摘要
针对汽车玻璃缺陷样本稀缺的问题,提出了一种基于潜在扩散模型的数据增强方法。该方法利用潜在扩散模型生成高质量的合成缺陷图像,再结合伪标签标注技术,显著提升了数据集的多样性。为解决原始YOLOv5s算法在跨尺度多目标检测任务中易漏检与错检的问题,通过在颈部区域引入BiFPN结构,实现多尺度特征的高效融合。实验结果表明,改进的算法对缺陷的识别精度明显提高,为汽车玻璃缺陷检测提供了可靠的技术支持与理论参考。
To address the issue of scarce defect samples in automotive glass,a data augmentation method based on a latent diffusion model is proposed.This method leverages the latent diffusion model to generate high-quality synthetic defect images,combined with pseudo-labeling techniques,significantly enhancing the diversity of the data sets.To resolve the problem of missed detection and false detection in the original YOLOv5s algorithm during cross-scale multi-object detection tasks,a BiFPN structure is introduced in the neck region to achieve efficient fusion of multi-scale features.Experimental results demonstrate that the improved algorithm significantly enhances defect recognition accuracy,providing reliable technical support and theoretical reference for automotive glass defect detection tasks.
作者
罗志学
黄诗浩
LUO Zhixue;HUANG Shihao(Fujian Provincial Key Laboratory of Automotive Electronics and Electric Drive,Fujian University of Technology,Fuzhou Fujian 350118,China;School of Electronic,Electrical Engineering and Physics,Fujian University of Technology,Fuzhou Fujian 350118,China)
出处
《莆田学院学报》
2025年第2期79-85,共7页
Journal of putian University
基金
福建省自然科学基金资助项目(2022J01950)。
作者简介
罗志学(1996-),男,福建三明人,2022级硕士研究生,主要从事机器视觉与图像处理等方面研究;通讯作者:黄诗浩(1985-),男,福建三明人,副教授,博士,主要从事光电信息材料与器件、机器视觉等方面研究。
