基于小样本学习的LCD产品缺陷自动检测方法

PDF下载HTML

Title:

An automatic small sample learning-based detection method for LCD product defects

作者:

马岭¹, 鲁越¹, 蒋慧琴¹, 刘玉敏²

1. 郑州大学 信息工程学院，河南 郑州 450001;
2. 郑州大学 商学院，河南 郑州 450001

Author(s):

MA Ling¹, LU Yue¹, JIANG Huiqin¹, LIU Yumin²

1. School of Information Engineering, Zhengzhou University, Zhengzhou 450001, China;
2. Business School, Zhengzhou University, Zhengzhou 450001, China

关键词:

缺陷诊断; 目标分类; 深度学习; 小样本学习; 卷积神经网络; 迁移学习; 深度卷积生成对抗网络; 继续学习

Keywords:

defect diagnostics; target classification; deep learning; small sample learning; convolution neural network; transfer learning; DCGAN; continue learning

分类号:

TP391

DOI:

10.11992/tis.201904020

摘要:

针对高分辨率液晶显示器产品(liquid crystal display, LCD)质量在线检测需求，基于深度学习提出一种LCD缺陷自动检测方法。通过设计自适应浅层特征提取层，并引入稀疏卷积结构，多维度、多尺度的提取深层特征，采用迁移学习和深度卷积生成对抗生网络扩充数据强化训练，构建基于小样本学习的LCD表面缺陷检测模型。其特征在于，采用设计的自动分割与定位预处理软件将高分辨率图像划分成适于卷积神经网络学习的图像子块，并根据模型对图像子块的判定类别和定位坐标，同时获取多类型缺陷检测结果。实验结果表明，本文模型可以有效提高检出率，并减少漏检率。

Abstract:

Aiming at the demands of high-resolution liquid crystal display (LCD) product quality online inspection, we propose an automatic deep learning-based detection method for LCD defects. An LCD surface defect detection model is constructed based on small sample learning by designing an adaptive shallow feature extraction layer and introducing a sparse convolution structure to extract the multi-dimensional and multi-scale deep features. Furthermore, the model is trained and enhanced using transfer learning and a deep convolutional generative adversarial network (DCGAN) for data expansion, and the LCD surface defect detection model is built based on small sample learning. The original high-resolution image is segmented into sub-images suitable for convolutional neural network (CNN) learning by designing automatic segmentation and location pretreatment software. The detection results of multi-type defects according to the category and location coordinates of the classification model’s output are obtained. The experimental results show that the model can effectively improve the detection rate and reduce the missed detection rate.