面向图像去噪的深度双层群稀疏编码网络  

Deep Two-Layer Group Sparse Coding Network for Image Denoising

作　　者：方祯煜 尹海涛 FANG Zhenyu;YIN Haitao(College of Automation and College of Artificial Intelligence,Nanjing University of Posts and Telecommunications,Nanjing 210023,China)

机构地区：[1]南京邮电大学自动化学院、人工智能学院,南京210023

出　　处：《计算机工程与应用》2023年第21期195-203,共9页Computer Engineering and Applications

基　　金：国家自然科学基金(61971237)。

摘　　要：当前基于深度学习的图像去噪方法主要是利用深度神经网络将噪声图像直接映射到干净图像,忽略了图像去噪任务知识。针对该问题,提出了一种基于双层群稀疏编码的深度图像去噪网络。为了充分利用图像中相似结构以及对相似块之间的特异性有效表示,提出了双层群稀疏编码图像去噪模型,并表示成l_(2,1)-l_(1)范数优化问题。利用算法展开技术将所提去噪模型的优化解转化成“端到端”的深度神经网络。为了进一步提高网络训练的稳定性,所提去噪网络中引入一种改进的残差连接。在BSD68、Set12、CBSD68、Kodak24和Urban100等常用数据集上的实验结果表明,所提算法在主观视觉质量和客观评价指标上优于一些主流的去噪方法。特别地,针对噪声等级为75,所提算法在CBSD68数据集上比经典的DnCNN算法平均PSNR指标提高了1.3 dB。The current deep learning based image denoising methods directly map the noisy image to its clean one using by deep neural network,which ignore the domain knowledge of image denoising task.According to this problem,this paper proposes a two-layer group sparse coding based image denoising network.Firstly,to effectively exploit the similarity structure of image and represent the specificity among similarpatterns,a two-layer group sparse coding is developed for image denoising,modeled as l_(2,1)-l_(1)-norm optimization problem.Then,the optimization solution of this denoising model is converted into an end-to-end deep neural network through algorithm unrolling.Moreover,a modified residual connection is applied to further improve the stability of network training.The experimental results on BSD68,Set12,CBSD68,Kodak24 and Urban100 datasets demonstrate that the proposed method outperforms some popular denoising methods.Specifically,the proposed method achieves 1.3 dB average PSNR gain over the classical DnCNN on CBSD68 at noise level 75.

关 键 词：图像去噪 群稀疏编码 深度神经网络 算法展开 残差连接 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]