基于字典学习和原子聚类的图像去噪算法  

Image denoising based on dictionary learning and atom clustering

作　　者：孙挺[1] 王华东[1] 耿国华[2] 

机构地区：[1]周口师范学院计算机科学与技术学院,河南周口466001 [2]西北大学可视化研究所,西安710069

出　　处：《计算机应用研究》2016年第7期2236-2240,共5页Application Research of Computers

基　　金：河南省科技发展计划科技攻关项目(122400450356);河南省科技发展计划软科学项目(132400410927)

摘　　要：针对图像去噪过程中会导致细节和纹理结构信息丢失的不足,提出了基于字典学习和原子聚类的图像去噪算法。该算法利用含噪图像通过字典学习算法得到自适应的冗余字典,然后提取字典中每个原子的HOG特征和灰度统计特征构成特征集,并利用原子的特征集将冗余字典中的原子分成两类(不含噪原子和噪声原子),最后利用不含噪原子恢复图像,达到去噪的目的。实验结果表明,提出的算法无须知道噪声的先验信息,峰值信噪比好于现有的流行算法,且能较好地保持图像细节和纹理结构信息,提高了视觉效果。For the shortcoming of losing detail and texture structure with image denoising processing, this paper proposed the image denoising algorithm based on dictionary learning and atom clustering. Firstly, it obtained adaptive redundant dictionary by noised image over dictionary learning. Then structured the set of features by HOG features and histogram of gray features, and clustered atoms of redundant dictionary to two classes by the set of features. Finally, it restored image by denoised atoms and removed noise. The experimental results demonstrate that the prior knowledge of the noise is unnecessary and the peak signal to noise ratio （PSNR） value of the proposed algorithm is better than state-of-the-art algorithms, while the proposed algo- rithm can well preserved the detail information and texture structures in the denoised image, improving the visual effect.

关 键 词：字典学习 稀疏表示 冗余字典 K-均值聚类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]