基于卷积神经网络的真实图像去噪算法  

Real Image Denoising Algorithm based on Convolutional Neural Network

作　　者：朱晏梅 林国军 ZHU Yanmei;LIN Guojun(Sichuan Key Laboratory of Artificial Intelligence,Sichuan University of Science&Engineering,Yibin 644000,China;School of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin 644000,China)

机构地区：[1]四川轻化工大学人工智能四川省重点实验室,四川宜宾644000 [2]四川轻化工大学自动化与信息工程学院,四川宜宾644000

出　　处：《成都工业学院学报》2024年第6期45-50,62,共7页Journal of Chengdu Technological University

基　　金：四川省科技厅合作项目(2022YFSY0056);教育部产学合作协同育人项目(202102581011);四川轻化工大学人才引进项目(2019RC12)。

摘　　要：为解决原有去噪算法对真实图像去噪时发生过于平滑产生伪影导致图像细节丢失的问题,提出一种基于卷积神经网络的真实图像去噪算法,由特征提取模块和干净图像生成器组成。使用特征提取模块对输入的噪声图像进行编码和特征提取;对图像特征的表达能力进行增强处理,将提取的图像特征输入到干净图像生成器进行学习图像特征后解码恢复干净图像,减少恢复干净图像时的伪影,更好地保留图像细节,从噪声图像中分离出高质量干净图像。在SIDD和DND真实噪声数据集的测试结果表明,峰值信噪比分别为35.12,36.89 dB,结构相似度分别为0.951,0.945,说明该算法能够有效消除真实图像中的噪声,去噪结果在客观评价和主观评价上均有先进性。A real image denoising algorithm based on convolutional neural network was proposed in order to solve the problem of image detail loss caused by artifacts generated by the original denoising algorithm,which consisted of a feature extraction module and a clean image generator.The feature extraction module was used to encode and extract features from the input noisy image;The expression ability of image features was enhanced,and the extracted image features were input into the clean image generator for learning image features,decoding,and then restoring the clean image.When recovering clean images,artifacts are reduced,image details can be better preserved,and high-quality clean images can be separated from noisy images.The test results on the SIDD and DND real noise datasets show that the PSNR values are 35.12,36.89 dB,and the SSIM values are 0.951,0.945,respectively.The proposed algorithm in this paper can effectively remove noise from real images,and the denoising results have advanced performance in both objective evaluation and subjective evaluation.

关 键 词：深度学习 卷积神经网络 图像去噪 真实噪声 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]