Research on improved YOLOv8 fabric surface defect detection algorithm

To address the limitations of traditional fabric defect detection algorithms, such as low accuracy in complex backgrounds and high rates of missed and false detections, this paper proposes an improved detection method based on the YOLOv8 architecture. The backbone network is enhanced by integrating a mixed local channel attention (MLCA) mechanism into the C2f module, resulting in the C2f_MLCA module, which captures both spatial and channel features while fusing local and global information, thereby enhancing attention to defect regions and preserving critical details. To handle multiscale defects and complex background interference, a context-guided spatial feature reconstruction feature pyramid network (CGRFPN) is introduced in place of the original neck, improving detail extraction through dynamic interpolation and multifeature fusion. Furthermore, a task-aligned dynamic detection head (TADDH) is proposed to alleviate spatial misalignment and weak task interaction between classification and localization branches. The Focaler-IoU loss function is also employed to address sample imbalance during regression and improve localization precision, particularly for small defects. Experimental results show that the proposed method outperforms the baseline YOLOv8 by 2.9%, 6.8%, and 2.8% in terms of accuracy, recall, and mAP0.5, respectively, reaching 88.6%, 89.2%, and 91.1%, while achieving 3.2%, 2.6%, and 4.5% gains on the validation dataset. In addition, the algorithm achieves a real-time inference speed of 86.91 FPS, making it highly suitable for accurate and efficient fabric defect detection in industrial environments.