多背景下TFT-LCD mura缺陷视觉检测方法,Journal of the Society for Information Display

当前位置： X-MOL 学术 › J. Soc. Inf. Disp. › 论文详情

Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)

多背景下TFT-LCD mura缺陷视觉检测方法
Journal of the Society for Information Display ( IF 1.7 ) Pub Date : 2022-08-12 , DOI: 10.1002/jsid.1171
Mingfang Chen ₁ , Ping Chen ₁ , Sen Wang ₁ , Yu Cui ₁ , Yongxia Zhang ₁ , Songlin Chen ₁

Affiliation

薄膜晶体管液晶显示（TFT-LCD）面板的mura缺陷小，灰度变化未知。在人工检测中，mura很容易被忽略，耗时耗力。这些因素导致我们无法在现场检查中正确评估和区分单个图像上的多个 Mura 缺陷。针对上述问题，本文提出了一种多背景TFT-LCD Mura缺陷视觉检测方法。为获得高精度高速检测下的最优算法，克服YOLOV4-tiny算法的漏检问题，在原有框架的基础上增加了空间金字塔池化模块和squeeze-and-excitation模块. 一方面，本文改进算法的优势在于专注于关键信息，使得特征融合前的信息更加丰富。另一方面，在 BackBone 中加入空间金字塔池化使得网络能够对输入特征层进行多尺度池化和融合，大大增强了网络的感受野，使网络能够提取更丰富的特征信息。实验结果表明，改进后的YOLOV4-tiny算法在TFT-LCD mura检测中的准确率达到99.72%，比YOLOV4-tiny高0.99%，实时检测速度达到63.84帧/秒（FPS）。同时，我们克服了YOLOV4-tiny算法的漏检问题，满足了多背景环境下TFT-LCD mura检测任务的检测精度和实时性要求。在BackBone中加入空间金字塔池化使得网络能够对输入特征层进行多尺度池化和融合，极大地增强了网络的感受野，使得网络能够提取更丰富的特征信息。实验结果表明，改进后的YOLOV4-tiny算法在TFT-LCD mura检测中的准确率达到99.72%，比YOLOV4-tiny高0.99%，实时检测速度达到63.84帧/秒（FPS）。同时，我们克服了YOLOV4-tiny算法的漏检问题，满足了多背景环境下TFT-LCD mura检测任务的检测精度和实时性要求。在BackBone中加入空间金字塔池化使得网络能够对输入特征层进行多尺度池化和融合，极大地增强了网络的感受野，使得网络能够提取更丰富的特征信息。实验结果表明，改进后的YOLOV4-tiny算法在TFT-LCD mura检测中的准确率达到99.72%，比YOLOV4-tiny高0.99%，实时检测速度达到63.84帧/秒（FPS）。同时，我们克服了YOLOV4-tiny算法的漏检问题，满足了多背景环境下TFT-LCD mura检测任务的检测精度和实时性要求。极大地增强了网络的感受野，使网络能够提取更丰富的特征信息。实验结果表明，改进后的YOLOV4-tiny算法在TFT-LCD mura检测中的准确率达到99.72%，比YOLOV4-tiny高0.99%，实时检测速度达到63.84帧/秒（FPS）。同时，我们克服了YOLOV4-tiny算法的漏检问题，满足了多背景环境下TFT-LCD mura检测任务的检测精度和实时性要求。极大地增强了网络的感受野，使网络能够提取更丰富的特征信息。实验结果表明，改进后的YOLOV4-tiny算法在TFT-LCD mura检测中的准确率达到99.72%，比YOLOV4-tiny高0.99%，实时检测速度达到63.84帧/秒（FPS）。同时，我们克服了YOLOV4-tiny算法的漏检问题，满足了多背景环境下TFT-LCD mura检测任务的检测精度和实时性要求。每秒 84 帧 (FPS)。同时，我们克服了YOLOV4-tiny算法的漏检问题，满足了多背景环境下TFT-LCD mura检测任务的检测精度和实时性要求。每秒 84 帧 (FPS)。同时，我们克服了YOLOV4-tiny算法的漏检问题，满足了多背景环境下TFT-LCD mura检测任务的检测精度和实时性要求。

"点击查看英文标题和摘要"

TFT-LCD mura defect visual inspection method in multiple backgrounds

The defect of mura of thin film transistor liquid crystal display (TFT-LCD) panel is small, and the gray change is unknown. In manual detection, mura is easily overlooked, which is time-consuming and labor-intensive. These factors resulted in our inability to properly assess and distinguish multiple mura defects on a single image in field inspections. Aiming at the above problems, this article proposes a multibackground TFT-LCD mura defect visual inspection method. To obtain the optimal algorithm under high-precision and high-speed detection and to overcome the problem of missed detection of the YOLOV4-tiny algorithm, the spatial pyramid pooling module and squeeze-and-excitation module are added on the basis of the original framework. On the one hand, the advantage of the improved algorithm in this article is to focus on the key information, so that the information before feature fusion is more abundant. On the other hand, adding spatial pyramid pooling in BackBone enables the network to perform multiscale pooling and fusion on the input feature layer, which greatly enhances the receptive field of the network and enables the network to extract richer feature information. Experimental results show that the improved YOLOV4-tiny algorithm has an accuracy of 99.72% in TFT-LCD mura detection, which is 0.99% higher than YOLOV4-tiny, and the real-time detection speed reaches 63.84 frames per second (FPS). At the same time, we overcome the problem of missed detection of YOLOV4-tiny algorithm and meet the detection accuracy and real-time requirements of TFT-LCD mura detection tasks in multibackground environment.

更新日期：2022-08-12

点击分享查看原文

点击收藏

阅读更多本刊新发论文本刊介绍/投稿指南