尚光网 本站首页 本所首页 联系我们
chi20190523上海光机所刘德安557478In large-scale high-power laser driving devices, the number of optical components is large. Working at high energy densities, component damage problems will be inevitable. As the number of laser irradiation increases, the initial damage area will gradually expand, resulting in a decrease in the service life of the component, which further affects the quality of the output laser and causes the modulation effect of the spatial light, which has a destructive effect on the downstream optical components. Therefore, it is important to perform routine damage monitoring and strategic state assessment and repair of optical components in the system. The first step in studying the damage problem is to accurately identify the point of the disease on the component. In order to effectively curb the growth of damage and avoid the problems caused by this, this paper will focus on the identification of defects. At the same time, the micro-damage segmentation algorithm and the elimination of false damage areas are systematically studied. The main content is: Clarify the target accuracy that the damage detection work needs to achieve. Briefly describe the overall process of defect detection based on imaging. Combined with the dark field imaging and total reflection illumination technology design system image acquisition device. And the imaging resolution and imaging effect map are given. Analyze the current status of ICF device damage detection image processing technology. Point out the inadequacies of the existing segmentation algorithm. Analysis of defect classification and pseudo-injury rejection is important for evaluating the damage state of components. It is pointed out that machine learning related algorithms can effectively deal with image classification problems. Aiming at the problem of poor performance of the original local signal-to-noise ratio method, a local signal-to-noise ratio segmentation algorithm based on adaptive difference window is proposed. The adaptive difference window mainly acts on the binarization of the seed image. The principle is to correlate the gray value of the pixel to be judged with its neighbor point and the overall gray distribution of the image. If the point to be judged satisfies the corresponding condition, it is marked as seed candidate points. The algorithm solves the problem that the traditional method relies on the experience to select the threshold, and improves the efficiency of the damage segmentation. Experiments show that under the side illumination total reflection dark field imaging mode, the average recognition rate of the algorithm is 99.37% for defect areas with a diameter of 30μm or less. In order to further reduce the missed detection rate of the damage point, a local signal-to-noise ratio algorithm based on gray histogram is proposed. The experimental results show that under the same conditions, the average recognition rate of the algorithm is 99.61%, which effectively improves the segmentation accuracy. In order to more accurately evaluate the damage state of optical components and eliminate the interference of dust, stains and other stray light imaging on the detection results, a technical scheme of using machine learning algorithms to train classifiers to identify false damages is proposed. Firstly, according to the characteristics of the target area to be classified, the corresponding gray level and topographic feature parameters are extracted to construct the sample data set. Then use AdaBoost and decision tree algorithm to train the damage classifier, and discuss the algorithm principle and the improvement method of weight update and training efficiency in detail. In the experiment, after repeated iterations, the cross-validation technique was used to test. Finally, the average classification accuracy of the model was 94.7%, which can effectively eliminate the false damage area and achieve the purpose of accurately evaluating the damage state of the optical component.201906atalunwen2195291192425Damage detection;Image segmentation; Adaptive thresholding;Region growing;Defect classificationResearch on Automatic Identification Technology of Micro Damage of Optical Components Based on Computer Vision基于计算机视觉的光学元件微小损伤自动识别技术研究大型高功率激光驱动装置中,光学元件数量繁多且工作在高能量密度下,损伤问题将不可避免。随着激光辐照次数的增加,初始损伤区域将逐渐扩大,造成元件使用寿命的降低,又进一步影响输出激光的质量并引起空间光的调制效应,对下游光学元件造成破坏性影响。因此,对系统中的光学元件进行常规的损伤监测与有策略的状态评估及修复工作是十分重要的。研究损伤问题的第一步,是准确的分割出元件上的疵病点。为有效遏制损伤增长并避免由此引发的种种问题,本文将重点围绕缺陷点的识别工作而展开研究,同时,系统研究了微小损伤分割算法以及后续虚假损伤区域的剔除问题,主要内容为: 明确本文损伤检测工作需要达到的目标精度,简述基于图像法的缺陷检测总体流程,并结合暗场成像与全反射照明技术设计系统的图像获取装置,给出成像分辨率及样品元件成像效果图。分析ICF装置损伤检测图像处理技术现状,指出现有区域分割算法的不足之处。分析缺陷分类及伪损伤剔除对评估元件损伤状态的重要意义,指出机器学习相关算法可以有效处理缺陷分类问题。 针对原始局部信噪比方法性能较差的问题提出基于自适应差异窗的局部信噪比分割算法。自适应差异窗主要作用于种子图像的二值化,原理是将待判断像素点的灰度值同其邻域点及图像整体灰度分布进行关联对比,若待判断点满足相应条件则标记为种子候选点。算法解决了传统方法依赖经验选取阈值的问题,提高了损伤分割的效率。实验表明,在侧照明全反射暗场成像方式下,针对直径30μm甚至更小的缺陷区域,算法的平均识别率为99.37%。为降低损伤漏检率,又进一步提出基于灰度直方图的局部信噪比算法。实验表明在相同条件下,算法的平均识别率为99.61%,有效提高了分割精度。 为更加准确地评估光学元件的损伤状态,排除灰尘、污渍及其他杂散光成像等对检测结果的干扰,提出使用机器学习算法训练分类器来鉴别虚假损伤的技术方案。首先根据待分类目标区域的特点,提取相应灰度及形貌特征参数来构建样本数据集。然后使用AdaBoost及决策树算法训练损伤分类器,并详细论述算法原理及针对权值更新与训练效率的改进方法。实验中,经多次迭代更新后,使用交叉验证技术进行测试,最后得到模型的平均分类正确率为94.7%,能够有效剔除虚假损伤区域,达到准确评估光学元件损伤状态的目的。损伤检测;图像分割;自适应阈值化;区域生长;缺陷分类中国科学院上海光学精密机械研究所唐如欲光学工程硕士
中文题目: 基于计算机视觉的光学元件微小损伤自动识别技术研究
外文题目: Research on Automatic Identification Technology of Micro Damage of Optical Components Based on Computer Vision
作者: 唐如欲
导师姓名: 刘德安
学位授予机构: 中国科学院上海光学精密机械研究所
答辩时间: 20190523
Damage detection;Image segmentation; Adaptive thresholding;Region growing;Defect classification
学位级别: 硕士
正文语种: chi
页面点击量: 16
文章下载量: 1
版权所有 © 2009 中国科学院上海光学精密机械研究所 沪ICP备05015387号
主办:中国科学院上海光学精密机械研究所 上海市嘉定区清河路390号(201800)