DLPW-YOLO: A Lightweight Fabric Defect Detection Model Based on the Improved YOLOv8n

Traditional fabric defect detection methods have shortcomings such as lower accuracy, high computational cost and insufficient practical suitability. In this work, a lightweight fabric defect detection model (DLPW-YOLO) is proposed based on the improved YOLOv8n. First, more accurate target localization is achieved by replacing the original C2f module with the C2f-DCNv4 module. It facilitates adaptive localization of convolutional kernel sampling. Second, the LSKA attention mechanism is integrated into the SPPF module. This modification enhancesmultiple-scale feature aggregation and fine details capturing capability. Third, the PConv convolution is introduced to the detection head to reduce size of training parameters and speed up defect detection. At last, Wise-IoUv3 loss function is employed to accurate track bounding box by taking advantage of its dynamic non-monotonic focusing capability. The proposed model is validated on four public defect fabric datasets. The experimental results demonstrated that DLPW-YOLO increased [email protected] by 2.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, and computational complexity and parameter count were reduced by 31.7% and 13.4% respectively compared to baseline YOLOv8n. The detection speed achieved 142 fps, which meets real-time requirement of fabric defect detection in textile production.