Enhanced YOLOv8 Method for Real-Time Detection of Complex Fabric Defects

Detecting fabric defects in the context of diverse fabrics, textures, and colors has always been a significant and challenging problem. Additionally, the scarcity of datasets in industrial applications limits the generalization and robustness of models, particularly in detecting small defects. To address these issues, this paper proposes an improved YOLOv8 method for textile defect detection. First, a weighted feature fusion method is used to enhance the network's focus on critical information, effectively addressing challenges in detecting small defects. To further improve detection performance, we introduce the DGSCSP feature fusion module, which integrates multi-level feature information, reduces redundancy from irrelevant data, and decreases computational complexity. Furthermore, the original bounding box loss function is replaced with an angle loss function to improve training stability. Finally, reparameterization techniques are applied to enhance model performance without increasing inference costs, maintaining the same model architecture. Experimental results on the Tianchi fabric dataset, detecting 24 defect types, show that the proposed algorithm achieves 92.2% precision, 82.8% recall, and 90.5% mAP50, with an inference time of only 2.2 milliseconds. These results demonstrate the exceptional performance and efficiency of the proposed algorithm in textile defect detection.