基于全变差模型与卷积神经网络的模糊图像恢复  

Fuzzy image restoration based on TV model and CNN

作　　者：杨琼 况姗芸[2] 冯义东[3,4] Yang Qiong;Kuang Shanyun;Feng Yidong(College of Information Technology,Qiongtai Normal University,Haikou 571127,China;School of Educational Information Technology,South China Normal University,Guangzhou 510631,China;School of Education,Hainan Normal University,Haikou 571158,China;Key Laboratory of Data Science and Intelligence Education of Ministryof Education,Hainan Normal University,Haikou 571158,China)

机构地区：[1]琼台师范学院信息科学技术学院,海南海口571127 [2]华南师范大学教育信息技术学院,广东广州510631 [3]海南师范大学教育学院,海南海口571158 [4]海南师范大学数据科学与智慧教育教育部重点实验室,海南海口571158

出　　处：《南京理工大学学报》2022年第3期277-283,共7页Journal of Nanjing University of Science and Technology

基　　金：海南省自然科学基金(60MS060,720RC614);海南省高校教育教学改革研究项目(Hnjg2019-98)。

摘　　要：为了提高模糊图像恢复性能,采用全变差(TV)正则模型进行粗粒度去模糊,运用卷积神经网络(CNN)算法进行模糊图像的像素恢复。首先,根据图像包含的噪声类型选择合适的TV模型,并针对每个像素点进行原始图像和模糊图像的TV正则最小值求解,以实现图像去模糊操作。然后,建立CNN图像恢复优化模型,将经过TV正则化后的分块图像样本作为CNN输入,结合图像信噪比(SNR)增益阈值,通过训练获得图像恢复结果。实验结果表明,采用TV正则策略及CNN的卷积优化,能够满足不同图像模糊核类别和尺寸,以及不同噪声的图像恢复需求,有效提高模糊图像的复原性能。分别采用R-L算法、反向传播神经网络(BPNN)、生成对抗网络(GAN)和TV-CNN算法对5类图像样本集进行性能仿真。通过合理设置卷积核尺寸,相比于其他模糊图像恢复算法,TV-CNN算法能够获得更优的图像恢复质量,且能够有效应对不同模糊核尺寸和不同等级噪声所带来的图像恢复难的问题。In order to improve the performance of blurred image restoration,the total variation(TV)regular model is used for coarse-grained deblurring,and the convolutional neural networks(CNN)algorithm is used for pixel restoration of blurred images.Firstly,an appropriate TV model is selected according to the noise type containing in an image,and the TV regular minimum values of the original image and the blurred image are solved for each pixel to realize the image deblurring operation.Then,the CNN image restoration optimization model is established.The partitioned image samples after TV regularization are taken as CNN input,and the image restoration results are obtained through training combined with the image signal noise ratio(SNR)gain threshold.The experimental results show that the TV regularization strategy and CNN convolution optimization can meet the image restoration requirements of different types and sizes of image blur kernels and different noise,and effectively improve the restoration performance of blurred images.R-L algorithm,back propagation neural network(BPNN),generative adversarial networks(GAN)and TV-CNN algorithm are used to simulate the performance of five kinds of image sample sets.By reasonably setting the convolution kernel size,TV-CNN algorithm can obtain better image restoration quality compared with common fuzzy image restoration algorithms,and can effectively deal with the problem of difficult image restoration caused by different fuzzy kernel sizes and different levels of noise.

关 键 词：全变差模型 卷积神经网络 模糊图像 图像恢复 模糊核 反向传播神经网络 生成对抗网络 

分 类 号：TP392.1[自动化与计算机技术—计算机应用技术]