基于多尺度编解码器网络的运动图像盲去模糊  

Blind deblurring of motion image based on multi-scale encoder-decoder network

作　　者：齐永锋 段友放 火元莲[2] 侯璐璐 QI Yong-feng;DUAN You-fang;HUO Yuan-lian;HOU Lu-lu(College of Computer Science and Engineering,Northwest Normal University,Lanzhou,Gansu 730070,China;College of Physics and Electronic Engineering,Northwest Normal University,Lanzhou,Gansu 730070,China)

机构地区：[1]西北师范大学计算机科学与工程学院,甘肃兰州730070 [2]西北师范大学物理与电子工程学院,甘肃兰州730070

出　　处：《光电子．激光》2021年第2期157-165,共9页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(61561044)资助项目。

摘　　要：为避免因模糊核估计不准确而使得复原后的图像易造成欠去模糊或过度去模糊等问题,提出了一种基于多尺度编解码器网络去模糊模型。首先,在传统的编解码器网络中加入跳跃连接和多尺度循环连接,结合各层图像特征信息的同时使网络训练得更加稳定。其次,将提出的编解码器网络与改进的嵌套残差网络结合,采用由"粗"到"精"的方法进一步提取不同尺度的图像信息,以增强图像的纹理细节。最后,采用端到端的模式,将模糊图像直接进行盲去模糊操作,最终输出清晰图像。改进后的多尺度网络模型具有更少的参数,网络结构更加简单,不需要模糊核估计环节。在GoPro和K?hler数据集上的实验结果表明,本文方法相较于传统的SFD方法,PSNR分别提升约6.61 dB和2.13 dB,相较于MRF CNN、Multi-scale CNN以及DeblurGAN等基于CNN的方法,PSNR分别最高提升约5.61 dB和1.59 dB,实验结果均优于另外4种方法,复原后的图像更接近真实图像,纹理细节更加清晰,是一种有效的运动图像去模糊方法。In order to avoid the problem that the reconstructed image is easy to cause under-deblurred or over-deblurred due to inaccurate blur kernels estimation,a multi-scale encoder-decoder network deblurring model is proposed.Firstly,the skip connection and multi-scale recurrent connection are added to the traditional encoder-decoder network,and combine the image feature of each layer to make the network training more stable.Then,the proposed encoder-decoder is combined with the nested residual network.The image information of different scales is further extracted by the method of"coarse-to-fine"to enhance the texture details of the image.Finally,the image is directly blind deblurred by using the end-to-end mode to output a clear image.The multi-scale network model has fewer parameters,the network structure is simpler,and does not need blur kernels estimation.The experimental results on the GoPro and K?hler datasets show that the PSNR of the proposed method is improved by about 6.61 dB and 2.13 dB compared with the traditional SFD method.Compared with MRF CNN、Multi-scale CNN、DeblurGAN and other methods based on CNN,the maximum increase in PSNR is about 5.61 dB and 1.59 dB,respectively.The experimental results are better than the other four methods.The restored image is closer to the real image,and the texture details are clearer.It is an effective motion image deblurring method.

关 键 词：盲去模糊 卷积神经网络 编解码器 多尺度模型 残差网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]