Real-World Super-Resolution via Kernel Estimation and Noise Injection

Ji, Xiaozhong; Cao, Yun; Tai, Ying; Wang, Chengjie; Li, Jilin; Huang, Feiyue

doi:10.1109/cvprw50498.2020.00241

Cited by 282 publications

(278 citation statements)

References 37 publications

Supporting

Mentioning

278

Contrasting

Order By: Relevance

“…In order to ensure that the degraded images have a similar noise distribution as the source images , we extract the noise mapping patches directly from the source images in the training dataset. Due to the large variance of the patches with rick contents [38], and inspired by [40,45], when extracting noise mapping patches we control the variance within a specific range under the condition:…”

Section: Generation and Injection Of Noisementioning

confidence: 99%

“…To sum up, the process of generating LR images in ROI_LR from the source images in ROI_Src can be expressed as Equation ( 10), where i and j are randomly selected: = ( * ) ↓ + (10) [26] model. Because ESRGAN discriminator may introduce more artifacts [38], SR-D is designed on the basis of PatchGAN [44] model. The perceptual feature extractor is designed on the basis of VGG-19 [46], so as to introduce the perceptual loss [47] to enhance the visual effect of low-frequency features of the images.…”

Section: Generation and Injection Of Noisementioning

confidence: 99%

“…The proposed model DKN-SR-GAN and other compared models like EDSR8-RGB, RCAN, and RS-ESRGAN are run under Pytorch environment, using the modules provided by "sefibk/KernelGAN" project [39], "xinntao/BasicSR" project [48] and "Tencent/Real-SR" project [38] in the Github library. BiCubic can be obtained directly using Matlab functions to perform interpolation operations.…”

Section: Training Detailsmentioning

confidence: 99%

“…All the proposed models could enhance the 10m channel of Sentinel-2 to 2m. However, by using the unnatural low-high image pairs consisting of Sentinel-2 and Worldview images, and other models using the BiCubic downsampling to construct LHR image pairs [21,25,26,[34][35][36][37], the track details related to frequency will be lost [38]. In order to solve this problem, inspired by blind SR model KernelGAN [39] and the blind image denoising model [40], we explicitly estimate the degradation kernel of LHR image pairs of natural images through GAN, estimate the distribution of the degraded noises at the same time and degrade the 10m resolution images of Sentinel-2 to construct near-natural LHR image datasets.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Super-Resolution of Sentinel-2 Images at 10m Resolution without Reference Images

Y¹,

B²

2021

Preprint

View full text Add to dashboard Cite

Sentinel-2 can provide multi-spectral optical remote sensing images in RGBN bands with a spatial resolution of 10m, but the spatial details provided are not enough for many applications. WorldView can provide HR multi-spectral images less than 2m, but it is a commercial paid resource with relatively high usage costs. In this paper, without any available reference images, Sentinel-2 images at 10m resolution are improved to a resolution of 2.5m through super-resolution (SR) based on deep learning technology. Our model, named DKN-SR-GAN, uses degradation kernel estimation and noise injection to construct a dataset of near-natural low-high-resolution (LHR) image pairs, with only low-resolution (LR) images and no high-resolution (HR) prior information. DKN-SR-GAN uses the Generative Adversarial Networks (GAN) combined of ESRGAN-type generator, PatchGAN-type discriminator and the VGG-19-type feature extractor, using perceptual loss to optimize the network, so as to obtain SR images with clearer details and better perceptual effects. Experiments demonstrate that in the quantitative comparison of the non-reference image quality assessment (NR-IQA) metrics like NIQE, BRISQUE and PIQE, as well as the intuitive visual effects of the generated images, compared with state-of-the-art models such as EDSR8-RGB, RCAN and RS-ESRGAN, our proposed model has obvious advantages.

show abstract

Section: Generation and Injection Of Noisementioning

confidence: 99%

Section: Generation and Injection Of Noisementioning

confidence: 99%

Section: Training Detailsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Super-Resolution of Sentinel-2 Images at 10m Resolution without Reference Images

Y¹,

B²

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…2) For the perceptual track, the second stage was trained using the following combined loss: where λ 1 is 1e-3, λ 2 is 1 and λ 3 is 5e-3. They used a patch discriminator as in [18]. The rest of the hyperparameters are the same as in [49].…”

Section: Kirinukmentioning

confidence: 99%

AIM 2020 Challenge on Video Extreme Super-Resolution: Methods and Results

Fuoli

Huang

et al. 2020

Lecture Notes in Computer Science

View full text Add to dashboard Cite

show abstract

Multiscale generative adversarial network for real‐world super‐resolution

Sun

Yang

Tao

et al. 2021

Concurrency and Computation

View full text Add to dashboard Cite

Summary Recently, most deep convolutional neural networks used for image super‐resolution have achieved impressive performance on ideal datasets. However, these methods always fail in real‐world super‐resolution, and the results are blurred and structurally deformed. In this paper, a multiscale generative adversarial network (MGAN) is proposed to alleviate these issues. The model's multiscale loss function can effectively reduce the solution space and obtain the best features to reconstruct the image. The degraded framework based on kernel estimation and noise injection is mainly applied to obtain LR images that share the same domain with real‐world pictures. Moreover, the gradient branch is presented to provide other structural priors for SR processing. Simultaneously, to obtain better visual effects, LPIPS is used for perceptual losses instead of Visual Geometry Group (VGG). The competitive results show that our MGAN model outperforms the state‐of‐the‐art methods, resulting in lower noise and better visual quality, and reflects the superiority in image structure restoration.

show abstract

Real-World Super-Resolution via Kernel Estimation and Noise Injection

Cited by 282 publications

References 37 publications

Super-Resolution of Sentinel-2 Images at 10m Resolution without Reference Images

Super-Resolution of Sentinel-2 Images at 10m Resolution without Reference Images

AIM 2020 Challenge on Video Extreme Super-Resolution: Methods and Results

Multiscale generative adversarial network for real‐world super‐resolution

Contact Info

Product

Resources

About