|
发布时间: 2022-12-16 |
图像处理和编码 |
|
|
收稿日期: 2021-07-26; 修回日期: 2021-12-15; 预印本日期: 2021-12-22
基金项目: 国家重点研发计划资助(2019YFE0105300);国家自然科学基金项目(61573299);中国高校产学研创新基金重点项目(2019ITA01016)
作者简介:
李潇瑶,女,博士研究生,主要研究方向为图像处理和计算机视觉。E-mail: houye0731@hnu.edu.cn
王炼红,通信作者,女,副教授,主要研究方向为信号处理、数据挖掘技术、现代网络与通信技术。E-mail: 292386791@qq.com 周怡聪,男,教授,主要研究方向为图像处理、计算机视觉、机器学习和多媒体信息安全。E-mail: yicongzhou@um.edu.mo 章兢,男,教授,主要研究方向为智能控制理论与应用、复杂系统工业控制、节能减排管控一体化和智能系统。E-mail: zhangj@hnu.edu.cn *通信作者: 王炼红 292386791@qq.com
中图法分类号: TN911.73
文献标识码: A
文章编号: 1006-8961(2022)12-3450-11
|
摘要
目的 许多彩色图像去噪算法未充分利用图像局部和非局部的相似性信息,并且忽略了真实噪声在彩色图像不同区域内分布的差异,对不同图像块和不同颜色通道都进行同等处理,导致去噪图像中同时出现过平滑和欠平滑现象。针对这些问题,本文提出一种自适应非局部3维全变分去噪算法。方法 利用一个非局部3维全变分正则项获取彩色图像块内和块间的相似性信息,同时在优化模型的保真项内嵌入一个自适应权重矩阵,该权重矩阵可以根据每次迭代得到的中间去噪结果的剩余噪声来调整算法在每个图像块、每个颜色分量以及每次迭代中的去噪强度。结果 通过不同的高斯噪声添加方式得到两个彩色噪声图像数据集。将本文算法与其他6个基于全变分的算法进行比较,采用峰值信噪比(peak signal-to-noise ratio, PSNR)和结构相似性(structural similarity, SSIM)作为客观评价指标。相比于对比算法,本文算法在两个噪声图像数据集上的平均PSNR和SSIM分别提高了0.16~1.76 dB和0.12%~6.13%,并获得了更好的图像视觉效果。结论 本文去噪算法不仅更好地兼顾了去噪与保边功能,而且提升了稳定性和鲁棒性,显示了在实际图像去噪中的应用潜力。
关键词
彩色图像去噪; 高斯噪声; 非局部相似性; 3维全变分; 自适应权重
Abstract
Objective Images are often distorted by noise during image acquisition, transmission and storage process. The generated noise can degrade image quality and affect image processing, such as edge detection, image segmentation, image recognition and image classification. Image denoising technique plays a key role in image pre-processing for image details preservation. Current Gaussian noise removal denoising techniques is often based on variational model like the total variation (TV) method. It can realize image smoothing through minimizing the corresponding energy function. However, TV-based denoising methods have their staircase effects and detail loss due to local gradient information only. Many researchers integrate the non-local concept into the total variation model after the non-local means was proposed. The existing non-local TV-based methods take advantages of the non-local similarity to denoise the image while keeping the image structure information. Unfortunately, many existing TV-based color image denoising methods fail to fully capture both local and non-local correlations among different image patches, and ignore the fact that the realistic noise varies in different image patches and different color channels. These always lead to over-smoothing and under-smoothing in the denoising result. Our newly TV-based color image denoising method, named adaptive non-local 3D total variation (ANL3DTV), is developed to deal with that. Method 1) Decompose the noisy color image into K overlapping color image patches, search for the m most similar neighboring image patches to each center image patch and then group the m image patches together. 2) Vectorize every color image patch in each image patch group and stack them into a 2D noisy matrix. 3) Obtain the corresponding 2D denoised matrices via ANL3DTV. To get the inter-patch and intra-patch correlations, our ANL3DTV takes advantages of a non-local 3D total variation regularization. On the basis of embedding an adaptive weight matrix into the fidelity term of the optimization model, it can automatically control the denoising strength on different color image patches and different color channels in each iteration. The weight matrix is correlated with the estimated noise level of each image patch. 4) Aggregate all the denoised 2D matrices to reconstruct the denoised color image. Result According to different ways to add Gaussian noise, there are two cases in the denoising experiment. In Case 1, the noisy images are corrupted with Gaussian noise with the same noise variance in all color channels. The selected noise levels are σ= 10, 30 and 50. In Case 2, we add Gaussian noise with different noise variances to each color channel. The noise levels are [σR, σG, σB] = [5, 15, 10], [40, 50, 30], [5, 40, 15] and [40, 5, 25]. ANL3DTV is compared to 6 existing TV-based denoising methods. The peak signal-to-noise ratio (PSNR) and structure similarity (SSIM) are adopted to denoising evaluation. The averaged PSNR/SSIM results of ANL3DTV in Case 1 are 32.33 dB/92.99%, 26.92 dB/81.68 and 24.57 dB/73.57%, respectively, and the quantitative results of ANL3DTV in Case 2 are 31.62 dB/92.88%, 24.49 dB/73.02%, 27.47 dB/85.94% and 26.81 dB/81.00%, respectively. Compared with other competing methods, ANL3DTV improves PSNR and SSIM by about 0.16~1.76 dB and 0.12%~6.13%. As can be seen from the denoised images, some competing methods oversmooth the images and lose many structure information. Some mistake noise pattern for the useful edge information and yield obvious ringring artifacts. Our ANL3DTV can remove more noise, preserve more details and suppress more artifacts than the competing methods. Conclusion We demonstrate an adaptive non-local 3D total variation model for Gaussian noise removal (ANL3DTV). To capture the inter-patch and intra-patch gradient information, ANL3DTV is focused on the non-local 3D total variation regularization. To adaptively adjust the denoising strength on each image patch and each color channel, an adaptive weight matrix into the fidelity term is introduced. To guarantee the feasibility of ANL3DTV mathematically, we develop the iterative solution of ANL3DTV and validate its convergence. The visual results demonstrate our ANL3DTV potentials in noise removal and detail preserving. Furthermore, ANL3DTV achieves more robustness and stablizes noise removal more under different noise levels.
Key words
color image denoising; Gaussian noise; non-local similarity; 3D total variation; adaptive weight
0 引言
图像去噪的目的是在消除噪声的同时最大程度地保留图像细节信息。图像在采集、传输和存储过程中经常受到噪声污染,导致图像质量降低,甚至影响图像增强(李健等,2021)、识别(贺敏雪等,2021)、分类(尹红等,2019)和分割(江宗康等,2020)等后续处理工作。因此,图像去噪是一项非常重要的图像预处理工作。
滤除高斯噪声,常见的空间域算法是基于变分法的去噪模型,可以通过最小化能量函数达到平滑图像的目的。Rudin等人(1992)提出一种非线性滤波算法,即全变分算法(total variation,TV),可以通过减小噪声图像的绝对梯度的积分来滤除噪声。Goldstein和Osher(2009)利用分裂Bregman快速算法提出了各向异性全变分和各项同性全变分。Adam和Paramesran(2019)为了更好地平衡消除阶梯效应和保留边缘信息之间的关系,提出一种结合重叠群稀疏和混合非凸高阶全变分模型(hybrid non-convex higher order TV with overlapping group sparsity,HNHOTV-OGS)。Parisott等人(2020)设计了一种高阶方向性全变分(total direction variation,TDV),在高阶梯度中加入各项异性方向信息。然而,由于全变分完全依赖于图像的局部特征,在边缘部分或梯度较大处容易受到噪声干扰而丢失细节信息,从而产生阶梯效应。
针对局部滤波这一缺点,Buades等人(2005)提出了非局部均值算法(non-local means,NLM),主要思想是将图像块作为一个处理单元,计算中心图像块与所有邻域图像块之间的相似性权重,最后得到的去噪图像块即邻域图像块的加权平均值。在NLM算法取得显著成效后,人们对基于图像块的非局部去噪算法进行改进。例如,将全变分与非局部相似性相结合,Gilboa和Osher(2007)提出了非局部全变分(non-local total variation,NLTV),Liu等人(2014)设计了非局部广义相对全变分(non-local version of the generalized relative TV,NLGRTV),Li等人(2017)设计了正则化非局部全变分(regularized non-local total variation,RNLTV),Wang等人(2019)提出了基于结构相似性的非局部全变分。不同于局部全变分模型仅通过图像像素灰度值的变化来除噪,这类非局部全变分算法利用了邻域内图像块结构的相似性,可以更好地保留图像的细节信息。
值得注意的是,现有许多去噪算法原本是针对灰度图像提出的,在处理彩色图像时只是简单地对3个颜色分量进行单独去噪。此外,大部分去噪算法忽略了真实噪声在图像不同区域内的分布差异,而是简单假定整幅图像的噪声强度是均匀的,并设置同样的参数,导致这些算法在每个颜色分量和每个图像块上,甚至每次迭代中的去噪强度都是一样的,从而造成过平滑和欠平滑共存现象。
为了解决以上问题,本文提出一种自适应非局部3维全变分(adaptive non-local 3D total variation,ANL3DTV)去噪模型,利用一种非局部3维全变分正则项来捕获图像块内和块间的相关性信息,同时引入一个自适应权重矩阵,根据每次迭代后去噪图像中残留噪声的特征,调整算法在不同颜色分量和不同图像块上的去噪强度。本文从数学上对ANL3DTV模型进行推导求解,分析该模型的收敛性,最后通过仿真去噪实验验证了算法的有效性。
1 原理与方法
1.1 非局部均值算法
给定噪声图像
$\boldsymbol{X}_i=\sum\limits_{j \in S_i} \omega_{i j} \boldsymbol{Y}_j $ | (1) |
$ \omega_{i j}=\frac{1}{W} \cdot \exp \left(-\frac{\left\|\boldsymbol{Y}_i-\boldsymbol{Y}_j\right\|_{\mathrm{F}}^2}{h^2}\right) $ | (2) |
式中,归一化参数
1.2 全变分模型
全变分去噪模型(Rudin等,1992)可以表示为
$\min\limits _{\boldsymbol{X}} \frac{1}{2}\|\boldsymbol{Y}-\boldsymbol{X}\|_{\mathrm{F}}^2+\lambda\|\nabla \boldsymbol{X}\|_{\ell} $ | (3) |
式中,
$ \|\nabla \boldsymbol{X}\|_1=\left\|\nabla_x \boldsymbol{X}\right\|_1+\left\|\nabla_y \boldsymbol{X}\right\|_1 $ | (4) |
$ \|\nabla \boldsymbol{X}\|_2=\sqrt{\left\|\nabla_x \boldsymbol{X}\right\|_2^2+\left\|\nabla_y \boldsymbol{X}\right\|_2^2} $ | (5) |
式中,
1.3 非局部全变分模型
已知噪声图像
$ \min\limits _{\boldsymbol{X}} \frac{1}{2}\|\boldsymbol{Y}-\boldsymbol{X}\|_{\mathrm{F}}^2+\lambda\left\|\nabla_\omega \boldsymbol{X}\right\|_{\mathrm{F}} $ | (6) |
$ \left\|\nabla_\omega \boldsymbol{X}\right\|_{\mathrm{F}}=\sum\limits_{i \in \boldsymbol{\varOmega}} \sqrt{\sum\limits_{j \in \boldsymbol{S}_i}\left(x_i-x_j\right)^2 \omega_{i j}} $ | (7) |
$ \omega_{i j}=\frac{1}{W} \cdot \exp \left(-\frac{\left\|\boldsymbol{X}_i-\boldsymbol{X}_j\right\|_{\mathrm{F}}^2}{h^2}\right) $ | (8) |
$ W=\sum\limits_{j \in \boldsymbol{S}_i} \exp \left(-\left\|\boldsymbol{X}_i-\boldsymbol{X}_j\right\|_{\mathrm{F}}^2 / h^2\right) $ | (9) |
式中,
2 本文去噪算法
2.1 去噪模型
ANL3DTV算法的主要思想是通过计算彩色图像块内的局部梯度和图像块间的非局部梯度,并利用不同图像块和颜色分量上的噪声强度差异,最终实现自适应去噪。ANL3DTV去噪算法流程如图 1所示,具体如下:
1) 将给定的噪声图像
2) 将图像块组中的每个彩色图像块拉直成长度为3
3) 对
4) 将所有去噪后的
步骤3)中,对
$ \min\limits _{\dot{\boldsymbol{X}}_i, \boldsymbol{W}_i} \frac{1}{2}\left\|\boldsymbol{W}_i \circ\left(\boldsymbol{Y}_i-\boldsymbol{X}_i\right)\right\|_{\mathrm{F}}^2+\lambda\left\|\nabla \boldsymbol{X}_i\right\|_1 $ | (10) |
式中,
$\boldsymbol{w}_{i j}=\left[w_{i j, R} \boldsymbol{I}_{p^2 \times 1} ; w_{i j, G} \boldsymbol{I}_{p^2 \times 1} ; w_{i j, B} \boldsymbol{I}_{p^2 \times 1}\right] $ | (11) |
$ w_{i j, c}=\frac{1}{\sigma_{i j, c}+\varepsilon}, c \in\{\mathrm{R}, \mathrm{G}, \mathrm{B}\} $ | (12) |
$\sigma_{i j, c}=\sqrt{\max \left(0, \sigma_c^2-\left\|\boldsymbol{y}_{i j, c}-\boldsymbol{x}_{i j, c}\right\|_2^2 / p^2\right)} $ | (13) |
式中,
2.2 模型求解
为了求解ANL3DTV去噪模型,引入辅助变量
$\min\limits _{\varTheta=\left\{\boldsymbol{X}_i, \boldsymbol{H}\right\}, \varPhi, \beta} \mathfrak{J}(\boldsymbol{\varTheta} ; \boldsymbol{\varPhi}, \boldsymbol{\beta}) $ | (14) |
式中
$\begin{gathered} \mathfrak{J}(\boldsymbol{\varTheta} ; \boldsymbol{\varPhi}, \beta)=\frac{1}{2}\left\|\boldsymbol{W}_i \circ\left(\boldsymbol{Y}_i-\boldsymbol{X}_i\right)\right\|_{\mathrm{F}}^2+\lambda\|\boldsymbol{H}\|_1+ \\ \left\langle\boldsymbol{\varPhi}, \boldsymbol{H}-\nabla \boldsymbol{X}_i\right\rangle+\frac{\beta}{2}\left\|\boldsymbol{H}-\nabla \boldsymbol{X}_i\right\|_{\mathrm{F}}^2 \end{gathered} $ | (15) |
式中,
1) 求解
关于
$\min \limits_{\boldsymbol{x}_i}\left\|\hat{\boldsymbol{W}}_i\left(\boldsymbol{y}_i-\boldsymbol{x}_i\right)\right\|_{\mathrm{F}}^2+\left\|\boldsymbol{h}^{(t)}-\boldsymbol{D} \boldsymbol{x}_i+\boldsymbol{\phi}^{(t)} / \beta^{(t)}\right\|_{\mathrm{F}}^2 $ | (16) |
$ \begin{aligned} & \boldsymbol{x}_i^{(t+1)}=\left[\hat{\boldsymbol{W}}_i^2+\beta^{(t)} \boldsymbol{D}^{\mathrm{T}} \boldsymbol{D}\right]^{-1} \cdot \\ & {\left[\hat{\boldsymbol{W}}_i^2 \boldsymbol{y}_i+\boldsymbol{D}^{\mathrm{T}}\left(\boldsymbol{\beta}^{(t)} \boldsymbol{h}^{(t)}+\boldsymbol{\phi}^{(t)}\right)\right]} \end{aligned} $ | (17) |
式中,
2) 求解
关于
$\min\limits _{\boldsymbol{H}} \frac{\lambda}{\beta^{(t)}}\|\boldsymbol{H}\|_1+\frac{1}{2} \| \boldsymbol{H}-\nabla \boldsymbol{X}_i^{(t+1)}+\boldsymbol{\varPhi}^{(t)} /\left.\beta^{(t)}\right|_{\mathrm{F}} ^2 $ | (18) |
计算为
$\boldsymbol{H}^{(t+1)}=S_{\lambda / \beta^{(t)}}\left(\beta^{(t)} \nabla \boldsymbol{X}_i^{(t+1)}-\boldsymbol{\varPhi}^{(t)}\right) $ | (19) |
式中,
3) 求解
$ \boldsymbol{\varPhi}^{(t+1)}=\boldsymbol{\varPhi}^{(t)}+\beta^{(t)}\left(\boldsymbol{H}^{(t+1)}-\nabla \boldsymbol{X}_i^{(t+1)}\right) $ | (20) |
4) 求解
$ \beta^{(t+1)}=\tau \beta^{(t)} $ | (21) |
2.3 收敛性分析
定理1 式(12)中,
$\lim\limits _{t \rightarrow \infty}\left\|\boldsymbol{H}^{(t+1)}-\nabla \boldsymbol{X}_i^{(t+1)}\right\|_{\mathrm{F}}=0 $ | (22) |
证明 首先,利用式(18)和式(19)得到
$ \begin{gathered} \left\|\boldsymbol{\varPhi}^{(t+1)}\right\|_{\mathrm{F}}=\left\|\boldsymbol{\varPhi}^{(t)}+\beta^{(t)}\left(\boldsymbol{H}^{(t+1)}-\nabla \boldsymbol{X}_i^{(t+1)}\right)\right\|_{\mathrm{F}}= \\ \beta^{(t)} \| S_{\lambda / \beta^{(t)}}\left(\beta^{(t)} \nabla \boldsymbol{X}_i^{(t+1)}-\boldsymbol{\varPhi}^{(t)}\right)- \\ \left(\beta^{(t)} \nabla \boldsymbol{X}_i^{(t+1)}-\boldsymbol{\varPhi}^{(t)}\right) \|_{\mathrm{F}} \leqslant \\ \beta^{(t)} \cdot \lambda / \beta^{(t)} \cdot \sqrt{9 m p^2}=3 p \lambda \sqrt{m} \end{gathered} $ | (23) |
所以,
因为
令
$ \begin{gathered} \mathfrak{I}\left(\boldsymbol{\varTheta}^{(t+2)} ; \boldsymbol{\varPhi}^{(t+1)}, \beta^{(t+1)}\right) \leqslant \\ \mathfrak{I}\left(\boldsymbol{\varTheta}^{(t+1)} ; \boldsymbol{\varPhi}^{(t+1)}, \beta^{(t+1)}\right)=\mathfrak{I}\left(\boldsymbol{\varTheta}^{(t+1)} ; \boldsymbol{\varPhi}^{(t)}, \beta^{(t)}\right)+ \\ \frac{\beta^{(t+1)}+\beta^{(t)}}{2\left(\beta^{(t)}\right)^2}\left\|\boldsymbol{\varPhi}^{(t+1)}-\boldsymbol{\varPhi}^{(t)}\right\|_{\mathrm{F}}^2 \leqslant\\ \mathfrak{J}\left(\boldsymbol{\varTheta}^{(t+1)} ; \boldsymbol{\varPhi}^{(t)}, \beta^{(t)}\right)+\frac{\beta^{(t+1)}+\beta^{(t)}}{2\left(\beta^{(t)}\right)^2} a^2 \leqslant \\ \Im\left(\boldsymbol{\varTheta}^{(1)} ; \boldsymbol{\varPhi}^{(0)}, \beta^{(0)}\right)+\frac{a^2}{\beta^{(0)}} \sum\limits_{t=0}^{\infty} \frac{\tau+1}{2 \tau^t}<+\infty \end{gathered} $ | (24) |
所以,得到的函数序列
然后,利用强凸函数的定义和性质(Boyd和Vandenberghe,2013)可得,存在
$\begin{gathered} \frac{a_1}{2}\left\|\boldsymbol{X}_i^{(t+1)}-\boldsymbol{X}_i^{(t)}\right\|_{\mathrm{F}}^2+\frac{a_2}{2}\left\|\boldsymbol{H}^{(t+1)}-\boldsymbol{H}^{(t)}\right\|_{\mathrm{F}}^2 \leqslant \\ \mathfrak{I}\left(\boldsymbol{\varTheta}^{(t)} ; \boldsymbol{\varPhi}^{(t)}, \beta^{(t)}\right)-\mathfrak{I}\left(\boldsymbol{\varTheta}^{(t+1)} ; \boldsymbol{\varPhi}^{(t+1)}, \beta^{(t+1)}\right)+ \\ \frac{\beta^{(t)}+\beta^{(t+1)}}{2\left(\beta^{(t)}\right)^2}\left\|\boldsymbol{\varPhi}^{(t+1)}-\boldsymbol{\varPhi}^{(t)}\right\|_{\mathrm{F}}^2 \end{gathered} $ | (25) |
由于
$\begin{array}{r} \sum\limits_{t=0}^{\infty}\left\{\frac{a_1}{2}\left\|\boldsymbol{X}_i^{(t+1)}-\boldsymbol{X}_i^{(t)}\right\|_{\mathrm{F}}^2+\right. \\ \left.\frac{a_2}{2}\left\|\boldsymbol{H}^{(t+1)}-\boldsymbol{H}^{(t)}\right\|_{\mathrm{F}}^2\right\}<+\infty \end{array} $ | (26) |
因此,序列
$ \begin{gathered} \lim _{t \rightarrow \infty}\left\|\boldsymbol{H}^{(t+1)}-\nabla \boldsymbol{X}_i^{(t+1)}\right\|_{\mathrm{F}}= \\ \lim _{t \rightarrow \infty} \frac{1}{\beta^{(t)}}\left\|\boldsymbol{\varPhi}^{(t+1)}-\boldsymbol{\varPhi}^{(t)}\right\|_{\mathrm{F}}=0 \end{gathered} $ | (27) |
3 实验结果与分析
3.1 实验设置
为了测试ANL3DTV算法的去噪效果,与TV、TDV、NLTV、RNLTV、NLGRTV和HNHOTV-OGS等算法进行对比。软件平台采用MATLAB R2016b,硬件平台采用Intel Core (TM) i7-7700U CPU处理器,16 GB内存,Windows 10操作系统。测试数据为从TID2013(tampere image database 2013)数据库(Ponomarenko等,2013)和一些常用的测试图像(例如Lena、Peppers、Baboon等)中随机选取的20幅尺寸为341 × 256像素和256 × 256像素的彩色图像,并用两种方式给这些图像添加高斯噪声。第1种噪声情况是给整幅图像添加同一强度的高斯噪声,
3.2 实验结果
表 1列出了第1种噪声情况下ANL3DTV和对比算法在测试数据集上的平均PSNR和平均SSIM结果。由表 1可知,当
表 1
第1种噪声情况下的平均PSNR和平均SSIM
Table 1
The average PSNR and average SSIM in the first noise case
算法 | ||||||||
PSNR/dB | SSIM/% | PSNR/dB | SSIM/% | PSNR/dB | SSIM/% | |||
TV | 30.71 | 90.01 | 25.65 | 77.66 | 23.8 | 70.21 | ||
TDV | 31.79 | 92.27 | 26.45 | 80.15 | 24.38 | 73.06 | ||
NLTV | 31.99 | 92.32 | 26.6 | 81.09 | 24.23 | 71.98 | ||
RNLTV | 31.72 | 92.05 | 26.18 | 79.57 | 24.06 | 73.19 | ||
NLGRTV | 31.42 | 91.66 | 25.54 | 78.81 | 23.67 | 70.6 | ||
HNHOTV-OGS | 32.2 | 92.97 | 26.65 | 81.52 | 24.47 | 73.36 | ||
ANL3DTV(本文) | 32.33 | 92.99 | 26.92 | 81.68 | 24.57 | 73.53 | ||
注:加粗字体表示各列最优结果。 |
表 2列出了第2种噪声情况下ANL3DTV和对比算法在测试数据集上的平均PSNR和平均SSIM结果。由表 2可知,当[
表 2
第2种噪声情况下的平均PSNR和平均SSIM
Table 2
The average PSNR and average SSIM in the second noise case
算法 | [ |
[ |
[ |
[ |
|||||||
PSNR/dB | SSIM/% | PSNR/dB | SSIM/% | PSNR/dB | SSIM/% | PSNR/dB | SSIM/% | ||||
TV | 29.22 | 89.26 | 23.71 | 69.71 | 26.35 | 83.46 | 25.8 | 78.31 | |||
TDV | 28.99 | 87.73 | 23.47 | 69.10 | 25.38 | 75.85 | 25.47 | 75.64 | |||
NLTV | 31.57 | 92.90 | 23.78 | 72.02 | 27.22 | 86.16 | 26.64 | 80.98 | |||
HNHOTV-OGS | 30.44 | 91.28 | 24.2 | 73.01 | 27.11 | 85.18 | 26.43 | 80.49 | |||
ANL3DTV |
31.05 | 91.52 | 24.05 | 72.93 | 25.97 | 79.12 | 25.77 | 77.24 | |||
ANL3DTV(本文) | 31.62 | 92.88 | 24.49 | 73.02 | 27.47 | 85.94 | 26.81 | 81.00 | |||
注:加粗字体表示各列最优结果。 |
图 2展示了噪声强度为
图 3和图 4分别展示了噪声强度[
为了进一步说明去噪算法保持图像纹理的能力,图 5展示了噪声强度
图 6显示了噪声强度
3.3 模型分析
ANL3DTV中需要调试的参数为正则化参数
图 8展示了噪声强度
ANL3DTV和对比算法在341×256像素图像上的平均运行时间如表 3所示。可以看出,TV所需时间最短,RNLTV和ANL3DTV所需时间相对较长。主要是因为ANL3DTV需要为每一个图像块搜索相应的非局部相似图像块,并计算3维全变分,而TV只需处理像素的局部梯度信息。
表 3
不同算法的运行时间
Table 3
The run time of different methods
算法 | 时间/s |
TV | 4.17 |
TDV | 75.18 |
NLTV | 32.21 |
RNLTV | 261.26 |
NLGRTV | 7.84 |
HNHOTV-OGS | 7.89 |
ANL3DTV(本文) | 232.55 |
注:加粗字体表示最优结果。 |
4 结论
针对现有许多彩色图像去噪算法未充分利用彩色图像的非局部和局部相似性信息,并且忽略真实噪声在不同图像块和颜色分量间的分布差异,提出了一种自适应非局部3维全变分(ANL3DTV)去噪算法。ANL3DTV利用一个非局部3维全变分正则项捕获彩色图像块内和块间的梯度信息,同时引入一个自适应权重矩阵,可以根据噪声强度的变化来调节算法在不同图像块和不同颜色分量上的去噪强度。实验从定量评价和视觉效果两个方面比较ANL3DTV与6个基于全变分的算法在不同高斯噪声强度下的去噪性能。结果表明,相较于对比算法,ANL3DTV在PSNR和SSIM指标上分别增加了0.16~1.76 dB和0.12%~6.13%。同时,ANL3DTV提升了去噪视觉效果,对噪声变化具有更强的鲁棒性。
由于ANL3DTV在搜索相似性图像块时耗费时间相对较长,因此不利于算法拓展至实时应用。另外,虽然ANL3DTV在去除高斯噪声上取得了令人满意的结果,但对于非高斯噪声的处理存在局限性。在后续工作中,针对彩色图像混合噪声的分布特点,将考虑彩色图像空间的几何结构特征,结合一些较先进的图像处理技术,对相似性测量方式和去噪模型进行优化,进一步提升算法的综合性能。
参考文献
-
Adam T, Paramesran R. 2019. Image denoising using combined higher order non-convex total variation with overlapping group sparsity. Multidimensional Systems and Signal Processing, 30(1): 503-527 [DOI:10.1007/s11045-018-0567-3]
-
Boyd S and Vandenberghe L. 2013. Convex Optimization. Wang S N, Xu Y and Hang X L, trans. Beijing: Tsinghua University Press (Boyd S, Vandenberghe L, 著. 2013. 凸优化. 王书宁, 许鋆, 黄晓霖, 译. 北京: 清华大学出版社)
-
Buades A, Coll B and Morel J M. 2005. A non-local algorithm for image denoising//Proceedings of 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. San Diego, USA: IEEE: 60-65 [DOI: 10.1109/CVPR.2005.38]
-
Gilboa G, Osher S. 2007. Nonlocal linear image regularization and supervised segmentation. Multiscale Modeling and Simulation, 6(2): 595-630 [DOI:10.1137/060669358]
-
Gilboa G, Osher S. 2009. Nonlocal operators with applications to image processing. Multiscale Modeling and Simulation, 7(3): 1005-1028 [DOI:10.1137/070698592]
-
Goldstein T, Osher S. 2009. The split bregman method for L1-regularized problems. SIAM Journal on Imaging Sciences, 2(2): 323-343 [DOI:10.1137/080725891]
-
He M X, Yu Y, Cheng R Q. 2021. Vehicle logo recognition method based on feature enhancement. Journal of Image and Graphics, 26(5): 1030-1040 (贺敏雪, 余烨, 程茹秋. 2021. 特征增强策略驱动的车标识别. 中国图象图形学报, 26(5): 1030-1040) [DOI:10.11834/jig.200327]
-
Jiang Z K, Lyu X G, Zhang J X, Zhang Q, Wei X P. 2020. Review of deep learning methods for MRI brain tumor image segmentation. Journal of Image and Graphics, 25(2): 215-228 (江宗康, 吕晓钢, 张建新, 张强, 魏小鹏. 2020. MRI脑肿瘤图像分割的深度学习方法综述. 中国图象图形学报, 25(2): 215-228) [DOI:10.11834/jig.190173]
-
Li J, Zhang X D, Li S F, Wu Z Z. 2021. Underwater color image enhancement based on two-scale image decomposition. Journal of Image and Graphics, 26(4): 787-795 (李健, 张显斗, 李熵飞, 吴子朝. 2021. 采用双尺度图像分解的水下彩色图像增强. 中国图象图形学报, 26(4): 787-795) [DOI:10.11834/jig.200144]
-
Li Z, Malgouyres F, Zeng T Y. 2017. Regularized non-local total variation and application in image restoration. Journal of Mathematical Imaging and Vision, 59(2): 296-317 [DOI:10.1007/s10851-017-0732-6]
-
Liu Q G, Xiong B, Zhang M H. 2014. Adaptive sparse norm and nonlocal total variation methods for image smoothing. Mathematical Problems in Engineering, 2014: #426125 [DOI:10.1155/2014/426125]
-
Lou Y F, Zhang X Q, Osher S, Bertozzi A. 2010. Image recovery via nonlocal operators. Journal of Scientific Computing, 42(2): 185-197 [DOI:10.1007/s10915-009-9320-2]
-
Parisotto S, Masnou S, Schönlieb C B. 2020. Higher-order total directional variation: analysis. SIAM Journal on Imaging Sciences, 13(1): 474-496 [DOI:10.1137/19M1239210]
-
Ponomarenko N, Ieremeiev O, Lukin V, Jin L, Egiazarian K, Astola J, Vozel B, Chehdi K, Carli M, Battisti F and Kuo C C J. 2013. A new color image database TID2013: Innovations and results//Proceedings of the 15th International Conference on Advanced Concepts for Intelligent Vision Systems. Poznań, Poland: Springer: 402-413 [DOI: 10.1007/978-3-319-02895-8_36]
-
Rudin L I, Osher S, Fatemi E. 1992. Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena, 60(1/4): 259-268 [DOI:10.1016/0167-2789(92)90242-F]
-
Wang W, Li F, Ng M K. 2019. Structural similarity-based nonlocal variational models for image restoration. IEEE Transactions on Image Processing, 28(9): 4260-4272 [DOI:10.1109/TIP.2019.2906491]
-
Yan J, Lu W S. 2015. Image denoising by generalized total variation regularization and least squares fidelity. Multidimensional Systems and Signal Processing, 26(1): 243-266 [DOI:10.1007/s11045-013-0255-2]
-
Yin H, Fu X, Zeng J X, Duan B, Chen Y. 2019. Flower image classification with selective convolutional descriptor aggregation. Journal of Image and Graphics, 24(5): 762-772 (尹红, 符祥, 曾接贤, 段宾, 陈英. 2019. 选择性卷积特征融合的花卉图像分类. 中国图象图形学报, 24(5): 762-772) [DOI:10.11834/jig.180426]