蓝宝石LED衬底表面缺陷检测和综合评定方法研究

Quality of sapphire substrate is the key for ensuring GaN epitaxial growth and improving of LED performance. However, existing commercial instruments fail in giving a quantitative description for the characteristics of the surface defects of sapphire substrate, thus making it difficult to achieve automated and efficient elimination of defective goods. This project contains the measurement method for surface defects, the characterizing of feature parameters, the recognition of defects features and the construction of evaluation software. The purpose of the project is to accurately locate and quantitatively describe the surface defects through the accurate measurement of sapphire substrate and identification of surface defects. Defects data can be obtained by the combination of scattering and confocal microscopy technology. The surface defects are identified by means of the image spectrum analysis technique, fractal theory and wavelet method. The defects are classified by using the Hu moment invariants and the inverse neural network algorithm. The scratches, cracks, pits and so on are mainly discussed in this work. Based on quantitative characterization of defect features, a special software will be established for quantitative evaluation of sapphire substrate defects. This project provides technical rationale to improve existing commercial instruments that are unable to give comprehensive and quantitative description for characteristic of surface defects. Ultimately, this project will make substantial contribution to the optimization of sapphire substrate industry.

高精度的蓝宝石衬底表面质量，对在其表面生长GaN薄膜及LED发光二极管的性能至关重要。然而，现有商业仪器不能对具有多样化特征的各类缺陷进行特征定量表征，使得自动化生产线高效准确剔除次品难以实现。项目从表面缺陷测量方法、缺陷特征参数表征、缺陷特征识别及评定软件构建等四个方面开展研究工作，旨在实现对蓝宝石衬底表面几何参数准确测量和缺陷特征准确识别，达到缺陷精确定位和定量描述的效果。项目结合散射法和共焦显微测量方法，获取蓝宝石衬底表面缺陷测量数据；基于图像频谱分析技术、分形理论和小波方法对缺陷特征进行准确识别；采用Hu矩不变数和逆向神经网络算法对缺陷进行分类训练，重点展开对划痕、裂纹、气泡、坑点、崩角等缺陷的特征分类。基于缺陷特征定量表征，构建蓝宝石衬底缺陷定量评定专用软件系统。项目的研究将有助于弥补现有商业检测仪器不能全面给出缺陷特征定量个性化描述等不足，为蓝宝石衬底产业工艺优化提供依据。

本项目针对蓝宝石衬底自动缺陷检测与特征定量描述的发展需求，制定了衬底表面缺陷检测方案，研究了表面缺陷二维以及三维特征参数提取算法，并基于深度学习网络模型实现对表面缺陷的分类，最后分析参数评定功能需求，确定了缺陷评定系统的组成模块。主要内容包括：.（1）根据蓝宝石衬底材料特性和表面典型缺陷结构特性，讨论了蓝宝石衬底表面典型缺陷对光的散射机制；为了实现对不同规格蓝宝石衬底表面缺陷的有效检测，重点对检测系统的光源方案和回转工作台进行设计以满足测量要求。.（2）针对缺陷二维、三维形貌特征，研究不同特征参数的提取算法。通过Krisch边缘检测、去除连通域及曲线拟合等算法，提取缺陷的长度、位置坐标、圆形度及矩形度等二维特征参数；采用聚焦合成算法对层析图像进行三维重构，通过对层析图像进行景深合成，提取缺陷三维特征参数。.（3）基于ZF-net卷积神经网络模型，开展衬底表面缺陷分类算法研究。针对训练网络大数据需求，采用图像处理技术中的旋转（Rotation）、翻转(Flip)和深度学习生成对抗GAN网络进行数据集扩充。结合缺陷特征，根据预先设置的训练模型参数，对样本进行网络训练，获得学习模型。测试结果表明所构建的学习网络对不同类型缺陷分类的准确度优于96%。.（4）表面缺陷评定软件系统研究。根据功能需求分析以及算法要求，确定了表面缺陷评定系统的组成模块以及检测流程，完成表面缺陷评价分析功能。