YOLO-GFD: A Fast and Accurate Fabric Defect Detection Model

Fabric defect detection plays a pivotal role in ensuring quality management within the textile industry, enabling timely identification and remediation of defects. This research addresses key challenges in fabric defect detection, including slow detection speed, significant variations in defect scale, and the difficulty in detecting small defects. To tackle these challenges, this study presents the “Yolo-GFD” fabric defect detection model, a refined version of YOLO5, designed to strike a harmonious balance between detection speed and accuracy. Firstly, “G-bottleneck” is utilized instead of the original bottleneck module in the backbone and neck network, which greatly reduces the parameters of the model and improves the detection speed. In addition, we design a new feature fusion network that better fuses features across scales and effectively enhances the detection of small targets. Finally, we use the k-means algorithm to design a priori frames suitable for fabric blemish detection and introduce the feature enhancement module “DSE-Block” behind the backbone network to make our detection model more adaptable to the scale variation of fabric blemishes. Experimental results on the dataset used in this paper show that compared with the original algorithm, the mean average precision (map) is dramatically improved by 6.72%, and the detection speed reaches 22. 39 fps, which effectively meets the requirements of real-time defect detection in industrial environments.