基于稀疏编码字典学习的疵点检测  

Defect detection based on sparse coding dictionary learning

作　　者：刘绥美 李鹏飞[1] 张蕾[1] 张宏伟[1] 张缓缓[1] 景军锋[1] 

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西安工程大学学报》2015年第5期594-599,共6页Journal of Xi’an Polytechnic University

基　　金：国家自然科学基金资助项目(61301276)

摘　　要：为了快速准确地实现背景纹理复杂织物的疵点检测,改善传统算法计算量大的缺点,提出基于稀疏编码字典学习的疵点检测算法.首先利用Radon变化对图像进行倾斜矫正,减小像素信息处理误差,再使用Gabor滤波器对矫正后图像滤波,消除噪声影响.接着对预处理后的图像,以一定尺寸窗口,滑动选取图像块构建输入样本集,采用K-SVD算法对无瑕疵样本集合进行字典学习,得到稀疏系数并重构,进而取得水平、垂直投影特征矩阵.最后利用已得到的字典与稀疏系数对待检测样本重构,求得其相对应的特征矩阵,并用结构相似法最终确定疵点区域.实验表明,该算法检测时间短,效率较高,平均可达92.3%.In order to detect defect of the complex background texture fabric fast and accurate- ly, and decrease the amount of computation of traditional algorithm, an algorithm based on sparse coding dictionary learning is proposed. Firstly, Radon skewness correction is adopted to reduce the pixel information processing error, and then Gabor filter is used to eliminate noise after correction. Secondly, sliding window with certain size is applied to select the image block in preprocessed-image and build input sample matrix. Dictionary and sparse coefficient of flawless sample matrix are obtained via K-SVD. Horizontal and vertical projection feature matrixes are calculated after sparse reconstruction. Thirdly, test sample achieved sparse reconstruction through the dictionary that has been treated with flawless image. Projection matrixes of test image is obtained as the same way and used to build corresponding horizontal and vertical feature matrixes to flawless image, respectively. Then the defective region of image is spotted. Experiments show that the proposed algorithm could efficiently detect defects with shorter time and average detection accuracy reaches 92.8%.

关 键 词：疵点检测 稀疏编码 K-SVD字典学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]