DGL-YOLOv8 a Lightweight Yarn Detection Algorithm

Yarn, being a fundamental raw material in fabric manufacturing, directly influences the quality of the final product. To address potential challenges that may occur at various stages of yarn production, in this paper, we propose a lightweight and efficient recognition algorithm based on YOLOv8n, named DGL-YOLOv8. First, we enhanced the lightweight convolution DWConv by incorporating the MLCA attention mechanism to strengthen the connection between channel and spatial information. Then, we introduced the GhostModule lightweight module to refine the C2f structure, replacing the standard convolution layers with dynamic convolution layers. This allows the model to better capture features of objects at different scales. Finally, in the detection head, we adopted the concept of shared parameters, replacing the bulky three Detect modules of the original model with a single shared convolution, further reducing the model's parameters and computational cost. According to the experimental data, DGL-YOLOv8 achieved an mAP0.5 value of 90.1%, an improvement of 0.5% over YOLOv8n. The parameter count is 0.95M, and the computational load is 3.3 GFLOPs, representing reductions of 68% and 59%, respectively, compared to the base model. The results demonstrate that the improved model greatly increases the effectiveness of real-time object detection in complex environments, providing an effective solution for yarn production monitoring.