摘要: 提出一種基于偏微分方程的圖像盲去模糊超分辨率重建算法, 旨在未知模糊核的情況下, 將含噪聲的低分辨率模糊圖像重建為清晰的高分辨率圖像. 首先, 針對(duì)圖像退
化過(guò)程構(gòu)建變分問題, 并借助變分方法推導(dǎo)出偏微分方程模型. 其次, 結(jié)合交替方向法和數(shù)值差分方法, 通過(guò)設(shè)計(jì)時(shí)空全離散數(shù)值格式求解未知的模糊核和清晰的圖像. 再次, 通過(guò)
一系列數(shù)值實(shí)驗(yàn), 分析參數(shù)選擇對(duì)圖像重建效果的影響, 確定合適的參數(shù)設(shè)置. 最后, 針對(duì)若干遙感圖像進(jìn)行實(shí)驗(yàn), 實(shí)驗(yàn)結(jié)果證明了所給模型的有效性與可靠性.
關(guān)鍵詞: 偏微分方程; 盲去噪去模糊; 超分辨率重建; 變分方法
中圖分類號(hào): O241.82" 文獻(xiàn)標(biāo)志碼: A" 文章編號(hào): 1671-5489(2025)01-0035-06
Blind Deblurring and Super-resolution Reconstruction Algorithmand Experiment Based on Partial Differential Equation
XU Wenda, WEN Xin, MAO Zhongxuan, ZOU Yongkui
(College of Mathematics, Jilin University, Changchun 130012, China)
Abstract:" We proposed a blind image deblurring and super-resolution reconstruction algorithm based on partial differential
equations (PDE). The goal was to reconstruct clear, high-resolution images
from noisy, low-resolution blurred images without prior knowledge of the blur kernel. Firstly, we constructed a variational problem for the image degrad
ation process and derived a PDE model by using variational methods. Secondly, by combining the alternating direction method and numerical difference method
, we designed a spatiotemporal fully discrete numerical scheme to solve the unknown blur kernel and the clear image. Thirdly, through a series of numeri
cal experiments, we analyzed the impact of parameter selection on image reconstruction performance and" determined appropriate parameter settings. Finally, e
xperiments were conducted on several remote sensing images, and the experimental results proved" the effectiveness and reliability of the proposed model.
Keywords: partial differential equation; blind denoising and deblurring; super-resolution reconstruction; variational method
收稿日期: 2024-11-23.
第一作者簡(jiǎn)介:" 徐文達(dá)(1985—), 男, 漢族, 碩士, 工程師, 從事偏微分方程數(shù)值解的研究, E-mail: xuwenda@jlu.edu.cn. 通信作者簡(jiǎn)介:
溫 馨(1995—), 女, 漢族, 博士, 從事偏微分方程數(shù)值解的研究, E-mail: xinwen21@jlu.edu.cn.
基金項(xiàng)目: 吉林省科技廳項(xiàng)目(批準(zhǔn)號(hào): 20240301017GX).
衛(wèi)星采集的遙感圖像通常存在退化現(xiàn)象, 即同時(shí)伴有噪聲、 模糊及低分辨率的特征. 圖像的退化過(guò)程可建模為
u0=Dk*u+n,(1)
其中u表示清晰圖像, k為模糊核, D為下采樣算子," n為高斯白噪聲,
u0為含噪聲的低分辨率模糊圖像. 圖像盲去模糊超分辨率重建任務(wù)要求在僅已知u0的
前提下, 恢復(fù)出u和k, 并確保u具有足夠高的分辨率. 顯然, 與非盲(即模糊核已知)的去
模糊超分辨率重建任務(wù)相比, 盲去模糊超分辨率重建更具挑戰(zhàn)性[1-2], 也更符合實(shí)際圖像處理的需求.
為最大限度減少噪聲帶來(lái)的偏差, 該類問題可轉(zhuǎn)化為求解如下變分問題:
argminu,k‖Dk*u-u0‖2.(2)
顯然, (u,k)=(u0,δ)是一個(gè)平凡解, 其中δ為單位脈沖模糊核. 為得到合理的解, 需引入正則化項(xiàng), 不同的正
則化項(xiàng)會(huì)對(duì)圖像的還原效果有不同影響, 因此人們?cè)O(shè)計(jì)了許多類型的正則化項(xiàng). 例如: Rudin等[3]通過(guò)引入正則化項(xiàng)L(u)=∫Ωudx建立了TV(total variation)模型, 其中Ω表示圖像區(qū)域; You等[4]通過(guò)引入正則化項(xiàng)L(u)=∫Ωf(2u
)dx建立了YK(You-Kaveh)模型, 其中函數(shù)f為非負(fù)遞增函數(shù); Lysaker等[5]在YK模型的基礎(chǔ)上進(jìn)行改進(jìn), 提出了LLT(Lysaker-Lundervold-Tai)模型; Guidotti等[6]
利用分?jǐn)?shù)階導(dǎo)數(shù)作為邊緣檢測(cè)器, 進(jìn)一步改進(jìn)了YK模型; Zhang等[7]通過(guò)加權(quán)結(jié)合TV模型和YK模型, 設(shè)計(jì)了一種自適應(yīng)模型; 張雨[8]針對(duì)被乘性
Gamma噪聲嚴(yán)重污染圖像的去散斑問題, 提出了一個(gè)全局凸的自適應(yīng)全變差模型; 溫馨等[9]進(jìn)一步改進(jìn)了文獻(xiàn)[7]中的正則化項(xiàng).
但上述工作大部分聚焦于去噪任務(wù)和非盲的去噪去模糊問題, 而聚焦于去噪去模糊同時(shí)進(jìn)行的超分辨率重建工作目前尚未見文獻(xiàn)報(bào)道. 因此, 本文采用加權(quán)結(jié)合
的TV和YK正則化項(xiàng), 構(gòu)造用于盲去噪去模糊超分辨率重建的模型, 并借助變分方法推導(dǎo)出偏微分方程組模型, 最后構(gòu)造全離散格式求解.
2.4 與經(jīng)典TV模型和YK模型的對(duì)比
注意到在本文模型中, 如果令μ=1, 則模型退化為TV模型; 如果令μ=0, 則模型退化為YK模型. 表2列出了本文模型與TV模型和YK模型的對(duì)比實(shí)驗(yàn)結(jié)果.
綜上可見, 本文提出的基于偏微分方程的盲去模糊超分辨率重建算法在未知模糊核的情況下, 實(shí)現(xiàn)了從含噪聲的低分辨率模糊圖像到清晰高分辨率圖像的重建, 數(shù)值實(shí)驗(yàn)驗(yàn)證了其有效性和
可靠性, 為遙感圖像退化問題提供了一種高效解決方案.
參考文獻(xiàn)
[1] NIMISHA T M, SINGH A K, RAJAGOPALAN A N. Blur-Invariant Deep Learning for Bli
nd-Deblurring [C]//2017 IEEE International Conference on Computer Vision (ICCV). Piscataway, NJ: IEEE, 2017: 4762-4770.
[2] LAI W S, HUANG J B, HU Z, et al. A Comparative Study for Single Image Blind Deblurring [C]//2016 IEE
E Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, NJ: IEEE, 2016: 1701-1709.
[3] RUDIN L I, OSHER S, FATEMI E. Nonlinear Total Variation Based Noise Removal Algorithms [J]. Physica D: Nonlinear Phenomena, 1992, 60(1/2/3/4): 259-268.
[4] YOU Y L, KAVEH M. Fourth-Order Partial Differential E
quations for Noise Removal [J]. IEEE Transactions on Image Processing, 2000, 9(10): 1723-1730.
[5] LYSAKER M, OSHER S, TAI X C. Noise Removal Using Smoothed Normals and Surface
Fitting [J]. IEEE Transactions on Image Processing, 2004, 13(10): 1345-1357.
[6] GUIDOTTI P, LONGO K. Two Enhanced Fourth Order Diffu
sion Models for Image Denoising [J]. Journal of Mathematical Imaging and Vision, 2011, 40(2): 188-198.
[7] ZHANG X J, YE W Z. An Adaptive Fourth-Order Partial Differential Equation for
Image Denoising [J]. Computers amp; Mathematics with Applications, 2017, 74(10): 2529-2545.
[8] 張雨. 基于偏微分方程正則化的圖像復(fù)原模型研究 [D]. 哈爾濱: 哈爾濱工業(yè)大學(xué), 2023. (
ZHANG Y. Research on Image Restoration Models Based on Partial Differential Equation Regularization [D]. Harbin: Harbin Institute of Technology, 2023.)
[9] 溫馨, 李峰, 隋鵬, 等. 一類基于偏微分方程的遙感圖像盲去噪去模糊算法 [J]. 數(shù)值計(jì)算與計(jì)算機(jī)應(yīng)用, 2024, 45(3): 288-300. (
WEN X, LI F, SUI P, et al. Blind Denoising and Deblurring Algorithm for Remote Sensing Images Based on Partial Differential Equations [J]. Journal on Nume
rical Methods and Computer Applications, 2024, 45(3): 288-300.)
[10] KAUR J. Method of Rothe in Evolution Equations [M]//VOSMANSKY'" J, ZLáMAL M. Equadiff6. Lecture Notes in
Mathematics, Vol.1192. Berlin: Springer, 1986: 23-33.
[11] SETIADI D I M. PSNR vs SSIM: Imperceptibility Quality Assessment for Image Steganography [J].
Multimedia Tools and Applications, 2021, 80: 8423-8444.
(責(zé)任編輯: 李 琦)