机织物密度对字典学习纹理表征的影响

Influence of woven fabric density on texture representation based on dictionary learning

为探讨纺织品表观质量的客观、智能评定方法,使用不同密度的机织物图像,采用子窗口样本获取方式作为学习样本,以离散余弦字典作为初始学习字典,选择基于最小二乘的字典学习算法求解用于表征织物纹理图像的字典,再通过字典元素的线性组合对织物图像进行重构。以均方误差为指标,首先讨论织物图像灰度值分布对字典学习算法重构误差的影响,然后对图像灰度值进行标准化处理,在此基础上探讨织物经纬密度对重构图像误差的影响。实验结果发现,当字典个数等于9时,织物密度在150~360根/10 cm之间,随着织物密度的增加,平纹重构图像的均方误差先变大,以后不再增加,而斜纹重构图像的均方误差增大。

In order to discuss an smart evaluation method for objective evaluation on fabric appearance quality,patches extracted from woven fabric images with different densities were used as training samples and discrete cosine dictionary was used as the initial dictionary of learning algorithm based on the least square method. The original woven fabric image samples can be reconstructed well by the dictionary by a linear summation of its elements. To evaluate the reconstruction performance,mean square error was selected as evaluation index. The influence of gray distribution of fabric images on the reconstruction error was discussed,and then the influences of density on the reconstruction error were discussed with the normalized image gray value. The experimental results show that when the number of dictionary atoms equals to 9,the mean square error of plain increases firstly and then remains within a certain range and the mean square error of twill increases with the increasing of warp and weft density from 150 to360 yarns/10 cm.