基于非局部正则化的图像去噪

Image denoising based on non-local regularization

图像复原是图像处理中一个重要的研究课题。大部分图像复原算法,都只是利用图像单个像素点,或某一邻域内的灰度和梯度信息。如果在图像复原模型中能够超越邻域,而更大范围地利用图像内容本身的信息,将可能更有效地改善复原质量。实际上,大多数自然图像中,其内容具有自相似特性。基于非局部的图像处理方法,就是以这类自相似特性为出发点的。提出一种改进的基于非局部的正则化图像去噪算法,主要针对非局部处理方法中的权值计算作了改进处理。与原始算法不同,采用EMD来进行图像子块间的相似度计算,并结合图像的方向信息分类子块区域,以解决计算量过大的问题。

Image restoration is an important research subject in image processing. Image gray and gradient magnitude based on single pixel or a neighbour is used in most of image restoration algorithm. If these information, which came from more wideranging image region, can be used for image restoration model, it is possible to improving restoration quality. In reality, there are self-similarity about image content in natural image. It is a foundation for non-local image processing method. This paper presented an improved image denosing algorithm based on non-local regularization. Designed the improvement for computing weight which camed from non-local regularization. Used the earth mover＇ s distance to compute the similarity between image sub-region. At the same time, used to solving expensive compute, the image direction information to classify image sub-region.