Concerning the measurement of the similarity of non-local means, a method based on dictionary learning was presented. First, block matching based local pixel grouping was used to eliminate the interference by dissimilar image blocks. Then, the corrupted similar blocks were denoised by dictionary learning. As a further development of classical sparse representation model, the similar patches were unified for joint sparse representation and learning an efficient and compact dictionary by principal component analysis, so that the similar patches relevency could be well preserved. This similarity between the pixels was measured by the Euclidean distance of denoised image blocks, which can well show the similarity of the similar blocks. The experimental results show the modified algorithm has a superior denoising performance than the original one in terms of both Peak Signal-to-Noise Ratio (PSNR) and subjective visual quality. For some images whose structural similarity is large and with rich detail information, their structures and details are well preserved. The robustness of the presented method is superior to the original one.%针对非局部均值中相似度的衡量问题,提出了一种基于字典学习的度量算法.首先利用局部像素群块匹配方法消除不相似的图像块带来的干扰,然后对含有噪声的相似块采用字典学习的方法降噪.与经典的字典学习不同的是,对相似块采用联合稀疏编码的思想,利用主成分分析法学习一个高效紧字典,保留相似块间的相关性信息.采用降噪后的图像块间的欧氏距离计算像素间的相似度,能更好地反映相似块的相似性.实验结果表明,所提出的方法在峰值信噪比和视觉效果方面都优于传统算法,尤其对含有较多细节且结构相似性强的图像,细节和纹理部分的保持效果更好,算法的鲁棒性也优于传统算法.
展开▼
