高阶SVD和全变差正则的乘性噪声去除模型

Higherorder singular value decomposition-and total variation-regularized multiplicative noise removal model

光滑性、稀疏性和自相似性先验作为自然图像的重要特性被广泛应用于图像去噪.根据高阶奇异值分解和全变差正则的互补性,提出了一种能够同时利用光滑性、稀疏性和自相似性先验的乘性噪声去除新方法.该方法首先采用高阶奇异值分解方法对对数变换后图像中的相似块组进行去噪;然后结合考虑光滑性先验的全变差约束对结果进行迭代优化.实验结果表明,该方法在有效去除乘性噪声的同时,可以更好地保留图像的边缘和纹理区域的细节信息.

Smoothness, sparsity and self-similarity are the priors widely used in image denoising due to their importance in representing natural images. Motivated by the collaborative roles of higher order singular value decomposition and total variation regularization, a new approach that can simultaneously capture the above priors is proposed in this paper for removing the multiplicative noises. By taking advantages of local adaptiveness, sparsity and self-similarity realized by higher order singular value decomposition, the proposed approach starts with similar-patch-group-wise adaptive denoising on the log- transformed image, followed by the iterative optimization implemented by the total variation constraint which considers the prior of smoothness. Experiments demonstrate the advantages of the proposed approach in removing multiplicative noise and preserving the details near the edges and in the texture area.