TL;DR
This paper introduces RPDNet, a novel deep learning approach for automatic fabric defect detection. By combining a repeated pattern analysis algorithm with a semantic segmentation network (specifically DeepLabV3+ with GhostNet backbone), RPDNet effectively identifies pixel-level defects in textiles, even in complex, repeating patterns, without sacrificing detection speed.
What Problem Does This Paper Address?
Quality control in the textile industry heavily relies on detecting defects in fabrics. Manual inspection is prone to human error and fatigue, leading to inconsistent accuracy (60-75%). While automated methods exist, they often struggle with the intricate and repeating patterns common in textiles. Traditional methods rely on hand-crafted features, which are difficult to generalize to new designs or environments. Deep learning offers promise, but existing models aren't specifically designed to leverage the inherent periodic nature of many fabric textures, leading to challenges in accurately identifying anomalies within these patterns. This paper aims to bridge this gap by developing a method that explicitly incorporates knowledge of fabric periodicity into a deep learning framework for more robust and efficient defect detection.
Methods
The core of RPDNet lies in its two-stage approach:
- Repeated Pattern Analysis Stage: This stage uses a Convolutional Neural Network (CNN) to detect periodic patterns within fabric images. It identifies "activation peaks" in feature maps from different network layers and uses a Hough voting space to determine the most likely displacement vector, representing the size of the repeated pattern. This information is then used to generate "repeated pattern informative labels."
- Semantic Segmentation Stage: This stage employs a semi-supervised learning scheme. Instead of directly feeding the pre-calculated pattern labels into the main network, RPDNet uses a set of parallel DeepLabV3+ networks with GhostNet backbones. One network takes the original fabric image, while others take the fabric image and the generated repeated pattern labels. This allows the network to learn to segment defects and understand periodic features simultaneously. Crucially, during training, gradients are updated across all networks, but after training, only the necessary components for defect detection are retained, ensuring no additional computational overhead during inference. The use of DeepLabV3+ and GhostNet contributes to a lightweight and efficient architecture.
Datasets
The paper evaluates RPDNet on two publicly available datasets:
- Fabric Images (FI) database: Contains 106 images with six defect types (broken end, hole, knot, netting multiple, thick bar, and thin bar).
- TILDA Textile Texture Database: Includes eight common textile types. For this study, three types with repeated textures were selected, resulting in 750 images with four defect types (hole, color spot, thread error, and foreign body).
All images were manually labeled at the pixel level using one-hot encoding.
Key Insights and Ideas
- Leveraging Periodicity: The fundamental insight is that the periodic nature of fabric textures is a valuable cue for defect detection. By explicitly modeling and injecting this periodic information, the network can better distinguish between normal patterns and anomalies.
- Semi-Supervised Learning for Efficiency: The proposed semi-supervised approach is key to avoiding computational bottlenecks. Instead of requiring pre-calculated pattern information at inference time, the knowledge is learned during training, allowing the final model to operate independently without extra processing.
- Lightweight Architecture: The integration of DeepLabV3+ and GhostNet demonstrates a focus on achieving high performance with reduced computational cost, making the method suitable for practical industrial applications.
- Generalizability: The RPDNet framework is presented as a general strategy that can be extended to various mainstream semantic segmentation architectures, suggesting its broad applicability.
Key Results
RPDNet achieved state-of-the-art performance on both the FI and TILDA datasets.
- Quantitative Performance: On average, RPDNet outperformed the second-best method by significant margins: 7.9% in IoU, 5.7% in Recall, 2.9% in Precision, and 4.4% in F1-Measure across both datasets.
- Specific Metrics: For the FI dataset, RPDNet achieved an IoU of 0.766, Recall of 0.833, Precision of 0.905, and F1-Measure of 0.867. For the TILDA dataset, it achieved an IoU of 0.831, Recall of 0.886, Precision of 0.930, and F1-Measure of 0.907.
- Speed: The average detection time on a GTX 2080Ti GPU was 6.59 ms, enabling real-time detection (exceeding 150 FPS).
- Qualitative Results: Visual comparisons showed RPDNet producing segmentation contours that were closest to the ground truth, with fewer errors compared to other methods.
Limitations and Future Work
- Repeated-Pattern-Free Images: The paper acknowledges that RPDNet struggles with images that lack clear repeating patterns. In such cases, the repeated pattern detector fails to produce effective responses, and the method's performance is only marginally better than baseline models.
- Future Work: The authors suggest that extending the applicable range of the method to handle non-repeating patterns will be of paramount importance. This could involve further research into how to adapt the approach for a wider variety of fabric textures.
Summary updated