联合内容和质量约束的真实图像去噪  

Real Image Denoising with Joint Content and Quality Constraints

作　　者：朱皓晨 赵墨 曹刚 ZHU Haochen;ZHAO Mo;CAO Gang(School of Computer and Cyber Sciences,Communication University of China,Beijing 100024,China;State Key Laboratory of Media Convergence and Communication,Communication University of China,Beijing 100024,China)

机构地区：[1]中国传媒大学计算机与网络空间安全学院,北京100024 [2]中国传媒大学媒体融合与传播国家重点实验室,北京100024

出　　处：《信号处理》2024年第6期1141-1147,共7页Journal of Signal Processing

基　　金：国家自然科学基金(62071434);中央高校基本科研业务费专项资金(CUC24GT01)。

摘　　要：真实图像去噪指从真实的数字照片图像中移除噪声失真,以提升图像的视觉质量。当前性能最优的真实图像去噪模型普遍依赖于逐像素对应的干净/噪声图像对样本集,但其采集困难。针对该挑战,本文提出了一种联合内容和质量约束的真实图像去噪网络CQDeNet,仅利用图像级标注样本完成训练并实现有效去噪。CQDeNet主要包含图像内容约束和图像质量约束两个前后联结的子网络。内容约束网络采用单噪声图像训练骨干网络TECDNet,通过噪声图像中的潜在干净图像信号还原内容信息。质量约束网络利用无参考图像质量评估来约束生成图像的质量,协作引导内容约束网络的优化方向。通过联合这两个子网络,CQDeNet解除了当前无监督方法对输入图像噪声残差零均值假设的限制,因此具有更强的泛化能力和扩展性。测试结果表明,所提方法能有效去除真实图像中噪声,在SSID和DND数据集上的平均PSNR值分别达到34.83 dB和37.21 dB。Real image denoising aims to remove noise distortion from noisy digital photograph images to improve their visual quality.The state-of-the-art real image denoising models generally rely on sample set of pixel-wise corresponding clean-noisy image pairs,which are difficult to collect.To address this challenge,we propose a content-and quality-constrained real image denoising network(CQDeNet),a network for real image denoising constrained by content and quality,trained solely on image-level labeled samples to achieve effective denoising.CQDeNet consists of two sequential sub-networks:the image content-constrained network and the image quality-constrained network.The content-constrained network is trained in a self-supervised manner using only noisy images to restore the content information by extracting the latent clean image signal from the noisy images.The quality-constrained network utilizes a no-reference image quality assessment to constrain the quality of the generated images,collaboratively guiding the optimization direction of the content-constrained network.By combining these two sub-networks,CQDeNet overcomes the limitation of the zero-mean assumption of noise residuals in current unsupervised methods,enabling the model to have stronger gen‐eralization ability and scalability.Test results show that the proposed method could effectively remove the noise from real images.The average PSNR values obtained on the SSID and DND datasets are 34.83 dB and 37.21 dB,respectively.

关 键 词：真实图像去噪 深度学习 图像内容约束网络 图像质量约束网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]