Paper (NeurIPS'21) - One-to-many Approach for Improving Super-Resolution
https://arxiv.org/pdf/2106.10437.pdf
Code using Tensorflow - https://github.com/krenerd/ultimate-sr
Paper (Mathematical Problems in Engineering, 2020) - GAN-Based Image Super-Resolution with a Novel Quality Loss
https://www.hindawi.com/journals/mpe/2020/5217429/
A Tensorflow 2.x based implementation of EDSR, WDSR and SRGAN - https://github.com/krasserm/super-resolution
Super-resolution related papers, data and repositories - https://github.com/ChaofWang/Awesome-Super-Resolution#DL-based-approach
Paper (CVPR'16) - Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network
Python Code using PyTorch - https://github.com/pytorch/examples/tree/master/super_resolution
Paper (ECCV'20) - Lightweight Single-Image Super-Resolution Network with Attentive Auxiliary Feature Learning
Th...
Paper (CVPR'17) - Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
The discriminator network is learned to classify between the natural images and the images created by the generator network.
The generator network tries to pool the discriminator network by creating a high resolution image as close to the natural image as possible.
Using a perceptual loss function, not MSE. This makes the resulting high resolution image more perceptually plausible.
Python Code using PyTorch
- https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Super-Resolution
Paper (TPAMI) - Image Super-Resolution using Deep Convolutional Network
Conventional super resolution methods consists of three steps:
Patch extraction and representation for a low resolution input image
Non-linear mapping between patches of low and high resolution images
Reconstruction of a high resolution image by combining (usually averaging) the high resolution patches which is obtained from the second step
All these steps can be replaced by linear convolution operations, i.e., a convolutional neural network (CNN).
Python code using PyTorch - https://github.com/YeongHyeon/Super-Resolution_CNN-PyTorch
Paper (IEEE Access) - Super Resolution Reconstruction of Images Based on Interpolation and Full Convolutional Neural Network and Application in Medical Fields
How it works:
Input image is fed into the first CNN and an initial high resolution image is created. Also another high resolution is created by one of conventional interpolation methods, such as nearest neighbor and bicubic interpolation.
Two high resolution images are fed into the second CNN to make the final high resolution image.
The CNNs do not have fooling layers.
https://ieeexplore.ieee.org/document/8939373
Paper (CVPR'08) - Image Super-Resolution as Sparse Representation of Raw Image Patches
Based on example-based learning
A representative SR method before CNN-based methods emerge
"The sparse representation of a signal can be correctly recovered from its down-sampled signal."
"Linear relationships among high-resolution signals can be precisely recovered from their low-dimensional projections."
Key idea - using the sparsity as a prior for regularization to solve the ill-posed SR problem.