Visual quality control on raw textile fabrics is a vital process in weaving factories to ensure their exterior quality (visual defects or imperfection) satisfying customer requirements. Commonly, this critical process is manually conducted by human inspectors, which can hardly provide a fast and reliable inspection results due to fatigue and subjective errors. To meet modern production needs, it is highly demanded to develop an automated defect inspection system by replacing human eyes with computer vision.,As a structural texture, fabric textures can be effectively represented by a linearly summation of basic elements (dictionary). To create a robust representation of a fabric texture in an unsupervised manner, a smooth constraint is imposed on dictionary learning model. Such representation is robust to defects when using it to recover a defective image. Thus an abnormal map (likelihood of defective regions) can be computed by measuring similarity between recovered version and itself. Finally, the total variation (TV) based model is built to segment defects on the abnormal map.,Different from traditional dictionary learning method, a smooth constraint is introduced in dictionary learning that not only able to create a robust representation for fabric textures but also avoid the selection of dictionary size. In addition, a TV based model is designed according to defects' characteristics. The experimental results demonstrate that (1) the dictionary with smooth constraint can generate a more robust representation of fabric textures compared to traditional dictionary; (2) the TV based model can achieve a robust and good segmentation result.,The major originality of the proposed method are: (1) Dictionary size can be set as a constant instead of selecting it empirically; (2) The total variation based model is built, which can enhance less salient defects, improving segmentation performance significantly.

中文翻译：

---

具有鲁棒纹理表示和总变异的纺织织物缺陷分割

纺织原料的视觉质量控制是织造工厂确保其外观质量（视觉缺陷或缺陷）满足客户要求的重要过程。通常，这个关键过程是由人工检查员手动进行的，由于疲劳和主观错误，很难提供快速可靠的检查结果。为适应现代生产需求，迫切需要开发一种以计算机视觉代替人眼的自动化缺陷检测系统。织物纹理作为一种结构纹理，可以通过基本元素（字典）的线性求和来有效表示。为了以无监督的方式创建织物纹理的鲁棒表示，对字典学习模型施加了平滑约束。当使用它来恢复有缺陷的图像时，这种表示对缺陷是鲁棒的。因此，可以通过测量恢复版本与其自身之间的相似性来计算异常图（缺陷区域的可能性）。最后，建立基于总变异（TV）的模型来分割异常图上的缺陷。与传统的字典学习方法不同，字典学习中引入了平滑约束，不仅能够为织物纹理创建一个鲁棒的表示，而且避免选择字典大小。此外，根据缺陷的特征设计了基于TV的模型。实验结果表明：（1）与传统字典相比，具有平滑约束的字典可以生成更稳健的织物纹理表示；(2) 基于 TV 的模型可以实现稳健和良好的分割结果。, 该方法的主要独创性是： (1) 字典大小可以设置为常数而不是凭经验选择; (2) 建立了基于总变异的模型，可以增强不太显着的缺陷，显着提高分割性能。