基于变换域纹理模式消隐与图像重建的一致纹理表面缺陷检测理论与方法

Defect detection in textured materials such as fabric and leather is a crucial but challenging task due to high variation of texture features. Traditionally, this work depends mainly on skilled human inspectors. However, human vision based inspection can be exhausting, neglectful and error-prone. This project address the defect detection in textured surfaces based on a highly effective scheme consisting of pattern removal and image reconstruction from the transform-domain. The basic idea is to attenuate the normal homogeneous regions, while accentuate the defective regions. To this end, several methods based on the spectral analysis, such as the Gabor transform (GT), the wavelet transform (WT), and the Fourier transform (FT) will be studied. In our previous research, a generalized Gabor filter (GGF) has been proposed that aims to minimize the number of filters required in the inspection so that less computational costs can be achieved. A GGF can be tuned freely to have either a form of traditional Gabor filter or a form of shifted elliptic Gabor filter. In this project, we further study how to model the GGF-filtered outputs for different types of textures by a stochastic process, and the evaluation criteria for the optimal GGF will then be developed and combined into an optimization model so that it can be solved by a hybrid genetic simulated annealing algorithm (GASA). As GT is known to be computationally expensive, methods based on optimal wavelet bases and wavelet-domain hidden Markov tree model (HMT) will be addressed as well. An optimal wavelet base can better characterize specific textures and have very high sensitivity to the abrupt changes in the texture structure caused by defects, while an HMT models both the non-Gaussian statistics and the persistence property across scale of wavelet coefficients by applying tree-structured Markov chains across scales. Therefore, an improved performance can be expected. Another research interest of this project is the unsupervised problem which has been less addressed in the literature by far. Generally, the reliabilities of (semi-)supervised methods might be drastically affected by image distortions such as the rotation, the scaling and the inhomogeneous illumination conditions in a real industrial environment. An unsupervised method does not require any images as reference template so that it avoids all the above problems. As the Fourier spectrum has been known to be ideally suited for describing the periodicity of an image, we attempt to remove periodic pattern by masking the associated dominant frequency components based on the gradient or Gaussian curvature analysis in the Fourier domain, and separate potential defects from the resultant image by multi-channel analysis. We believe that this project will make a significant contribution to real-time texture defect detection applications.

针对织物、皮料等材质表面由于纹理特征变动性造成的自动化视觉检测难题，开展基于变换域纹理模式消隐与图像重建的缺陷在线视觉检测理论研究，消除人工目检所具有的低效、低精度、乏味枯燥等致命弱点。针对传统Gabor变换检测方法中需要的滤波器数目多、效率低的问题，提出一种具有灵活配置形式的新型“扩展Gabor滤波器”，为此将在深入分析该新型滤波器纹理响应特性及参数配置方法的基础上，研究最优扩展Gabor滤波纹理缺陷检测算法。其次，将结合小波分析的高效性研究纹理匹配最优小波基的构造、基于小波域隐马尔科夫树模型的纹理表征机制，解决现有方法未能较好地实现纹理模式可辨尺度下的缺陷检测问题。同时，针对有监督、半监督检测方法对环境变化敏感、通用性差的问题，项目还将积极探索无监督缺陷检测问题，提出以Fourier谱几何特征为基础，结合频谱梯度与高斯曲率分析，实现自适应纹理模式消隐及检测。

纹理广泛存在于各类产品表面。纹理表面的缺陷视觉检测是一个重要且复杂的课题，目前仍亟待突破。项目研究内容主要包括：（1）基于扩展Gabor滤波器的检测方法。提出一种可配置成多种形态的新型扩展Gabor滤波器，建立了基于fisher准则的纹理响应评价指标，研究了滤波器的优化配置，结果表明该新型滤波器对纹理类型具有更好的适应能力，对于周期性、随机性纹理样本总体检测精度由传统方法的88.33%提高到96.67%。（2）基于小波域隐马尔科夫树模型（HMT）的织物纹理缺陷检测。通过对正常纹理与缺陷的小波系数分布规律的研究，发现纹理小波系数可以用HMT有效刻画，据此可估计各小波系数属于正常纹理的概率值，据此对小波系数进行修剪及重建，可将纹理与缺陷充分解耦，实现缺陷的检测与定位。对多种纹理样本的测试证实了该方法的有效性。（3）基于无抽样小波分解（UDWT）与Gumbel分布的缺陷检测方法。利用UDWT突破传统DWT缺乏平移不变性的局限，同时根据人类视觉发现机理，提出一种高效的小波域纹理特征表征方法，对特征分布规律的研究表明可将问题域映射到低维Gumbel参数空间，实现对缺陷的准确分离。与现有的基于Gabor小波、Fourier分析的无监督检测算法对比，该方法具有更高的检测精度及更低的误报率、漏检率，并具有满意的运算效率。（4）针对周期性纹理，提出一种以Fourier谱几何特征为基础的自适应纹理模式消隐及检测方法。无需参考样本，通过频谱高斯曲率分析，可直接自动分离出纹理的主频成分，得到去除背景的重建图像，从而将缺陷检测问题转化成普通的阈值分割问题。（5）基于深度学习的智能检测方法。提出一种支持纹理精确重建的新型DCGAN网络结构，可使重建结果仅包含缺陷信息，有效克服了传统方法对纹理和缺陷类型适应性差的缺点。项目资助发表SCI论文3篇，已投稿在审的SCI论文1篇；申请发明专利2项，培养硕士生2名（在读）。项目投入经费21万元，支出13.7558万元，各项支出基本与预算相符。经费结余7.2442万元，计划用于项目的后续研究支出。