基于PU-Faster R-CNN的手机屏幕缺陷检测算法研究
2023,31(7):99-106
摘要：手机屏幕缺陷检测是手机生产的重要环节，实现准确而高效的屏幕缺陷检测对于提高手机工业产能具有重要意义。在实际生产过程中，手机屏幕图像缺陷特征隐晦、缺陷尺寸差异大等问题，加大了手机屏幕缺陷检测的难度。为解决上述问题，提出了一种基于Preprocessing operations are combined with U-Net-Faster R-CNN（PU-Faster R-CNN）的手机屏幕缺陷检测模型。针对手机屏幕图像的特征信息隐晦的问题，提出多层特征增强模块，有效的对目标缺陷特征信息进行增强。构建多尺度特征提取网络，有效提取多尺度的缺陷特征信息。为了生成拟合性更好的Anchor box，提出了自适应区域建议网络，通过自迭代聚类算法生成尺寸更准确的Anchor box模板。实验结果表明，基于PU-Faster R-CNN的手机屏幕检测框架在手机屏幕数据集上优于主流的手机屏幕缺陷检测框架。
关键词：手机屏幕；缺陷检测；PU-Faster R-CNN；多层特征增强模块；自适应区域建议网络；
PU-Faster R-CNN Based Defect Detection Modelfor Mobile Phone Screen
Abstract：Mobile phone screen defect detection is an important part of mobile phone production. To achieve accurate and efficient defect detection is of great significance for improving the productivity of mobile phone industry. In the actual production process, the screen defect image features are not obvious and the defect size difference is large, which increases the difficulty of mobile phone screen defect detection. A mobile phone screen defect detection model based on PU-Faster R-CNN was proposed to solve the above problems. For the problem of obscure feature information of cell phone screen images, a multi-layer feature enhancement module was proposed to effectively enhance the target defect feature information. A multi-scale feature extraction network was constructed to effectively extract multi-scale defect feature information. In order to generate Anchor boxes with better fitting performance, an adaptive region proposal network was proposed to generate Anchor box templates with more accurate size by self-iterative clustering algorithm. The experimental results showed that the framework was superior to the mainstream mobile phone screen defect detection framework in mobile phone screen datasets.
Key words：Mobile Phone Screen; Defect Detection; PU-Faster R-CNN; Multi-layer Feature Enhancement Module; Adaptive Region Proposal Network;
收稿日期：2023-01-17
基金项目：科技创新2030-“新一代人工智能”国家级重大项目(2020AAA0108304)
