产品的表面质量是其商品价值和使用价值的重要保障，也与企业的利益和信誉息息相关。基于机器视觉的表面缺陷检测技术能在高速生产线上剔除不合格产品，在保证产品质量的同时节省大量的人力成本。所以，视觉表面缺陷检测技术有着重要的研究意义和很高的实用价值，也是智能制造技术的重要组成部分。本文针对视觉缺陷检测中尚未得到解决的若干关键技术问题展开研究，在考虑缺陷检测算法准确性的同时也充分考虑了高速生产线对实时性的要求。本文的主要工作如下： (1)研究显著性检测模型在缺陷检测领域的应用。提出了一个使用增强、互补或削弱等策略，融合图像多个线索的缺陷显著性检测模型。与传统的显著性检测模型相比，该方法具有更强的特征描述能力，可以检测出复杂多变的缺陷。针对磁瓦缺陷检测问题，改进U-Net卷积神经网络检测缺陷。在磁瓦缺陷检测任务中，该深度学习模型优于传统缺陷检测模型。 (2)针对卷积神经网络参数规模大、计算量大、依赖于昂贵硬件等问题，提出一种紧凑型卷积神经网络。使用小卷积核和深度方向卷积修改残差瓶颈结构，减少了网络参数量、计算量和时间代价，使网络能在低功耗计算机中运行，降低缺陷检测系统的硬件成本。本章还使用了数据增广方法并采用预训练策略来解决缺陷图像数据严重不足的问题，采用了渐进式训练策略改善网络不收敛和微小缺陷识别率低的问题。本文提出的端到端的紧凑模型不需要人工设计特征、预处理、后处理等繁琐操作，且能在低性能的CPU上进行实时缺陷检测。在热轧带钢、铁轨等多个表面缺陷数据集上的测试结果表明，与多种相关方法相比，本文提出的模型在最少的参数量的情况下，可以获得最高的准确率和最快的检测速度。 (3)针对大幅面、高分辨率图像的微小缺陷检测问题，设计了一种多任务型卷积神经网络。利用快速下采样层，使网络在保留更多的图像细节的同时，有效地减少网络对显存的占用；采用多任务监督策略，使得缺陷二分类和多分类的性能都得以提升；使用常规增广、随机掩模和条件Inpaint等数据增广方法，解决了高分辨图像样本不均衡和数据不足等问题。本章还分析了不同类型的损失函数对网络性能的影响，在高分辨率的布匹缺陷数据集上验证了本文提出的多任务型卷积神经网络的有效性。与其它多种相关方法相比，在有限的硬件资源下，本章所提出的方法获得了最高的准确率，且算法的运算速度达到了实时检测高分辨率图像的需求。 (4)针对无法通过二维图像进行可靠检测的微小三维形变缺陷，设计了两种不同的高精度线结构光三维视觉传感器。通过改进Steger算法和使用FPGA硬件加速等方法，提高了三维视觉传感器的扫描频率。本文提出的由粗到精的重心法能在普通CPU上达到280 Hz的扫描频率，可以较好地满足高频率扫描应用场景的需求。本章还提出了基于特征直线加速的倒角匹配方法和改进FDCM方法来配准三维深度图的方法，可以检测亚毫米级别的三维形变缺陷并大幅减少FDCM方法的计算量，使三维缺陷检测系统能够达到实时检测的要求。 本文的最后还简单总结了本文的工作成果，并讨论了在本文的研究基础上可以进一步开展的研究工作。 关键词：视觉表面缺陷检测，显著性检测，卷积神经网络，线结构光，三维配准

 The surface quality of products guarantees the commodity value and use-value of the products, and it is closely related to the interests and reputation of enterprises. Machine vision based surface defect inspection technology can be used to filter out unqualified products on high-speed production lines, and this technology also saves a lot of labor cost on quality inspection. This dissertation focuses on several key technical problems that have not been solved in visual defect inspection. We are committed to improving the accuracies of defect inspection algorithms, and also fully consider the realtime requirements of high-speed production lines. The major works and contributions of this dissertation are as follows: (1)The feasibility of applying the saliency detection model to surface defect inspection is studied. A defect detection model fusing multiple saliency cues with strategies including enhancement, complementation, or weakening is proposed. Compared with the traditional saliency detection model, this method has stronger representation ability and can detect complex and variable defects. And an improved U-Net is used to inspect magnetic tile defects. This deep learning approach achieves the best detection result in the magnetic tile defect inspection task, against other traditional models. (2)To solve the problem of Convolutional Neural Networks (CNNs) which have large parameter size, high computational cost and dependency on expensive hardware devices, a compact CNN is proposed. We modify residual bottlenecks with small kernels and depthwise convolutions, which significantly reduce network parameters, computation and time cost, making the defect inspection network run in low-cost computers. This chapter also uses the data augmentation method and the pretraining to solve the severe shortage of defect images. And a gradual training strategy is adopted to handle the network non-convergence and improve the low accuracy of tiny defects. The proposed end-to-end compact model does not require extracting hand-craft features, preprocessing, and post-processing. It can also inspect surface defects on low-performance CPU in real-time. We test the proposed model in several surface defect datasets such as hot-rolled strip and rails. The results show that the proposed model achieves the highest accuracy and the fastest inspection speed in the case of minimum parameter size against the related methods. (3)A multi-task CNN is designed to detect small defects in high-resolution images. The fast down-sampling layer effectively reduces the graphics memory footprint while retaining image details; multi-task supervision improves the binary and multi-class defect classification performance; and data augmentation methods such as conventional augmentation, random mask, and conditional inpaint solve the problem of data imbalance and data insufficiency. The effect of different loss functions on network performance is also analyzed. The effectiveness of the proposed multi-task CNN is tested on the high-resolution fabric defect dataset. Compared with other related methods, the proposed model achieves the highest accuracy with limited hardware resources, and the inference speed meets the needs of real-time inspection of high-resolution images. (4)Two kinds of high-precision linear structured light 3D vision sensors are designed to solve the problem that 2D images cannot reliably detect 3D slight deformation defects. By improving the Steger algorithm or using the FPGA acceleration, the scanning frequency of the 3D vision sensor is improved. The proposed method can achieve 280 Hz scanning frequency on a common CPU and can satisfy the requirement of highfrequency scanning scenarios. In this chapter, characteristic line chamfer matching and improved FDCM method are proposed for registration of 3D depth images, which can inspect sub-millimeter level 3D deformation defects and greatly reduce the computational complexity of FDCM method, so that the 3D defect inspection system can meet the real-time requirements. At the end of this dissertation, the conclusion is given and further research work is analyzed. Keywords: Vision based surface defect inspection, Saliency detection, Convolutional neural network, Linear structured light, 3D registration