Fabric Defect Detection and Reduction Using Vision-Based Deep Learning Techniques
Samin Yeasar B.Sc in Textile Engineering Department of Apparel Engineering Campus: Textile Engineering College, Chattogram Bangladesh University of Textiles Email: saminyeasar123@gmail.com
Abstract
Fabric defect detection is a critical quality control task in textile manufacturing, as surface defects can reduce fabric value by 45–65% and significantly increase production costs. Traditional manual inspection methods typically achieve only 60–75% accuracy due to operator fatigue, limited inspection speed, and subjective judgment. Recent advances in computer vision and deep learning have enabled automated fabric inspection systems capable of real-time detection with accuracy exceeding 95%. This article presents an evidence-based analysis of fabric defect detection using experimental results reported in six peer-reviewed research papers. Instead of descriptive discussion, the article explains why deep learning models achieve high accuracy by linking network architecture, feature-learning mechanisms, dataset design, and optimization strategies directly to quantitative results. Detection methods, defect reduction strategies, and complete industrial solution procedures are discussed with proper reference indexing.
Fabric inspection is one of the most important quality assurance processes in textile manufacturing. During spinning, weaving or knitting, dyeing, and finishing operations, defects such as holes, broken yarns, missing picks, oil stains, needle lines, and color inconsistencies may occur. These defects directly affect fabric usability and market value.
Industrial investigations indicate that fabric defects can reduce the selling price of textile products by approximately 45–65% [1]. Manual fabric inspection is still widely used in many factories, where operators visually examine fabrics under strong illumination. However, manual inspection accuracy is typically limited to 60–75% due to fatigue, loss of concentration, and increasing production speed [2].
A real-time industrial study conducted in Ethiopian textile factories demonstrated that inspection accuracy depends heavily on operator mental condition and lighting environment, making consistent quality control difficult under high-speed production conditions [3]. To overcome these limitations, automated vision-based inspection systems have been introduced.
With the development of deep learning, convolutional neural networks and real-time object detection models now provide detection accuracies above 95%, offering reliable alternatives to traditional inspection systems.
This article addresses three research questions based strictly on experimental evidence reported in the provided literature:
How are fabric defects detected using vision-based systems?
What technical factors improve detection accuracy and reduce defects?
How can an effective industrial solution be implemented?
2. Traditional Fabric Defect Detection Methods
Early fabric inspection systems relied on classical image-processing techniques.
Statistical methods such as Gray Level Co-occurrence Matrix (GLCM), Local Binary Patterns (LBP), histogram analysis, and autocorrelation functions analyze variations between defective and non-defective texture regions. These techniques assume that fabric background texture remains statistically uniform.
However, experimental evaluation shows that statistical features fail when defects present smooth transitions or appear within complex patterned textures [2].
Spectral approaches using Fourier transforms, wavelet transforms, and Gabor filters analyze fabric textures in the frequency domain. Although effective for periodic fabrics, these methods require strict illumination control and are sensitive to noise. Erdogan and Dogan demonstrated that spectral methods degrade significantly under real-time industrial vibration and illumination variation [4].
Model-based methods, such as Gaussian mixture models, depend heavily on prior distribution assumptions and show limited adaptability to unseen defect types [2]. As a result, traditional approaches rarely exceed 85–90% detection accuracy in practical environments.
3. Deep Learning-Based Fabric Defect Detection
3.1 Hierarchical Feature Learning
Convolutional neural networks automatically learn hierarchical features from raw fabric images. According to Almeida et al. [2], shallow layers capture low-level information such as edges and yarn directions, intermediate layers model texture irregularities, and deeper layers encode semantic defect patterns.
This hierarchical learning mechanism enables CNNs to detect subtle texture disturbances that cannot be represented using handcrafted features.
3.2 Residual Learning and Texture Preservation
Deep residual networks improve detection accuracy by preventing feature degradation. Mewada et al. employed a ResNet50 architecture optimized using cuckoo search optimization [5]. Their experimental results showed a classification accuracy of 95.36% and an F1-score of 95.35%.
Residual skip connections preserve low-level texture information across deep layers, which is critical because fabric defects often represent minor disturbances rather than distinct objects.
3.3 Transfer Learning
Textile datasets are often limited in size and class balance. Transfer learning mitigates this limitation by using pretrained models trained on large-scale datasets.
Both Mewada et al. [5] and Durgaprasad et al. [6] reported faster convergence, improved generalization, and higher classification accuracy when pretrained ImageNet weights were used instead of training networks from scratch.
3.4 Attention Mechanisms
Fabric defects usually occupy a very small region of the image. Without guidance, neural networks may focus excessively on repetitive background texture.
Lu et al. integrated global attention and shuffle attention mechanisms into YOLOv8 [7]. Their enhanced model achieved a mean average precision of 96.6%, representing a 3.6% improvement over the baseline YOLOv8 network.
Attention mechanisms enhance defect saliency by amplifying informative features and suppressing irrelevant background regions.
3.5 Multi-Scale Feature Fusion for Tiny Defects
Tiny defects such as broken filaments and pinholes are difficult to detect due to limited pixel representation.
Yue et al. improved YOLOv4 by adding additional prediction layers, optimizing anchor boxes using k-means clustering, and integrating attention modules [8]. Their experiments demonstrated a 12% increase in average precision for tiny defect detection and a 3% improvement in overall mean average precision.
These results confirm that multi-scale feature fusion significantly enhances fabric defect localization.
3.6 Dataset Augmentation and Orientation Robustness
Fabric defects may appear at arbitrary orientations due to fabric skew and machine vibration.
Mewada et al. applied rotation, horizontal flipping, and vertical flipping to improve orientation robustness and reduce overfitting [5]. Erdogan and Dogan further confirmed that orientation-invariant feature extraction improves detection stability under real-time conditions [4].
3.7 False Negative Reduction
False negatives represent the most costly inspection error.
Almeida et al. demonstrated that their CNN-based system achieved approximately 75% accuracy in automatic mode, which increased to 95% after applying false-negative reduction strategies and limited operator verification [2]. This 20% improvement highlights the importance of decision-level optimization in inspection systems.
4. Fabric Defect Reduction Strategies
Experimental results across studies indicate that defect reduction depends on early-stage inspection, real-time monitoring, and process feedback.
Real-time vision systems allow immediate corrective actions, preventing defect propagation across production stages. Chaka et al. reported an average industrial inspection accuracy of approximately 89% using a CNN-based real-time vision system [3].
Defect classification statistics further enable root-cause analysis, helping identify machine malfunction, yarn quality issues, and process instability.
5. Industrial Solution Procedure
Based on validated experimental evidence, an effective fabric defect detection system follows the steps below:
Image acquisition using high-resolution industrial cameras and uniform illumination.
Dataset construction with labeled defect classes and augmentation.
Model selection based on application requirements.
Training using transfer learning and hyperparameter optimization.
Real-time deployment with continuous performance monitoring and retraining.
6. Conclusion
This article presented an analysis of fabric defect detection based on six peer-reviewed studies. Experimental evidence confirms that deep learning models outperform traditional inspection techniques due to hierarchical feature learning, residual connections, attention mechanisms, multi-scale feature fusion, dataset augmentation, and false-negative reduction strategies.
Modern vision-based systems consistently achieve detection accuracies between 95% and 97%, enabling reliable real-time inspection and significant reduction in textile production losses. The integration of intelligent inspection systems is therefore essential for sustainable, high-quality textile manufacturing.
7. References
[1] T. Almeida, F. Moutinho, and J. P. Matos-Carvalho, “Fabric Defect Detection With Deep Learning and False Negative Reduction,” IEEE Access, vol. 9, 2021.
[2] T. Almeida et al., “Fabric Defect Detection With Deep Learning and False Negative Reduction,” IEEE Access, 2021.
[3] K. T. Chaka, A. A. Shiferaw, and S. T. Sharew, “Inspection of Cotton Woven Fabrics Produced by Ethiopian Textile Factories Through a Real-Time Vision-Based System,” Journal of Natural Fibers, 2023.
[4] M. Erdogan and M. Dogan, “Enhanced Curvature-Based Fabric Defect Detection Using Gabor Transform and Deep Learning,” Algorithms, vol. 17, no. 506, 2024.
[5] H. Mewada et al., “Fabric Surface Defect Classification Using a Cuckoo Search Optimized Deep Residual Network,” Engineering Science and Technology, 2024.
[6] P. Durgaprasad et al., “Fabric Defect Detection Using Deep Learning,” IJRASET, 2023.
[7] G. Lu, T. Xiong, and G. Wu, “YOLO-BGS Optimizes Textile Production Processes,” Sustainability, vol. 16, 2024.
[8] X. Yue et al., “Research on Tiny Target Detection Technology of Fabric Defects Based on Improved YOLO,” Applied Sciences, vol. 12, 2022.
Founder & Editor of Textile Learner. He is a Textile Consultant, Blogger & Entrepreneur. Mr. Kiron is working as a textile consultant in several local and international companies. He is also a contributor of Wikipedia.
