YOLO-DLL: an effective detection method for fabric defects

Owing to the complex fabric structure and appearance variations, the diversity of defect types and scales, as well as the impact of lighting and environmental conditions on fabric appearance and detection results, automated fabric defect detection remains a challenging issue in the industry. To address the problems of complex background textures and limited hardware resources in online fabric defect detection, this paper proposes a high-precision and lightweight fabric defect detector: YOLO with dilation and lightweight attentive learning (YOLO-DLL). First, the dilation-wise residual module (DWR) is proposed to optimize the backbone network of YOLOv8. This module can capture multiscale contextual information from the input feature maps, and enhance the model's ability to integrate multiscale information. Second, the spatial pyramid pooling - fast (SPPF) module in the backbone network is optimized by using large separable kernel attention (LSKA). This enhancement can increase the network's focus on important features, and achieve a balance between high-precision detection and fast inference. Furthermore, to address the deployment of the detector on edge computing devices, the lightweight shared convolution detection (LSCD) head is designed. This LSCD head can use the scale layer to scale features, and further enhance the model's feature extraction capability while reinforcing its lightweight characteristics. Finally, extensive experiments were conducted on the large circular knitting fabric defect dataset and the fiberglass fabric defect dataset. The experimental results show that the proposed YOLO-DLL defect detection model achieves an accuracy of 99.2%, with 17.5% reduction in computational load and 28% increase in detection speed compared with the original model. The YOLO-DLL defect detection model not only improves the extraction of small target defect features and the ability to resist interference, but also achieves lightweight characteristics, which can meet the online detection needs of the textile industry for fabric defects.