1.Department of Mathematics, Xinzhou Normal University, Xinzhou 034000, China
2.School of Mathematics and Statistics, Shanxi Datong University, Datong 037009, China
3.Department of Computer Science and Technology, Xinzhou Normal University, Xinzhou 034000, China
Show less
文章历史+
Received
Accepted
Published
2024-12-18
2025-04-25
Issue Date
2026-04-24
PDF (3042K)
摘要
随着现代成像技术的发展,一个成像设备瞬间可获得多幅同样内容的图像。然而,这些图像在获取、传输、存储和处理中有时难免遭受噪声污染。因此,从多幅退化图像出发,恢复一幅干净图像就成为一个现实而有意义的研究课题。本文提出一种低维流形先验正则的三重余量Wasserstein距离优化模型,并将该模型应用于图像去噪。首先,采用嵌套在高维空间中的低维流形图像先验信息,构建图像去噪模型的正则项;利用源于最优传输理论的Wasserstein距离,迫使被恢复图像的余量分布逼近参照余量分布,实现退化图像噪声估计。其次,所提模型证实,图像的低维流形正则与余量Wasserstein距离分布约束是相互补充,而非孤立无缘的;二者的巧妙结合,共同促成图像恢复性能的提升。最后,直方图匹配与加权非局部Laplacian的交替迭代优化算法,具有恢复图像效果好、实现算法效率高的特点。数值实验显示,与近年来的图像去噪方法相比,所提方法在主客观评价方面都具有优势。结果表明,本文算法比去噪性能极好的Wasserstein驱动低维流形模型(Wasserstein Driven Low-Dimensional Manifold Model, W-LDMM)和多重余量Wasserstein驱动模型(Multiple Residual Wasserstein Driven Model, MRWM)平均峰值信噪比(Peak Signal to Noise Ratio, PSNR)分别提高了1.23%和0.73%,且运算时间分别缩短了25.58%和93.21%。
Abstract
With the development of modern imaging technology, an imaging device can instantly obtain multiple images containing the same content. However, these images are sometimes difficult to avoid noise during acquisition, transmission, storage, and processing. Therefore, it is a realistic and meaningful research topic to restore a clean image from multiple degraded images. In this paper, a triple residual Wasserstein distance optimization model with low dimensional manifold prior regularization is proposed, and the model is applied to image denoising. Firstly, the regularization term of the image denoising model is constructed by utilizing the image prior information of the low-dimensional manifold nested in the high-dimensional space. By using the Wasserstein distance derived from the optimal transmission theory, the residual distribution of the restored image is forced to approximate the reference residual distribution, achieving noise estimation of degraded images. Secondly, the proposed model confirms that the regularization of low dimensional manifold in image and the constraint of residual Wasserstein distance distribution complement each other, rather than being isolated and unrelated. The clever combination of the two contributes to the improvement of image restoration performance. Finally, the alternating iterative optimization algorithm of histogram matching and weighted non local Laplacian has the characteristics of good image restoration effect and high algorithm efficiency. Numerical experiments show that compared with image denoising methods in recent years, the proposed method has advantages in both subjective and objective evaluation. The results indicate that the algorithm proposed in this paper has improved the average Peak Signal to Noise Ratio (PSNR) by 1.23% and 0.73% respectively compared to Wasserstein Driven Low-Dimensional Manifold Model (W-LDMM) and Multiple Residual Wasserstein Driven Model (MRWM), which have excellent denoising performance. Additionally, the computation time has been reduced by 25.58% and 93.21% respectively.
随着现代成像技术的不断提升,一个成像设备(如手机)瞬间就可获取多幅同样内容的图像。然而这些图像在获取、传输、存储和处理时难免被噪声污染。于是,从多帧退化图像出发恢复一幅干净图像,就成为一个有意义的研究内容。基于有限样本的多线性数据,学者提出多帧图像去噪的张量奇异值分解(Tensor-Singular Value Decomposition, t-SVD)[24],给出基于张量核范数惩罚的多维数据恢复算法。在三维块匹配算法基础上,借助图像帧间连接策略,三维块匹配的扩展版(Extension of the BM3D,E-BM3D)[25]展现了同一场景多幅图像的去噪方法。利用小波帧的属性获取图像的多尺度、多决策函数,用波的邻近特征捕捉图像的多频熵,提出基于多尺度熵的多帧图像复原方法[26]。通过图像总变差先验知识与多残差Wasserstein分布近似相结合,呈现了用于多帧图像去噪的多重余量Wasserstein驱动模型(Multiple Residual Wasserstein Drive Model, MRWM)[27]。不同于这些方法所用的图像先验信息,本文主要利用自然图像嵌套在高维空间中的低维流形结构特征[17-18],结合三重余量Wasserstein距离分布逼近作用,构建优化模型来进行多帧图像去噪。
WANGT, ZHANGK H, SHAOZ Q, et al. Gridformer: Residual Dense Transformer with Grid Structure for Image Restoration in Adverse Weather Conditions[J]. Int J Comput Vis, 2024, 132(10): 4541-4563. DOI: 10.1007/s11263-024-02056-0 .
[2]
KAURA, DONGG F. A Complete Review on Image Denoising Techniques for Medical Images[J]. Neural Process Lett, 2023, 55(6): 7807-7850. DOI: 10.1007/s11063-023-11286-1 .
[3]
HER Q, FENGX C, ZHUX L, et al. RWRM: Residual Wasserstein Regularization Model for Image Restoration[J]. Inverse Probl Imag, 2021, 15(6): 1307. DOI: 10.3934/ipi.2020069 .
[4]
ZHAOC P, FENGX C, JIAX X, et al. Root-transformation Based Multiplicative Denoising Model and Its Statistical Analysis[J]. Neurocomputing, 2018, 275: 2666-2680. DOI: 10.1016/j.neucom.2017.11.048 .
[5]
MEIY Q, FANY C, ZHANGY L, et al. Pyramid Attention Network for Image Restoration[J]. Int J Comput Vis, 2023, 131(12): 3207-3225. DOI: 10.1007/s11263-023-01843-5 .
[6]
GONGK, JOHNSONK, FAKHRI GEL, et al. PET Image Denoising Based on Denoising Diffusion Probabilistic Model[J]. Eur J Nucl Med Mol Imaging, 2024, 51(2): 358-368. DOI: 10.1007/s00259-023-06417-8 .
[7]
HER Q, FENGX C, WANGW W, et al. W-LDMM: A Wasserstein Driven Low-dimensional Manifold Model for Noisy Image Restoration[J]. Neurocomputing, 2020, 371: 108-123. DOI: 10.1016/j.neucom.2019.08.088 .
[8]
CHOIJ K, DONGB, ZHANGX Q. An Edge Driven Wavelet Frame Model for Image Restoration[J]. Appl Comput Harmon Anal, 2020, 48(3): 993-1029. DOI: 10.1016/j.acha.2018.09.007 .
[9]
SHENH, ZHAOZ Q, ZHANGW D. Adaptive Dynamic Filtering Network for Image Denoising[J]. Proc AAAI Conf Artif Intell, 2023, 37(2): 2227-2235. DOI: 10.1609/aaai.v37i2.25317 .
ZHANGK, ZUOW M, CHENY J, et al. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising[J]. IEEE Trans Image Process, 2017, 26(7): 3142-3155. DOI: 10.1109/TIP.2017.2662206 .
[12]
ZHANGK, ZUOW, ZHANGL. FFDNet: Toward a Fast and Flexible Solution for CNN-based Image Denoising[J]. IEEE Trans Image Process, 2018, 27(9): 4608-4622. DOI: 10.1109/TIP.2018.2839891 .
[13]
DABOVK, FOI A, KATKOVNIKV, et al. Image Denoising by Sparse 3-D Transform-domain Collaborative Filtering[J]. IEEE Trans Image Process, 2007, 16(8): 2080-2095. DOI: 10.1109/tip.2007.901238 .
[14]
DONGW, ZHANGL, SHIG, et al. Nonlocally Centralized Sparse Representation for Image Restoration[J]. IEEE Trans Image Process, 2013, 22(4): 1620-1630. DOI: 10.1109/tip.2012.2235847 .
[15]
ZUOW M, ZHANGL, SONGC W, et al. Gradient Histogram Estimation and Preservation for Texture Enhanced Image Denoising[J]. IEEE Trans Image Process, 2014, 23(6): 2459-2472. DOI: 10.1109/TIP.2014.2316423 .
[16]
GUS H, XIEQ, MENGD Y, et al. Weighted Nuclear Norm Minimization and Its Applications to Low Level Vision[J]. Int J Comput Vis, 2017, 121(2): 183-208. DOI: 10.1007/s11263-016-0930-5 .
[17]
CARLSSONG, ISHKHANOVT, DE SILVAV, et al. On the Local Behavior of Spaces of Natural Images[J]. Int J Comput Vis, 2008, 76(1): 1-12. DOI: 10.1007/s11263-007-0056-x .
[18]
PEYREG. A Review of Adaptive Image Representations[J]. IEEE J Sel Top Signal Process, 2011, 5(5): 896-911. DOI: 10.1109/JSTSP.2011.2120592 .
[19]
OSHERS, SHIZ, ZHUW. Low Dimensional Manifold Model for Image Processing[J]. SIAM Imaging Sciences, 2017, 10(4): 1669-1690. DOI: 10.1137/16M1058686 .
[20]
SHIZ Q, OSHERS, ZHUW. Generalization of the Weighted Nonlocal Laplacian in Low Dimensional Manifold Model[J]. J Sci Comput, 2018, 75(2): 638-656. DOI: 10.1007/s10915-017-0549-x .
[21]
LIZ, SHIZ Q, SUNJ. Point Integral Method for Solving Poisson-type Equations on Manifolds from Point Clouds with Convergence Guarantees[J]. Commun Comput Phys, 2017, 22(1): 228-258. DOI: 10.4208/cicp.111015.250716a .
[22]
SHIZ Q, SUNJ, TIANM H. Harmonic Extension on the Point Cloud[J]. Multiscale Model Simul, 2018, 16(1): 215-247. DOI: 10.1137/16m1098747 .
[23]
SHIZ Q, OSHERS, ZHUW. Weighted Nonlocal Laplacian on Interpolation from Sparse Data[J]. J Sci Comput, 2017, 73(2): 1164-1177. DOI: 10.1007/s10915-017-0421-z .
[24]
ZHANGZ M, ELY G, AERONS, et al. Novel Methods for Multilinear Data Completion and Denoising Based on Tensor-SVD[C]//2014 IEEE Conference on Computer Vision and Pattern Recognition. Columbus: IEEE. 2014: 3842-3849. DOI:10.1109/CVPR.2014.485 .
[25]
BODDUNAK, WEICKERTJ. Enhancing Patch-Based Methods with Inter-Frame Connectivity for Denoising Multi-Frame Images[C]//2019 IEEE International Conference on Image Processing. Taipei: IEEE. 2019: 2414-2418. DOI: 10.1109/ICIP.2019. 8803344 .
[26]
DHABLIYAD, KUPPURAJT, SHARMAN, alet, Multi-Scale Entropy Based Image Denoising for Multi-frame Image Restoration[C]//2024 International Conference on Computing Communication and Networking Technologies. Kamand: IEEE. 2024:1-6. DOI: 10.1109/ICCCNT61001.2024.10725726 .
[27]
HER Q, LANW S, LIUF. MRWM: A Multiple Residual Wasserstein Driven Model for Image Denoising[J]. IEEE Access, 2022, 10: 127397-127411. DOI: 10.1109/ACCESS.2022.3226331 .
[28]
RABINJ, PEYRÉG. Wasserstein Regularization of Imaging Problem[C]//2011 IEEE International Conference on Image Processing. Brussels: IEEE. 2011: 1541-1544. DOI: 10.1109/ICIP.2011.6115740 .
[29]
PEYRÉG, FADILIJ, RABINJ. Wasserstein Active Contours[C]//2012 19th IEEE International Conference on Image Processing. New York: IEEE, 2012: 2541-2544. DOI: 10.1109/ICIP.2012.6467416 .
[30]
PELEO, WERMANM. Fast and Robust Earth Mover's Distances[C]//2009 IEEE 12th International Conference on Computer Vision. New York: IEEE, 2009: 460-467. DOI: 10.1109/ICCV.2009.5459199 .
[31]
VILLANIC. Optimal Transport: Old and New[M]. New York: Springer Science and Business Media, 2008.
[32]
GHOSHA, SUBUDHIB N, GHOSHS. Object Detection from Videos Captured by Moving Camera by Fuzzy Edge Incorporated Markov Random Field and Local Histogram Matching[J]. IEEE Trans Circuits Syst Video Technol, 2012, 22(8): 1127-1135. DOI: 10.1109/TCSVT.2012.2190476 .
[33]
THURMANS T, FIENUPJ R. Noise Histogram Regularization for Iterative Image Reconstruction Algorithms[J]. J Opt Soc Am A Opt Image Sci Vis, 2007, 24(3): 608-617. DOI: 10.1364/josaa.24.000608 .
[34]
FANJ Y, CHENT, CAOJ W. Image Tampering Detection Using Noise Histogram Features[C]//2015 IEEE International Conference on Digital Signal Processing (DSP). New York: IEEE, 2015: 1044-1048. DOI: 10.1109/ICDSP.2015.7252037 .
[35]
SMITHS M, BRADYJ M. SUSAN: A New Approach to Low Level Image Processing[J]. Int J Comput Vis, 1997, 23(1): 45-78. DOI: 10.1023/A:1007963824710 .