基于时间序列分形特征的织物瑕疵检测

为充分利用织物纹理特点，并大幅度减少计算量，将织物纹理图像的灰度值分别沿纵、横方向投影得到一维时间序列，提取分形特征。根据Fisher判别准则优选了2个具有一定互补性而又能最大限度区分正常与瑕疵样本的时间序列盒维数组成特征向量，对7种具有不同纹理背景的织物进行瑕疵检测。试验结果表明，在一定的阈值范围内，误检率和漏检率可控制在10%以内，表明采用这种方法在大幅减少维数估算计算量的同时能够有效区分正常与瑕疵纹理。

Fabric defect detection based on fractal feature of time series

In order to reduce the amount of computation,fractal features were extracted after the gray value of fabric texture image was projected on one-dimensional time series along vertical and lateral direction.Fisher's criterion was used to evaluate the discrimination between the defect samples and the normal samples.Two features of certain complementarity and great discrimination were chosen to detect defects on seven fabrics of different texture background.The experimental result showed that the false alarm tate and missing rate could be controlled below 10% within a certain range of threshold values,which indicated the proposed method was efficient in reducing the amount of fractal dimension computation greatly as well as differentiating normal and defect texture.