NTIRE 2018 Challenge on Single Image Super-Resolution: Methods and Results
read more
Citations
ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks
Image Super-Resolution Using Very Deep Residual Channel Attention Networks
Image Super-Resolution Using Very Deep Residual Channel Attention Networks
ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks
The 2018 PIRM Challenge on Perceptual Image Super-Resolution
References
Deep Residual Learning for Image Recognition
Adam: A Method for Stochastic Optimization
Image quality assessment: from error visibility to structural similarity
Densely Connected Convolutional Networks
A theory for multiresolution signal decomposition: the wavelet representation
Related Papers (5)
Frequently Asked Questions (13)
Q2. How many residual blocks are used to compute the approximation coefficients of HR image?
The network uses the first 32 residual blocks and a 1 × 1 convolution layer to compute the detail coefficients of HR image and utilize another 32 residual blocks to compute the approximation coefficients of HR image.
Q3. How much PSNR does curriculum learning add?
Curriculum learning adds an average of 0.07dB PSNR on the validation set of DIV2K for 2×/4×/8× scales compared to 0.03dB using normal multiscale training.
Q4. How long does it take to generate the HR image?
On a GTX 1080Ti GPU, it takes 6.75s for rainbow, while 35s are necessary for Toyota-TI per LR image to generate the HR image, including self-ensemble for both methods.
Q5. How many residual modules are used in the proposed EUM?
Compared with the original upscaling layer in MDSR, which uses only one convolution layer without an activation function to increase the number of features, they introduce four residual modules and concatenate the outputs of the modules to increase the number of feature maps.
Q6. What is the common setting in the SR literature?
Track 1: Classic Bicubic ×8 uses the bicubic downscaling (Matlab imresize, default settings), the most common setting from the recent SR literature, with factor ×8.
Q7. What was the first challenge of its kind?
It was the first challenge of its kind with tracks employing standard bicubic degradation and ‘unknown’ operators (blur and decimation) on the 1000 DIVerse 2K resolution images from DIV2K [1] dataset.
Q8. What is the blur kernel of the SRMD?
In order to apply SRMD to tracks 2, 3 and 4, the blur kernel of which is unknown, HIT-VPC centers the blur kernels based on the largest values to align the LR image and HR image, and calculates the mean (aligned) degradation maps for each track.
Q9. What is the way to estimate the residual of a LR image?
Since each sub-band map of HR wavelet coefficients are with the same size of LR image, the proposed network (see Fig. 15) do not need deconvolution or subpixel layers.
Q10. How many residual blocks were used in each residual module?
For track 1, they used 48 residual blocks and 2 residual blocks in each residual module for feature extraction and upscaling, respectively.
Q11. What are the two main methods used in the NTIRE 2017 SR challenge?
The deep residual net (ResNet) architecture [10] and the dense net (DenseNet) architecture [11] are the basis for most of the proposed methods.
Q12. Why do the authors use the SSIM and PSNR?
Because of the pixel shifts and scalings, for Tracks 2, 3, and 4 the authors consider all the translations ∈ [−40, 40] on both axes, compute PSNR and SSIM and report the most favorable scores.
Q13. What are the objectives of the NTIRE 2018 challenge?
The objectives of the NTIRE 2018 challenge on example-based single-image super-resolution are: (i) to gauge and push the state-of-the-art in SR; (ii) to compare different solutions; and (iii) to promote realistic SR settings.