【作者】 李明；

【导师】 贾华宇；

【作者基本信息】 太原理工大学 ， 控制工程（专业学位）， 2020， 硕士

【摘要】 近年来,我国的激光科学技术正处于迅猛发展阶段。在社会全面发展、产业经济提升、国防安全应用和经济结构转型以及国家竞争性发展等各方面,激光科技的全面创新和产业应用的转型发展都有着十分重要的战略意义。激光器是激光科学与技术的核心,然而在激光芯片的生产过程中,不可避免地会出现损伤缺陷。传统检测方法耗时、严重依赖模型、需要丰富的操作经验,而且无法实现缺陷定位。本文中我们主要研究卷积神经网络在激光芯片缺陷检测中的应用,并做了以下几方面的工作:第一、将采集到的激光芯片缺陷图像进行标记和图像预处理,通过将经过预处理的缺陷图像进行旋转、翻转、错切等一系列图像处理操作,人为地实现数据扩充,从而构建激光芯片缺陷图像数据集;第二、将卷积神经网络和类激活映射技术相结合,在Alex Net模型的基础上对其进行改进并得到Laser Chip Net网络;第三、使用经过图像预处理后的激光芯片缺陷图像数据集对Laser Chip Net网络进行训练;第四、设计了三组实验来验证图像预处理操作对模型准确率的提升作用、模型在激光芯片缺陷分类识别和缺陷区域定位方面的优良性能。最终的结果表明,本文提出的网络模型分类准确度高,能在不使用区域级人工注释的情况下进行缺陷区域定位,而且分类和定位功能在一个网络中就能实现,十分的快速和方便。可见,本文提出的模型对于激光芯片缺陷检测是有积极意义的。更多还原

【Abstract】 In recent years,China’s laser science and technology is in a stage of rapid development.In all aspects of social development,industrial economic improvement,national defense security applications and economic structural transformation,and national competitive development,the comprehensive innovation of laser technology and the transformation and development of industrial applications have very important strategic significance.Laser is the core of laser science and technology.However,in the production process of laser chips,damage defects will inevitably occur.Traditional inspection methods are timeconsuming,rely heavily on models,require extensive operating experience,and cannot locate defects.In this article,we mainly study the application of convolutional neural network in laser chip defect detection,and do the following work:First,mark and pre-process the collected laser chip defect images.By rotating,flipping,and miscutting the preprocessed defect image,a series of image processing operations are performed to artificially expand the data to construct the laser chip defect Image data set;Second,the combination of convolutional neural network and class-activated mapping technology to improve it based on the Alex Net model and obtain the Laser Chip Net network;Third,the use of laser chip defect image data after image preprocessing Set to train the Laser Chip Net network;Fourth,design three sets of experiments to verify the effect of image preprocessing on the accuracy of the model,and the model’s excellent performance in laser chip defect classification and recognition and defect area location.The final results show that the network model proposed in this paper has high classification accuracy,and can locate defect regions without using manual annotation at the regional level,and the classification and positioning functions can be realized in a network,which is very fast and convenient.It can be seen that the model proposed in this paper is of positive significance for the defect detection of laser chips.更多还原