为解决传统彩色图像去噪算法容易出现细节模糊、伪色彩及去噪效果不佳等问题, 文中提出了一种基于本征图像分解的稀疏表示彩色图像去噪算法。利用本征图像分解良好的色彩保持和细节恢复等优点, 将含噪彩色图像分解成反映图像真实颜色特征的反射率部分和反映图像亮度特征的光照率部分。一方面, 反射率部分仅含有部分孤立噪声点且是具有分段平滑特性的彩色图像, 因此文中采用在去除彩色图像轻度污染方面表现良好的基于稀疏表示的彩色图像去噪算法对其进行处理。另一方面, 光照率部分包含了主要噪声成分且是具有较强稀疏性的灰度图像, 因此文中采用能够保持图像细节的非局部集中稀疏表示灰度图像去噪算法对其进行处理。为了有效地求解所提算法, 文中结合正交匹配追踪法和软阈值法设计了一种新的数值解法。数值实验结果表明, 新算法明显优于经典的彩色图像去噪算法。以256×256的Boat图像为例, 在噪声方差等于20时, 新算法的PSNR值比K-SVD算法和NCSR算法分别提高了1.7 dB和067 dB, SSIM值比K-SVD方法和NCSR算法分别提高了0.11和0.09。文中所提算法在提高彩色图像去噪效果的同时能够有效地保留图像细节, 在视觉效果和客观评价指标等方面均优于传统的ROF算法、K-SVD算法和NCSR算法。

In order to solve the problems of traditional color image denoising algorithm, such as detail blur, false color and poor denoising effect, a sparse representation color image denoising algorithm based on intrinsic decomposition is proposed. By utilizing the intrinsic image decomposition advantages of good color retention and detail recovery,the noisy color image is decomposed into reflectance and shading which reflect the true color feature of the image and the image brightness feature,respectively. On the one hand, the reflectance is a color image that only contains some isolated noise points with segmentation smoothing characteristics. Therefore, the color image denoising algorithm based on sparse representation which performs well in removing the light pollution of color images is processed. On the other hand, the shading is a grayscale image with strong sparsity which contains the main noise component. Accordingly, the non-local concentrated sparse representation gray image denoising algorithm (NCSR)which can maintain the image details is processed. At the aim of computing the proposed algorithm effectively, a new numerical solution is designed by combining orthogonal matching pursuit algorithm and soft threshold algorithm. The numerical experiments show that the new algorithm is superior to the classical color image denoising algorithm. Taking the 256×256 Boat image as an example, when the noise variance is 20, the PSNR of the new algorithm is 1.7 dB and 0.67 dB higher than the K-SVD algorithm and the NCSR algorithm, respectively.And the SSIM were improved by 0.11 and 0.09 compared with K-SVD and NCSR, respectively. The numerical experiments indicate that the proposed algorithm can preserve the image details availably while improving the denoising effect of color images. It is superior to the traditional ROF algorithm, K-SVD algorithm and NCSR algorithm in terms of visual effects and objective evaluation indicators.