基于Mask R-CNN的触摸屏玻璃疵病检测与识别

Detection and Identification of Glass Defects on the Touch Screen Based on Mask R-CNN

为快速有效地检测与识别触摸屏玻璃疵病,提出了一种基于Mask R-CNN技术的检测与识别方法。Mask R-CNN是Faster R-CNN技术的扩展,其在边界框识别的现有分支上添加一个并行的用于预测对象掩膜的分支,是基于R-CNN系列、FPN、FCIS等工作之上的一种技术,相对于传统的目标检测和分割算法有很大提升。使用暗场散射成像、多图片堆栈等方法采集数据样本,同时采用几何变换方法扩增样本库。为提升训练速率和检测精确率,根据训练样本差异性,通过测试不同的VGG16、ResNet50+FPN以及ResNet101+FPN主干网络提取图像特征。实验表明,ResNet50+FPN的识别精确率较高,达到96.7%。基于Mask R-CNN的检测与识别方法不仅能对疵病快速检测与识别,还能提高识别精确率。

In order to detect and identify glass defects on the touch screen quickly and effectively,a screen glass defect detection and identification method based on Mask R-CNN technology is proposed.Mask R-CNN is Faster R-CNN technical extension,on the existing branches of applied to identify boundary box a parallel branch is added to predict object mask based on R-CNN series,FPN,FCIS.Compared with traditional target detection and segmentation algorithm,the algorithm has been greatly improved.This paper uses dark field scattering imaging,multiple image stacks and other methods to collect data samples,and uses geometric transformation methods to expand the sample library.In order to improve the training rate and detection accuracy,according to the differences of the training sample,this paper tests different backbone networks of VGG16,ResNet50+FPN,ResNet101+FPN to extract the image features,and finds that the recognition accuracy is higher and the accurate rate reaches 96.7%.As result,the detection and recognition method based on Mask R-CNN can not only detect and identify defects quickly and effectively,but also improve the recognition accuracy rate.