【作者】 王宁；

【导师】 胡广华；

【作者基本信息】 华南理工大学 ， 机械工程， 2021， 硕士

【摘要】 大量实际产品表面均具有某种天然/人工纹理模式。从局部视野来看,纹理模式可表现出复杂的像素灰度空间分布,导致如何从纹理背景中定位和分离出各种潜在缺陷成为一个具有挑战性的问题。同时,生产技术的进步及人工成本的升高使得从工业现场收集足量类型齐全的缺陷样本几乎不可行,意味着现有的有监督检测方法存在较大的局限性。为此,本文基于频谱分析及深度学习,探索了若干无监督纹理缺陷检测方法,具体研究内容如下:(1)针对规则度较高、具有一定周期性变化的纹理表面缺陷,研究了一种基于频域梯度分析的缺陷检测方法,能够在没有先验知识的情况下直接检测不同类型的纹理表面缺陷。首先利用待测图像在频域中的梯度信息,屏蔽周期性纹理的频谱分量。接着利用环形Gabor滤波器组进一步提取频谱图像中的缺陷纹理特征,综合各个滤波器所提取到的缺陷纹理特征,获取检测结果图像。在不同规则程度的数据集上进行了实验,验证了方法的有效性。(2)针对基于正常样本无监督学习的缺陷检测方法在训练过程中容易过拟合的问题,研究了一种基于迁移学习的自监督学习缺陷检测方法。利用预训练网络模型作为编码器,提取更为抽象的图像特征信息。再将提取到的特征输入到解码器中,生成重建图像。为了增强网络模型对于异常样本的泛化能力,通过在正常样本中加入噪声作为异常样本输入到网络模型中,使网络模型能够生成与输入的正常样本相同的重建图像。通过在训练过程中联合对于正常样本与异常样本的重建,使得网络模型具有更强的泛化能力。与现有的缺陷检测方法进行对比,研究的方法表现出了较为优异的检测性能。(3)针对现有的网络模型难以有效检测复杂、欠规则的纹理表面缺陷的问题,研究了一种基于图像修复的无监督学习缺陷检测方法。在训练过程中人为地在待测样本中设置缺失区域,由网络模型来预测、填补缺失区域的图像内容,使得网络模型能够有效习得正常纹理的分布,从而避免网络模型对于缺陷区域高度相似的重建,导致难以分割缺陷区域。在测试阶段中逐块地遮掩输入的样本图像,再输入到网络模型中进行重建。经过实验验证,研究的方法能够有效检测欠规则纹理表面的缺陷。更多还原

【Abstract】 A large number of real product surfaces have some kind of natural/ artificial texture pattern.From the local field of view,the texture pattern can represent a complex spatial distribution of pixel grayscale,which makes it a challenging problem to locate and isolate various potential defects from the texture background.At the same time,the progress of production technology and the increase of labor cost make it almost impossible to collect sufficient defect samples with complete types from the industrial site,which means that the existing supervised detection methods have great limitations.Therefore,based on spectrum analysis and deep learning,several unsupervised texture defect detection methods are explored in this paper.The specific research contents are as follows:(1)For the texture surface defects with high regularity and certain periodic changes,a defect detection method based on frequency domain gradient analysis is studied,which can directly detect different types of texture surface defects without prior knowledge.Firstly,the spectral component of periodic texture is shielded by using the gradient information of the measured image in the frequency domain.Then,utilizing ring Gabor filter banks the defect texture features are further extracted from the spectral images,and the detection results are obtained by synthesizing the defect texture features extracted by various filters.The validity of the method is verified by experiments on data sets of different rule degrees.(2)Aiming at the problem that defect detection method based on normal sample unsupervised learning at the training stage is easy to overfit,a self-supervised learning defect detection method based on transfer learning is studied.The pretrained network model is used as the encoder to extract more abstract image feature information.Then the extracted features are input into the decoder to generate the reconstructed image.In order to enhance the generalization ability of the network model for abnormal samples,noise is added to the normal sample as the abnormal sample to input into the network model,so that the network model can generate the same reconstructed image as the input normal sample.By combining the reconstruction of normal samples and abnormal samples in the training process,the network model has stronger generalization ability.Compared with the existing defect detection methods,the method presented in this paper shows a better performance.(3)Aiming at the problem that existing network models are difficult to detect complex and irregular texture surface defects effectively,an unsupervised learning defect detection method based on image repair was studied.In the training process,missing areas are artificially set in the sample to be tested,and the image content of the missing areas is predicted and filled by the network model,so that the network model can effectively learn the distribution of normal texture,thus avoiding the reconstruction of highly similar defect areas by the network model,which leads to the difficulty in segmenting the defect areas.In the test stage,the input sample image is obscured block by block,and then input into the network model for reconstruction.The experimental results show that the proposed method can effectively detect the defects on irregular textured surfaces.更多还原