《led芯片检测系统视觉图像分析技术研究》由会员分享，可在线阅读，更多相关《led芯片检测系统视觉图像分析技术研究（81页珍藏版）》请在金锄头文库上搜索。
1、杭州电子科技大学 硕士学位论文 LED芯片检测系统视觉图像分析技术研究 姓名：李蒙 申请学位级别：硕士 专业：检测技术与自动化装置 指导教师：薛凌云 20091201 杭州电子科技大学硕士学位论文 I 摘 要 LED 芯片检测与分选是 LED 芯片生产中的重要工序。全自动 LED 芯片检测、分拣设 备中的关键技术之一就是机器视觉图像对准定位。视觉定位的过程是通过对待检测、分选 的 LED 芯片图像进行处理与分析，提取并输出图像特征（如边缘、轮廓及标识等） ，通过 这些特征信息的辨析，获取芯片晶粒的形状、位置、姿态等参数，并给伺服控制器提供运 动控制参数，完成芯片对准。 本文对 LED 芯片图像
2、分析，以期获得适用于 LED 视觉定位任务的图像处理技术与方 法。LED 芯片图像主要表现为含噪背景、芯片片基、电极三类景物，要实现 LED 视觉定 位任务，需对其进行图像分割、目标识别与定位。图像分割是为了获取芯片片基和电极两 类感兴趣区域。目标识别即分类过程，需特征提取与分析匹配。定位则是在提取的特征基 础上做参数计算，获取目标位置。本文具体研究内容如下： (1) LED 图像特征分析。探讨 LED 芯片图像特征，统计了 LED 芯片图像的灰度特征。 应用几何参数、区域与边缘的 7 个 Hu 不变矩开展了对两类电极和片基的特征提取与分析 工作。通过数据对比分析，选取可以很好的区分两类电极目
3、标以及片基目标的特征周 长、面积以及边缘的不变矩做为分类依据。 (2) 基于区域的图像分割技术。根据 LED 芯片图像的区域特性，应用 Otsu 双阈值和 基于直方图势函数标记的分水岭分割方法对 LED 图像进行分割。 实验结果表明， 分水岭分 割较 Otsu 双阈值分割效果更好。Otsu 双阈值分割能够将电极、片基和背景三类区域提取 出来，但片基区域内部含有误判为背景的区域。标记分水岭分割方法采用直方图多阶势函 数自动衰减实现标记提取，在标记基础上的分水岭分割能够将电极、片基和背景区域分离 提取， 片基内部不会出现误判为背景的区域。 而且标记分水岭分割在 LED 芯片图像的应用 中较好的抑制
4、了过分割现象。 (3) 基于边缘检测的图像分割技术。 应用 Canny 边缘检测算子对 LED 芯片图像进行边 缘检测，获取 LED 芯片图像的边缘信息，并在 Canny 边缘基础上对目标区域进行种子填 充与区域生长，提取区域信息。利用面积误分率评判分割效果，由 Canny 算子所得分割结 果比采用基于区域的 Otsu 双阈值分割和直方图势函数标记的分水岭分割结果面积误分率 更低。 (4) 基于Hough变换的特征描述方法。应用Hough变换（HT）实现对LED芯片图像中直 线和圆的检测。研究随机Hough变换（RHT）机制，在RHT基础上提出了RHT与最小二乘 法（LSM）相结合的算法（RH
5、T- LSM）。该算法首先通过RHT选取合适的点集，再对选 取的点集用LSM拟合，使得该算法具有RHT的较强抗干扰能力和LSM对点集拟合的残差最 小特性。使用RHT- LSM对已知参数的理想图和缺陷图进行直线和圆的检测实验，验证了 杭州电子科技大学硕士学位论文 II 算法的有效性和稳定性。在LED芯片图像的应用中，采用RHT-LSM可获得精确的LED芯片 形位参数。 本文通过对LED芯片检测系统视觉任务相关的图像分析技术的研究，能够实现LED芯 片图元分类，计算较精确的形位参数。 关键词：图像分析，LED 芯片，RHT-LSM，分水岭 杭州电子科技大学硕士学位论文 III ABSTRACT M
6、apping and sorting is an important procedure in LED die production. Positioning through machine vision is one of the key technologies for automatic LED mapping sorter. The process of LED wafer visual inspection is that, the LED die features such as edge, contour, mark and so forth are extracted and
7、exported through images processing and analyzing, then LED die parameters such as shape, position, posture and motion control parameters used for servo controller are obtained ,thus LED die alignment is completed. According to the analysis of LED die image, this paper explores image processing techn
8、iques and methods for LED visual location tasks. LED images involve noisy background, light-emitting zone and electrodes. Image segmentation, recognition and positioning should be done to LED die image in order to achieve the vision task. Image segmentation aims at two areas of interest: light-emitt
9、ing zone and electrodes. Recognition is the process of classify, which combines feature extraction, analysis and matching. Positioning is parameters computing with the features extracted. The main contents of this paper are as follows: (1) Feature analysis on LED die image. The features of LED die i
10、mage are discussed. Gray features are compiled and analyzed. The features on two kinds of electrodes are extracted and analyzed with the application of geometric parameters, 7 Hu Invariant Moments of region and edge. According to experiments, distinctive perimeters (by types of electrodes and light-
11、emitting zone), area and edge Invariant Moment features, are selected to provide a foundation for target classification. (2) Region-based Image Segmentation. According to the regional features of LED die image, Otsu dual-threshold segmentation and watershed segmentation, marked by histogram potentia
12、l function, are applied to LED image. The simulation results indicate that the latter method is better than the former one. Although the regions of electrode, light-emitting zone and background could be extracted by Otsu dual-threshold, there are partly aliasing of light-emitting zone and background
13、. Multi-stage potential function of histogram, which can be automatically attenuated, is involved in watershed segmentation to obtain credible markers. The result of watershed segmentation can partition electrodes, light-emitting zone and background area with the markers. Moreover, the marked waters
14、hed method is better in suppressing the phenomenon of over-segmentation in the LED die image. (3) Edge-Based Image Segmentation. The Canny operator is applied in LED die image to obtain the edges information. At the same time, seeds are filled based on the Canny edges and 杭州电子科技大学硕士学位论文 IV regions a
15、re grown, thus the regional information is extracted. Percentage area mis-classified (PMAC) is used to evaluate the segmentation results, indicating that Canny method obtained lower PMAC than dual-Otsu and marked-watershed method. (4) Feature description method based on Hough Transform. Hough Transf
16、orm (HT) is used to detect straight line and circles in LED die image. The principles of Random Hough Transform are discussed, and then the method combined with the RHT and Least Squares Method (RHT-LSM) is proposed. The RHT-LSM chooses proper pixel sets with RHT and uses LSM to fit the pixel sets so that the RHT-LSM combines the strong anti-noise ability of RHT and the small residual error property of LSM. It is verified efficie
