王曄 孫志寬 李征



現有的圖像超分辨率重建方法都較少考慮真實低分辨率圖像中包含的噪聲信息,因此會影響圖像的重建質量.受真實圖像去噪算法的啟發,本文引入一個噪聲分布收集網絡來收集低分辨率圖像的噪聲分布信息,并采用生成對抗網絡的模型設計,提高含噪聲圖像的重建質量.噪聲分布信息會分別輸入到超分辨率重建網絡和判別網絡,在重建過程中去除噪聲的同時保證有用高頻信息的恢復,另外由于判別網絡的能力對整個模型的性能有著重要影響,選擇使用 U-Net 網絡來獲得更好的梯度信息反饋.與經典圖像超分辨率重建算法的對比以及消融實驗表明,使用噪聲收集網絡和 U-Net 判別網絡后,本文模型在噪聲低分辨率圖像重建任務中獲得了更好的性能.
圖像超分辨率; 生成對抗網絡; 真實圖像; 噪聲分布
TP391A2023.032001
收稿日期: 2022-06-20
基金項目: 國家重點研發計劃項目(2020YFA0714003); 國家重大項目(GJXM92579); 四川省科技廳重點研發項目(2021YFQ0059)
作者簡介: 王曄(1997-), 女, 山西晉城人, 碩士研究生, 主要研究領域為計算機視覺與圖像處理. E-mail: yesica1230@163.com
通訊作者: 李征. E-mail: lizheng@scu.edu.cn
An image super-resolution reconstruction method based onGenerative Adversarial Network and noise distribution
WANG Ye1, SUN Zhi-Kuan1, LI Zheng1, 2
(1.College of Computer Science(College of Software), Sichuan University, Chengdu 610065, China;
2.Tianfu Engineering-Oriented Numerical Simulation & Software Innovation Center, Sichuan University, Chengdu 610207, China)
Existing image super-resolution reconstruction methods take less into account the noise information contained in real low-resolution images, which will affect the quality of image reconstruction. Inspired by the real image denoising algorithm, this paper introduces a noise distribution collection network to collect noise distribution information of low-resolution images, and adopts a model design of Generative Adversarial Network to improve the reconstruction quality of noisy images. The noise distribution information will be input to the super-resolution reconstruction network and the discriminant network respectively. During the reconstruction process, the noise is removed during while ensuring the recovery of useful high-frequency information, because the ability of the discriminant network has an important impact on the performance of the entire model, the U-Net network is selected to obtain better gradient information feedback. Comparison with the classical image super-resolution reconstruction methods and ablation experiments,the resluts show that the proposed model obtains better performance in the noisy low-resolution image reconstruction task after using the noise collection network and the U-Net discriminant network.
Image super-resolution; Generative Adversarial Network; Real images; Noise distribution
1 引 言
隨著科技的發展,作為重要信息載體的圖像擁有了越來越高的分辨率.更高的像素密度展示了更多的紋理細節,蘊涵更豐富的信息,在很多場景中高分辨率圖像都極具價值,比如CT圖像、核磁共振圖像、監控視頻和衛星遙感圖像等.但高分辨率的圖像對成像設備和環境因素要求很高.在實際生活中,人們往往……