MicroLED表面缺陷的快速高精度检测

打开文本图片集

关键词：深度学习；MicroLED;缺陷检测；动态卷积；注意力机制中图分类号：TP394.1；TH691.9 文献标识码：Adoi：10.37188/0PE.20253309.1434 CSTR：32169.14.OPE.20253309.1434

Rapid and high-precision detection on surfacedefectsofMicroLED

ZHAO Tianyuan 1，2 ，DONG Dengfeng 1，2* ，WANG Guoming 1，2 ，WANG Bo 1 ， ZHOU Weihu 1，2 （1. Institute ofMicroelectronics， Chinese Academy of Sciences，Beijing lOoo29，China; 2.University of Chinese Academy of Sciences，Beijing1OoO49，China） * Corresponding author， E-mail： dongdengfeng@ime. ac. cn

Abstract：To address the demands for real-time and high-precision Micro LED defect detection，this study introduces LED-YOLO，a rapid and accurate detection algorithm that integrates a lightweight architecture with enhanced feature extraction capabilities.An image acquisition system was designed to simulate industrial interference，and various data augmentation techniques were employed to increase the diversity of training data. To overcome the limited discriminative power for Micro LED defects，a Lightweight Dynamic Fusion Module（LDFM）was developed，combining dynamic convolution，deep convolution， and channel mixing operations;this approach maintains model compactness while enhancing feature extraction.Furthermore，an Enhanced Coordinated Attention Module （ECAM）was proposed to improve defect localization by integrating channel and spatial attention mechanisms alongside residual connections， thus refining feature extraction accuracy. Given the minimal aspect ratio variation in Micro LED images，a dynamic focusing mechanism was incorporated，and a DIoU_W regression loss function was introduced to accelerate convergence and improve robustness. Experimental results demonstrate that LED-YOLO surpasses the state-of-the-art YOLOvlls in detection accuracy，recall， mean average precision （mAP），and F1 score.Despite a reduction of 1.6 milion parameters，LED-YOLO achieves substantial improvements in detection speed and accuracy，effectively fulfilling the quality inspection requirements of Micro LED panel manufacturing.

Key words： deep learning；Micro LED;defect detection; dynamic convolution;attention

1引言

MicroLED是一种新兴的显示技术，由微型发光二极管组成，以高亮度、宽色域和低功耗等优点，正迅速成为显示技术的主流选择。（剩余15419字）