中图分类号：TP3文献标识码：ADOI:10.19339/j.issn.1674-2583.2026.04.007 中文引用格式：于邈，孙俊，彭逸飞，等. 基于改进YOLOv13的纺织品瑕疵检测方法分析［J］.集成电路应用，2026，43（4）：39-44. 英文引用格式：Yu Miao， Sun Jun，Peng Yifei,et al. Research on textile defect detection method based on improved YOLOv13［J］.Application of IC，2026，43（4）：39-44.
Research on textile defect detection method based on improved YOLOv13
Yu Miao，Sun Jun，Peng Yifei, Lu Zhiqiang
School of Electrical and Information Engineering, Jiangsu University
Abstract： To address the issues of strong interference from complex texture backgrounds, weak features of small targets, and the loss of detail information during downsampling in textile defect detection, this paper proposes a textile defect detection model named WHA-YOLO based on an improved YOLOv13. In the backbone network, a Wavelet Residual Downsampling Module (WRDM) is introduced to preserve edge and texture details. In the multi-scale feature enhancement stage, a Hypergraph-based Context Enhancement Module (HyperCEM) is designed to improve long-range context modeling capability. In the neck network, an Adaptive Feature Modulation and Aggregation Module (AFMAM) is designed to achieve joint enhancement of spatial details and channel relationships. Experimental results show that the proposed WHA-YOLO achieves 94.2% mAP@0.5 on a mini-public dataset containing four types of defects, which is 3.4 percentage points higher than the baseline model YOLOv13n, and that on a public dataset containing ten types of defects, it achieves 75.1% mAP@0.5, which is also 2.3 percentage points higher than the second-best YOLOv13s method.
Key words : textile defect; object detection; YOLOv13; deep learning; feature enhancement
