基于字典学习的机织物疵点图像等级评定研究

Study on image grading of woven fabric defect based on dictionary learning

针对机织物疵点图像等级自动评定问题,在应用字典学习方法对疵点图像稀疏表达的基础上提出一种等级自动评定方法,采用该方法分别对样本大小和字典原子个数进行优化,首先在机织物纹理图像上截取特定尺寸的疵点纹理图像和正常纹理图像,对正常织物纹理图像进行K-SVD字典学习,然后用学习得到的字典对疵点纹理图像进行重构,最后根据重构效果进行等级评定。实验结果表明:最佳子样本尺寸为128像素×128像素,最佳字典原子个数为256。该方法的自动评定结果相对于人工评定结果准确率达到了83.61%。

In view of automatic evaluation of the image grades of woven fabric defect,an automatic method for evaluating grades was processed based on the application of dictionary learning method to sparse expression of defective images.In this method,the sample size and the number of dictionary atoms were optimized respectively.Firstly,defect texture image and normal texture image of certain size were cropped on the woven fabric texture image,the K-SVD dictionary of the normal fabric texture was studied,and then the defective texture image was reconstructed by the learning dictionary.Finally,the grading was carried out according to the reconstruction results.The experimental results showed that the best subsample size was 128×128 pixels and the optimal number of dictionary atoms was 256,the accuracy of this method was 83.61%compared with the accuracy of the manual evaluation.