Abstract:

Image denoising is a typical ill-posed problem, the latent clean image corrupted by noise is unknow, to solve the problem, prior information of the image should be considered in the target function. In order to use the prior information of external clean image for image denoising, an image denoising algorithm based on non-local clustering with external clean image and internal noise image is proposed. The patches of external clean image are combined with the internal noise patches to obtain a clustering sparse dictionary. A global block matching method is adopted to obtain the estimation of sparse coefficients of the ideal image. Based on clustering dictionary and estimated sparse coefficients, image denoising is done by sparse reconstruction based on compressed sensing technology. The experimental results indicate that the proposed method provides a better ability on suppression of the mosaic effect caused by the partition operation, preserve more details of image and is visually more pleasant. The combination of ideal image priori source further improves the ability of the algorithm, making it suitable for image denoising task even under heavy-noise deterioration.

Key words: sparse prior, nonlocal clustering, image denoising, heavy noise

摘要：

图像去噪是一类典型的病态（ill-posed）逆问题求解，噪声掩盖下的真实图像并不确定，需要引入先验信息缩小病态问题的求解范围。为了将外部干净图像的先验信息引入去噪进程，提出了一种基于外部干净图像与内部噪声图像稀疏先验的非局域聚类图像去噪算法，通过联合外部干净图像与内部噪声图像的图像块得到类稀疏化表达字典；通过全局的相似块匹配，得到理想图像的稀疏系数估计；基于类字典和估计的稀疏系数，采用压缩感知技术的稀疏重建方法实现图像去噪。实验表明，与传统的非局域稀疏聚类图像去噪方法相比，所提算法显著降低去噪块效应，在保留更多细节的同时，图像平坦区域过渡更加自然；而理想图像先验来源的扩展则进一步提高了算法在强噪声下的去噪性能，对强噪声具有更强的抑制能力。

关键词: 稀疏化先验, 非局域聚类, 图像去噪, 强噪声