无监督多重非局部融合的图像去噪方法

非局部均值去噪 (Non-local means, NLM) 算法利用图像的自相似性, 取得了很好的去噪效果. 然而, NLM 算法对图像中不相似的邻域块分配了过大的权重, 此外算法的搜索窗大小和滤波参数等通常是固定的且无法根据图像内容的变化做出自适应的调整. 针对上述问题, 本文提出一种无监督多重非局部融合 (Unsupervised multi-non-local fusion, UM-NLF) 的图像去噪方法, 即变换搜索窗等组合参数得到多个去噪结果, 并利用 SURE (Stein's unbiased risk estimator) 对这些结果进行无监督的随机线性组合以获得最终结果. 首先, 为了滤除不相似或者相似度较低的邻域块, 本文引入一种基于可微分硬阈值函数的非局部均值 (Non-local means with a differential hard threshold function, NLM-DT) 算法, 并结合快速傅里叶变换 (Fast Fourier transformation, FFT), 初步提升算法的去噪效果和速度; 其次, 针对不同的组合参数, 利用快速 NLM-DT 算法串联生成多个去噪结果; 然后, 采用蒙特卡洛随机采样的思想对上述多个去噪结果进行随机的线性组合, 并利用基于 SURE 特征加权的移动平均滤波算法来抑制多个去噪结果组合引起的抖动噪声; 最后, 利用噪声图像和移动平均滤波后图像的 SURE 进行梯度的反向传递来优化随机线性组合的系数. 在公开数据集上的实验结果表明: UM-NLF 算法去噪结果的峰值信噪比 (Peak signal to noise ratio, PSNR) 超过了 NLM 及其大部分改进算法, 以及在部分图像上超过了 BM3D 算法. 同时, UM-NLF 相比于 BM3D 算法在视觉上产生更少的振铃伪影, 改善了图像的视觉质量.

Unsupervised Multi-non-local Fusion Image Denoising Method

Non-local means (NLM) denoising algorithm achieves a good denoising effect with the self-similarity of images. However, NLM assigns excessive weights to dissimilar neighborhood patches in the image. Meanwhile the parameters of NLM, such as the search window size, filtering coefficient and so on, are usually fixed and could not make adaptive adjustments based on changes of image content. In view of the above problems, this paper proposes an unsupervised multi-non-local fusion (UM-NLF) image denoising method, transforming the combined parameters such as the search window to obtain multiple denoising results and making an unsupervised stochastic linear combination for these results by using the Stein's unbiased risk estimator (SURE) to obtain the final result. Firstly, in order to eliminate dissimilar or low similarity neighborhood patches, this paper proposes a non-local means with a differentiable hard threshold function (NLM-DT) algorithm, and then combines fast Fourier transformation to make the preliminary improvements in the denoising effect and speed of the algorithm; Secondly, the paper uses a fast NLM-DT algorithm to generate multiple denoising results in series for different combination parameters; Then this paper combines the above multiple denoising results randomly and linearly with the Monte Carlo random sampling, and uses a weighted moving average filtering algorithm based on SURE features to suppress the jitter noise caused by the combination of multiple denoising results; Finally, the paper uses the SURE of the noise image and the filtered image to optimize linear combination unsupervisedly by the back propagation of gradients. Experiments show that the peak signal to noise ratio (PSNR) of UM-NLF algorithm exceeds those of NLM and most of the improved algorithms of NLM on public datasets and exceeds BM3D on some images. Besides, UM-NLF produces fewer ringing artifacts than BM3D algorithm visually, which improves the visual quality of the image.