Patterned Fabric Defect Detection Based on Separate Convolutional Unet

The textile industry demands high-quality fabric production, necessitating automated inspection systems to enhance quality control. This study utilizes the ZJU-Leaper dataset, con-sisting of 28,739 images ranging from plain to floral-patterned fabrics, to evaluate an adapted version of the SCUNet architecture for semantic segmentation. Unlike previous research that primarily used grayscale images, our focus is on colored and patterned fab-rics, providing a more comprehensive analysis. Our experiments demonstrate that the proposed SCUNet architecture is highly effective in segmenting input images, achieving mean Intersection over Union (mIoU) of 71.69% and Dice Coefficient of 55.60%. These metrics reflect the model's ability to accurately segment defective regions in fabrics. Besides evaluating performance, this study addresses suggestions from prior research by extending the application of SCUNet to colored images and more complex fabric types. The results indicate that SCUNet can improve defect detection processes and accurately segment defect areas, even when the defects closely resemble the background. SCUNet also successfully detects small defects that are difficult for humans to identify. This research makes a significant contribution to developing robust, efficient, and scalable solutions for automated fabric inspection. By focusing on complex, colored, and patterned fabrics, our study provides valuable insights for real-world applications, ultimately benefiting the textile industry through improved quality control and cost efficiency. Advances in this inspection technology mark a significant step toward achieving more reliable and effective fabric defect detection.