DTM-YOLO: a texture-regularity-aware YOLO framework for textile defect detection under complex backgrounds

Aug 27, 2026

Upload by Maeve

Textile defect detection remains challenging because repetitive fabric textures often dominate feature responses, making subtle, elongated, and small-scale defects hard to distinguish from complex backgrounds. To address this issue, the paper presents DTM-YOLO, a texture-regularity-aware YOLO framework built on YOLOv11. It introduces a Texture Suppression Attention Module for residual-guided channel recalibration that enhances defect-sensitive signals while suppressing texture-dominant interference. Dynamic Snake Convolution in C3k2 blocks improves continuity for elongated defects, and a multi-scale fusion neck preserves shallow details for small defects.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Introduction

Textile defect detection challenges

Proposed DTM-YOLO contributions

Related works

Signal, Image and Video Processing (2026) 20:547 [https://doi.org/10.1007/s1](https://doi.org/10.1007/s1) 1760-026-05592-z DTM-YOLO: a texture-regularity-aware YOLO framework for textile defect detection under complex backgrounds Yuheng Pan1 · Chao Fan1 · Weijia Lu1 · Jia Cong1 · Guoyan Li1 Received: 29 December 2025 / Revised: 9 July 2026 / Accepted: 15 July 2026 / Published online: 3 August 2026 © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2026 Abstract Textile defect detection remains challenging because repetitive fabric textures often dominate feature responses, making subtle, elongated, and small-scale defects difﬁcult to distinguish from complex backgrounds. To address this problem, this paper proposes DTM-YOLO, a texture-regularity-aware detection framework based on YOLOv11 . Its core component is a Texture Suppression Attention Module, which models local statistical deviations from periodic texture patterns and performs residual-guided channel recalibration to enhance defect-sensitive responses while suppressing texture-dominant background features. Dynamic Snake Convolution is embedded into C3k2 blocks to improve continuity modeling of elongated defects, and a multi-scale assisted feature fusion neck is adapted to preserve shallow spatial details for small defect detection. The proposed framework is evaluated on two datasets: a self-constructed textile defect dataset and the public Tianchi fabric defect dataset. On the self-constructed dataset, DTM-YOLO improves mAP@0.5, mAP@0.5:0.95, and precision by 4.9%, 4.02%, and 5.7%, respectively, compared with YOLOv11n. Generalization results on the Tianchi dataset further indicate improved robustness under external textile data. Keywords Textile defect detection · YOLOv11 · Texture regularity · Texture suppression attention · Dynamic snake convolution · Multi-scale feature fusion 1 Introduction Fabric surface defect detection has gradually become an important research topic in both academia and industry. Traditional manual inspection methods rely heavily on inspectors’ experience and concentration. However, they suffer from high labor intensity, high omission rates, low reliability, and poor generalization, with defect detection accuracy typically ranging from 60 to 75% and a maximum inspection speed of only 12m per minute [1, 2] . Visual inspection is also affected by fatigue, experience, and subjective judgment [3, 4] . Since fabric defects can reduce product value by approximately 45–65%[5], manual inspection is increasingly being replaced by automated detection technologies. In recent years, deep learning-based object detection has become a key solution for automated textile quality inspection. B Weijia Lu[luweijia@tcu.edu.cn](luweijia@tcu.edu.cn) 1 School of Computer and Information Engineering, Tianjin Chengjian University, Tianjin 300384, China TheYOLO serieshasbecome a mainstream choice fortextile defect detection because of its end-to-end single-stage detection mechanism and high real-time performance [6] . Compared with two-stage detectors such as Faster R-CNN, YOLO simultaneously performs object localization and classiﬁcation in a single forward pass, reducing computational cost and supporting fast industrial deployment. Nevertheless, textile defect detection still faces several challenges. First, normal fabric regions usually contain periodic or quasiperiodic textures, and such repetitive patterns may dominate feature responses. Second, defects such as Thread error and Ribbon yarn are slender, elongated, and direction-varying, which requires continuous shape modeling. Third, small or low-contrast defects are easily weakened during deep feature extraction and multi-scale fusion. In this study, we propose DTM-YOLO (Dynamic-convolution and Texture-suppressionMulti-scale Attention YOLO) to enhance textile defect detection under complex backgrounds. The main contributions are summarized from t

FAQs

Why are textile defects difficult to detect in complex backgrounds?

Repetitive or quasi-periodic fabric textures often dominate feature responses, and subtle defects such as slender elongated thread/ribbon yarn and small or low-contrast flaws are easily weakened during deep feature extraction and fusion.

What is the main idea behind DTM-YOLO?

DTM-YOLO enhances YOLO-based detection by explicitly modeling texture regularity, suppressing periodic background interference, adapting shape continuity modeling for elongated defects, and preserving shallow details for small defects.

How does DTM-YOLO improve detection accuracy and robustness?

Experiments on a self-constructed dataset show improvements in mAP@0.5, mAP@0.5:0.95, and precision over YOLOv11n, and results on the Tianchi dataset indicate better robustness under external textile data.

Aug 27, 2026

Upload by Maeve