YOLO-FIX: Improved YOLOv11 with Attention and Multi-Scale Feature Fusion for Detecting Glue Line Defects on Mobile Phone Frames
Tianrun Ye, Shize Huang, Weiwei Qin, Haiyang Tu, Ping Zhang, Yafei Wang, Chunming Gao, Yanli Gong
Abstract
This paper presents YOLO-FIX, an improved intelligent detection model based on YOLOv11, designed to identify glue line defects in mobile phone frames. The model addresses the challenges of complex glue line morphology, background interference, and illumination transformation. YOLO-FIX enhances the extraction of local and global features to optimize the detection accuracy by integrating advanced attention mechanisms and multi-scale feature fusion modules specifically Deformable Large-Kernel Attention (De-formable-LSKA) and Mamba-Like Linear Attention (MLLA). Experimental evaluations demonstrate that YOLO-FIX achieves a mean Average Precision (mAP50) of 95.2%, an 8.6% improvement over the baseline YOLOv11 model while maintaining a real-time detection speed of 189 FPS. It effectively identifies five common defect types: broken glue, wall climbing, glue dropping, single-tip wall climbing, and collapsed glue, showcasing exceptional robustness and generalization across varying production environments. These results affirm YOLO-FIX as a highly accurate and efficient solution for automated defects in industrial applications.