范先友,过 峰,俞建峰,化春键,蒋 毅,钱陈豪.基于改进 YOLOv7 的液晶面板电极缺陷视觉检测技术研究[J].电子测量与仪器学报,2023,37(9):225-233 |
基于改进 YOLOv7 的液晶面板电极缺陷视觉检测技术研究 |
Research on visual detection technology for liquid crystal panel electrode defect by improved YOLOv7 |
|
DOI: |
中文关键词: 缺陷检测 改进 YOLOv7 液晶面板 电极 |
英文关键词:defect detection improved YOLOv7 liquid crystal panel electrode |
基金项目:国家自然科学基金(51905215)项目资助 |
|
Author | Institution |
Fan Xianyou | 1. School of Mechanical Engineering, Jiangnan University, 2. Jiangsu Key Laboratory of Advanced Food Manufacturing Equipment and Technology |
Guo Feng | 3. Textile Industrial Products Testing Center of Nanjing Customs District |
Yu Jianfeng | 1. School of Mechanical Engineering, Jiangnan University, 2. Jiangsu Key Laboratory of Advanced Food Manufacturing Equipment and Technology |
Hua Chunjian | 1. School of Mechanical Engineering, Jiangnan University, 2. Jiangsu Key Laboratory of Advanced Food Manufacturing Equipment and Technology |
Jiang Yi | 1. School of Mechanical Engineering, Jiangnan University, 2. Jiangsu Key Laboratory of Advanced Food Manufacturing Equipment and Technology |
Qian Chenhao | 1. School of Mechanical Engineering, Jiangnan University, 2. Jiangsu Key Laboratory of Advanced Food Manufacturing Equipment and Technology |
|
摘要点击次数: 880 |
全文下载次数: 827 |
中文摘要: |
电极质量对液晶面板的显示效果极其重要,针对液晶面板电极缺陷种类多、尺度小、背景复杂而导致难以检测的问题,
本文提出了一种基于改进 YOLOv7 的液晶面板电极缺陷视觉检测方法。 首先,将 CBAM 注意力模块嵌入到 YOLOv7 骨干网络
中,抑制背景信息干扰,强化缺陷特征;其次,采用跨层级连接操作,实现浅层网络与深层网络特征信息的融合;然后,将 C2f 模
块融入特征金字塔网络中以轻量化模型,提高训练速度;最后,使用 WIoU 替换 YOLOv7 模型的损失函数,减小低质量标注产生
的有害梯度,提高对缺陷的定位性能。 在自定义的电极缺陷数据集上进行测试,结果表明,该算法对电极大划伤、划伤、磕伤以
及脏污 4 类缺陷的平均检测精度达 67. 8%,单张检测时间为 5. 6 ms。 |
英文摘要: |
The quality of electrode is extremely important for the display effect of liquid crystal panel. To solve the problem of difficult
detection due to the variety of electrode defects, small scale and complex background, an electrode defect detection method for liquid
crystal panels is proposed based on improved YOLOv7 algorithm. Firstly, the CBAM attention module was embedded into the YOLOv7
backbone network to suppress background information interference and strengthen defect features. Secondly, the feature information of
the shallow networks and deep networks was fused by cross-level connection operation. Then, the C2f module was integrated into the
feature pyramid network to lightweight the model and improve the training speed. Finally, the WIoU was used to replace the loss function
of the YOLOv7 model to reduce the harmful gradient caused by low quality labeling and improve the defect location performance. By a
customized electrode defect dataset, the results showed that the proposed algorithm was able to achieve an average detection accuracy of
67. 8% for large scratch, scratch, shell and dirt on electrodes with a per-sheet detection time of 5. 6 ms. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|