| 引用本文: | 任秉银,李智勇,代勇.手机屏幕轻微划痕检测方法[J].哈尔滨工业大学学报,2021,53(1):29.DOI:10.11918/202008109 |
| REN Bingyin,LI Zhiyong,DAI Yong.Method for detection of slight scratch of mobile phone screen[J].Journal of Harbin Institute of Technology,2021,53(1):29.DOI:10.11918/202008109 |
|
| 摘要: |
| 针对手机屏幕等产品光滑表面轻微划痕的自动检测问题,提出一种基于分类网络+Attention U-Net的小目标分割与微小缺陷检测方法. 论述基于经典的U-Net网络进行光滑表面缺陷检测的数据集准备、语义分割网络构建、评估指标、损失函数、正则化方法以及初始化方式, 分析应用经典的U-Net神经网络对微小缺陷误检测与漏检测的原因; 给出在分类网络中加入分割网络以及加入Attention机制对U-Net网络进行改进的方案; 搭建分类网络+Attention U-Net以改善小目标分割与微小缺陷检测效果. 结果表明:提出的改进网络方案对手机屏幕轻微划痕等微小缺陷检测的像素准确率达到0.997,能够很好地满足准确检测手机屏幕轻微划痕的实际需求, 也能为瓷砖等产品的光滑表面的轻微划痕与裂纹检测提供有益参考. |
| 关键词: 手机屏幕 缺陷检测 轻微划痕 深度学习 语义分割 小目标检测 |
| DOI：10.11918/202008109 |
| 分类号:TN911.7 |
| 文献标识码:A |
| 基金项目: |
|
| Method for detection of slight scratch of mobile phone screen |
|
REN Bingyin,LI Zhiyong,DAI Yong
|
|
(School of Mechatronics Engineering, Harbin Institute of Technology, Harbin 150001, China)
|
| Abstract: |
| Aiming at automatically detecting the slight scratch of smooth surface of mobile phone screen, a method for small target segmentation and insignificant defects detection based on “Classification net + Attention U-Net” is presented. The necessary components of smooth surface defect detection system based on the classic U-Net are proposed, including preparation of data set, semantic segmentation network, definition of evaluation standards and loss function, appropriate regularization methods and special initialization method. The reason for failing to detect and false detection of insignificant defects by the constructed classic U-Net are analyzed. An improved scheme of adding segmentation network to classification network and adding the Attention mechanism to the classic U-Net is presented. A network consists of “Classification network + Attention U-Net” is constructed to improve the effects of small target segmentation and insignificant defects detection. The results show that the pixel accuracy rate of defects detection of mobile phone screen arrives 0.997 by the presented improved net. The presented method can meet the actual requirements on detecting the slight scratch defects of the mobile phone screen. It can also provide valuable reference for similar scratch and crack detection of smooth surface of ceramic tiles. |
| Key words: mobile phone screen defect detection slight scratch deep learning semantic segmentation small target detection |
