YOLO11 based on the YOLO11n network. The model introduces a CSP-MultiScaleEdgeInformationEnhance module to enhance edge feature extraction capability; designs a Re-CalibrationFPN structure to achieve bidirectional fusion of high and low-level features; extends the P2(1/4) detection head to improve small defect detection ability; and adopts EMASlideLoss function to solve the sample imbalance problem. Experimental results show that the improved model achieves 54.1% mAP@0.5, an increase of 5.7 percentage points compared to the baseline model YOLO11n, with a parameter count of only 3.74 M. Ablation experiments and visualization results confirm the superior performance of this model in detecting tiny defects and in complex fabric backgrounds.