随着现代制造业的飞速发展，传统的人工缺陷检测由于存在效率低、可靠性差等不足，因此机器视觉扮演者越来越重要的角色。本课题是由工业上利用线阵相机做缺陷检测的需求产生的，主要针对传统面阵相机系统无法解决的高精度、高速度、低光照度场景下的缺陷检测问题。总的来说，本文的工作主要包括以下三个方面：1.线阵扫描检测系统整体的搭建，包括机械结构、运动控制、以及图像处理算法等方面。针对采用时间延迟积分（TDI）技术的特殊线阵相机，由于其特殊的成像原理，本文针对其成像特性，设计了与之对应的机械结构与运动控制和相机行频率同步算法来提高最终成像质量。2.在线阵扫描检测系统整体结构搭建完毕后，本文选用了两个具体应用场景来进行测试，手机玻璃盖板边缘缺陷检测和金属印刷板移位缺陷检测，并分别给出了缺陷检测结果。为了提升缺陷检测的效率，本文用Adaboost和级联分类器训练了一个目标区域检测器，用来迅速找到目标区域并进行缺陷检测。针对线阵扫描检测系统的数据量大、算法实时性要求高等特点，本文选用了图形处理单元（GPU）并行加速来尝试解决，并分析了算法性能瓶颈，比较了采用共享存储和全局存储的不同。 3.本文尝试了从图像复原去运动模糊的角度来增强最终系统的输出图片质量。由于本文中的线阵扫描检测系统中的退化函数卷积核可以较准确估算出来，因此去模糊方法采用了单幅图片非盲去卷积。本文分别分析了相机响应函数（CRF）的非线性与复原结果中振铃效应的关系，相机感光芯片的噪声特性，信号相关噪声的方差稳定化操作（VST），BM3D去噪算法与Lucy-Richardson解卷积算法在线阵扫描系统中的应用，并给出了测试对比结果。
With the rapid development of the manufacturing industry, the eager demand of the AOI (Automatic Optical Inspection) technology arises. The TDI-CMOS line-scan system is of great importance for the products which require a large field of view and extreme high-precision also high speed. At present, defects inspection is mainly by naked eyes of experienced workers, which is slow, a waste of time and of low accuracy. So this thesis discusses some key issues in the implementation of the line-scan defects inspection systems. Firstly, this thesis stated the equipment which is enhanced by a TDI (Time- Delay Integration) line-scan camera with 16k pixels. This kind of special line array visual sensor can help to get a more signal-to-noise ratio images compared to the common single line array sensor. To get a clear image, the line frequency of the camera should be synchronized with the feedback signal of the servo motor rotary encoders.To satisfy line-scan defects inspection systems which need a short time for processing a large amount of data-flow, we propose an accelerated method running on the NVIDA CUDA platform. Then we make an experiment on the image segmentation method otsu using both the global memory and the shared memory of GPU to analysis the difference. Last, to improve the quality of the output image, we discussed the non-blind motion deblurring solutions adjust to thins systems. In this chapter, the role that nonlinear camera response functions have on image deblurring is discussed. Also the restoration of blurred images corrupted by Poissson noise using variance-stablizing transformation is considered. Then we present an efficient implementation using the BM3D denoising filter. After that we propose a combined deblurring framework using methods mentioned above.
