【作者】 姜立芳；

【导师】 刘泊；

【作者基本信息】 哈尔滨理工大学 ， 测试计量技术及仪器， 2003， 硕士

【摘要】 本文从经济实用的角度出发，设计了一套工业视觉检查系统。工业视觉大致可分为用于检查、定位和装配三类，本系统属于检查类，是对工业零件进行分类、识别的自动检查装置。在工业生产中，工业环境和照明条件的确定性使得快速准确地判别分类出工业零件的想法成为可能，本系统就是在这种想法下应运而生的。由于在汽车零件生产线上分类产品并进而检查产品合格与否的场合中，要求整个过程在生产线上自动完成，不允许脱线做人工检查，这一检查工作可由本项目研制的工业视觉检查系统自动完成，使得生产更为高效连续。本系统采用图像传感器摄取原始信息，避免了与对象的直接接触和人工检查，所以更符合工业生产在线检查的要求。而且在编制软件时可根据工业零件的形状设计简化的步骤，可达到高效快速的目的。在本课题中，主要进行了对圆形和方形零件分类识别系统的研究和设计，针对的对象是工业零件的强度图像，采用人机对话方式来对零件进行识别。在系统软件设计部分中，首先是对所选零件进行模式识别，包括图像预处理、特征提取和分类器设计三个阶段，其中在图像预处理阶段本系统主要做的工作有：点运算、图像增强、正交变换、边缘提取和边缘增强、轮廓跟踪等。由于视觉系统本身就是一个神经系统，故本文所设计的分类器采用BP神经网络，其具有一些传统技术所没有的优点。但随着应用的广泛，BP网络存在的问题也日益显现出来，主要有：易形成局部极小而得不到整体最优、训练易陷入瘫痪、收敛速度很慢等缺点。本系统针对BP算法的局限性，给出了一种优化的BP算法，采用经过大量实验总结出的经验公式来确定隐层神经元的个数，并选取了一种新的误差平方和函数，该函数的特点是对一些可能的异常点的误差权值设计的较小，从而降低了异常值误差带来的影响，便于模拟出真实的函数关系。采用有导师学习方式，根据样本对在训练过程中的情况，系统地调整其参数。将上述算法应用于对工业零件的识别当中，相对于传统的神经网络可缩短识别时间2.8秒，而且识别正确率可达到81％以上。在对所选零件识别分类之后，本课题又设计了一套零件缺陷检测子系统，缺陷检测正确率可达到91％以上。实验结果表明，若将该系统应用于工业生产线上，可有效地提高生产率和保证产品的质量。更多还原

【Abstract】 From the view of economy and utility, In this paper, a suit of examination system of industrial vision is designed. Industrial vision can be divided into three classes: inspecting, orientation and setting-up. This system belongs to the class of inspecting, this system is a set of automatic inspecting devices, which is used to classify and recognize industrial parts. During the course of industrial manufacture, the determinacy of the industrial environment and illuminating conditions makes the ideas possible that industrial parts can be distinguished and classified rapidly and accurately. The system introduced below is the result of this idea. In automobile parts product line, if we want to classify products and check them whether they are up to grade or not, all process is required to be finished automatically in the process line and is not permitted to be checked manually on the off-line. The examination system of industrial vision can finished this work automatically and make production faster and more efficient. This system gets original information by image sensor, which avoids direct contact with objects being processed and manual inspection. So this system can meet the requirement that industrial parts should be inspected on-line. At the same time, when software is programmed, simple steps can be designed according to the shape of industrial parts to attain the high-speed and high-efficient purpose. In this task, we presented and designed a system that can classify and recognize circular and quadrate parts and the intension image of industrial parts is the object of the system, human-computer communication mode is adopted to identify the parts. In the part of software design, pattern recognition of selected parts is firstly carried out, it includes three phases: image pretreatment, feature extraction, systematizer design. In the phase of image pretreatment, the main jobs of this system includes dot operation, image swell, positive chiasma transform, edge extraction and edge swell, outline track, etc. Because the visual system itself is a neural system, systematizer designed in the paper adopts BP neural network to accomplish computer image identification, the system has some advantages over the traditional one, but with the extensive application of BP neural network, the problems existing in BPneural network come forth increasingly. There are some main problems, such as easily forming local extraordinary smallness and not having integral extraordinary excellence, training easily into paralysis, quite slow convergence speed, etc. Aiming at overcoming the limitation of the BP algorithm, this paper brings forward a sort of improved BP algorithm. A experienced equation which is summarized by many experiments is used to determine the number of mesosphere nerve cell and a sort of new square-sum function of errors is adopted. Its characteristic is that weight errors of possible exceptional point is less. Accordingly, the effect of errors of possible exceptional point is reduced, which make actual function relation simulation easier. With having tutor study style, this system adopts a sort of new error function to adjust its parameter systemically according to the status of sample in the course of training. When the algorithm is applied in the identification of industrial parts, comparison with the traditional BP neural network the recognition time will be shortened 2.8 second, and the recognition accuracy can reach more than 81%. After the classification and recognition of selected parts, a disfigurement inspection subsystem is also designed by this task. The accuracy of disfigurement inspection can reach more than 91%. The experiment result indicates, if applying this system to industrial production line can improve the productivity and guarantee the quality of products.更多还原