Abstract: Aiming at the problem of unobvious contrast between the complex background and the defect in color fabric images, which makes the defect detection difficult, a color fabric defect detection method based on local steering Gaussian kernel is proposed. The input RGB fabric image is converted into the CIEL*a*b color space and the gradient features in each color channel are extracted to construct a structure-adaptive modulation factor by Singular Value Decomposition（SVD）. The shape and the size of the Gaussian function are modulated by the obtained modulation factor, and local steering Gaussian kernel features in the local neighbor of each pixel are extracted to establish a feature matrix for representing the local structures of the input color fabric image. The matrix cosine similarity metric is employed to measure the similarity of kernel feature matrices for obtaining the desired defect map. The experimental results indicate that the proposed method can effectively reveal the singular structures in color fabric images and achieve compelling detection results in comparing with other competitors.

Key words: local steering kernel, matrix cosine similarity, fabric defect detection, defect map

摘要： 针对复杂色织物背景与疵点对比度不明显而导致疵点难以检测的问题，提出了一种基于局部可控高斯核的织物疵点检测方法。该方法将输入的RGB色织物图像转换到CIEL*a*b颜色空间，分别在L、A和B颜色通道下，提取各个通道下的图像梯度特征进行奇异值分解，以构造具有结构自适应的调制因子。利用获得的调制因子调制高斯核函数的形状和大小，提取每个像素局部邻域内的可控高斯核特征构成特征矩阵，以表征色织物图像的局部结构。利用矩阵余弦相似性度量核特征矩阵之间的相似性，建立织物疵点的映射图。实验结果表明，提出的疵点检测方法能有效表征色织物的奇异性结构，相比于其他方法，该方法对不同纹理类型的色织物疵点具有较好的检测性能。

关键词: 局部可控核（LCK）, 矩阵余弦相似性, 织物疵点, 映射图