基于卷积神经网络的TFT-LCD引脚导电性能检测

Conductive performance detection of TFT-LCD bumps based on convolutional neural network

针对导电粒子粘连严重影响TFT-LCD引脚质量检测精度问题,提出融入K均值聚类特征选择的卷积神经网络引脚导电性能二分类图像检测方法。基于引脚附着粒子数量合格与否的判别标准,建立引脚图像检测的合格与不合格二分类模型。根据监督学习机理,构建包含2 000张引脚图像的数据集。为了提高检测精度,针对不同引脚图像中导电粒子与背景之间灰度阈值存在差异的特点,提出融入K均值聚类特征选择的卷积神经网络引脚图像粒子数量分类算法。进一步应用10折交叉验证法对算法的有效性进行了评估,检测正确率达到96.0%,比极大类间方差法提高了8%,比分水岭法提高了5.5%。

Aiming at the phenomenon that the adhesion of conductive particles seriously affects the detection accuracy of thin film transistor liquid crystal display(TFT-LCD)bumps,a binary classification image detection method for the conductivity of bumps with convolutional neural network(CNN)incorporating K-means clustering feature selection was proposed.A binary classification model of qualified and unqualified bump image detection was established based on the standard of judging whether the number of particles attached to the bump was qualified or not.A data set containing 2000 bump images was constructed according to the supervised learning mechanism.In order to improve the detection accuracy,further aiming at the characteristics of differences in the gray threshold between conductive particles and the background in different bump images,a CNN bump image particle number classification algorithm incorporating K-means clustering feature selection was proposed.A 10-fold cross-validation method was applied to evaluate the effectiveness of the algorithm,and the algorithm accuracy rate reached 96.0%,which was 8%higher than that of Otsu’s method and 5.5%higher than that of the watershed method.