DCF-YOLO: a context-aware lightweight YOLO with mixed attention and dynamic pruning for real-time yarn defect detection

Abstract To address the dual challenges of high-precision detection and computational efficiency in real-time yarn quality detection for textile manufacturing, this paper proposes DCF-YOLO, an enhanced lightweight detection framework based on YOLOv8. Our approach introduces four key innovations: First, we develop DMAConv through integrating a Mixed Local-Channel Attention (MLCA) mechanism with Depthwise Convolution (DWConv), effectively reducing model parameters while enhancing multi-scale feature extraction. Second, we propose the C2f-CG module by incorporating contextual awareness and optimizing channel attention mechanisms, improving target-background discrimination through spatial-contextual feature fusion. Third, we design a Feature-Enhanced Lightweight Detection Head (FELDH) that replaces conventional multi-head detection with shared FEConv modules, achieving computational reduction while maintaining detection accuracy. Finally, we implement Layer-adaptive Magnitude-based Pruning (LAMP) with a 1.5 pruning ratio to optimize model compression. Extensive experiments indicate that DCF-YOLO significantly outperforms the YOLOv8 baseline, achieving improvements of 0.6% in precision, 0.8% in recall, and 1.9% in mAP. Moreover, DCF-YOLO reduces computational complexity by 70.3% and the number of parameters by 89.9%. Practical deployment on industrial inspection systems with six spinning positions shows significant efficiency gains for a spinning system with 12 nozzles, the average detection time per position was shortened by 4.8 s to 37.88 s. This work provides an effective solution for real-time textile quality control with balanced accuracy-efficiency tradeoffs.