Comparison of FPS of three models.

In smartphone production, detecting small defects on screens remains challenging due to low detection accuracy, high missed detection rates, and slow processing speeds. To address these issues, this paper presents a Lightweight Network Based on YOLOv8 (LNB-YOLO) for defect detection, with several key enhancements. First, a Feature Pyramid Network based on Context-Guided Spatial Feature Reconstruction (CGRFPN) is integrated to improve the perception of multi-level features and enhance small target recognition in complex backgrounds. Second, the Efficient Local Attention (ELA) module is incorporated into the Backbone’s C2F module to improve localization precision, while the Minimum Point Distance based IoU (MPDIoU) loss function is employed to prevent gradient explosion. Third, a lightweight Detail-Enhanced Convolution and Shared Convolutional Detection Head (LSDECD) is designed to capture fine details while reducing parameters and computational complexity. Finally, model pruning and knowledge distillation techniques are applied to further optimize efficiency. Experimental results on the PKU-Market-Phone dataset show that LNB-YOLO achieves a mAP@0.5 of 97.5% and a mAP@.5:.95 of 68.8%, surpassing the original YOLOv8 by 6.1% and 9.3%, respectively. The model also reduces parameters by 80% and computational cost by 63%, effectively meeting precision requirements for smartphone production lines.