Accurate Image Super-Resolution Using Very Deep Convolutional Networks
By Jiwon Kim, Jung Kwon Lee, Kyoung Mu Lee. 2015
7,077 citations recorded by the source corpus.
We present a highly accurate single-image superresolution (SR) method. Our method uses a very deep convolutional network inspired by VGG-net used for ImageNet classification [19]. We find increasing our network depth shows a significant improvement in accuracy. Our final model uses 20 weight layers. By cascading small filters many times in a deep network structure, contextual information over large image regions is exploited in an efficient way. With very deep networks, however, convergence speed becomes a critical issue during training. We propose a simple yet effective training procedure. We learn residuals only and use extremely high learning rates (104 times higher than SRCNN [6]) enabled by adjustable gradient clipping. Our proposed method performs better than existing methods in accuracy and visual improvements in our results are easily noticeable.
Subfields: Computer Vision and Pattern Recognition, Media Technology
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Ming-Hsuan Yang
- Dacheng Tao
- Chen Change Loy
- Xiaogang Wang
- Trevor Darrell
- Yi Yang
- Jitendra Malik
In Media Technology
Citations in the corpus
- Image Super-Resolution via Iterative Refinement (2021, background): “sion based and trained with an MSE loss [49], [52], [54], [55], [56].”
- Designing a Practical Degradation Model for Deep Blind Image Super-Resolution (2021, methodology): “For example, DnCNN [50] is able to handle SISR with different scale factors, JPEG compression deblocking with different quality factors and denoising for a wide range of noise levels, while still having a performance comparable to VDSR [21] for SISR.”
- The Unreasonable Effectiveness of Deep Features as a Perceptual Metric (2018, methodology): “We also evaluate several additional methods, including bicubic upsampling, and four of the top performing deep superresolution meth-ods [22, 54, 29, 45].”
- Dual Path Networks (2017, background): “Since then, different kinds of residual networks arose, concentrating on either building a more efficient micro-block inner structure [3, 21] or exploring how to use residual connections [9].”
- AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks (2020, neutral): “We compare AGD with a structural pruning (Li et al., 2016) baseline, and also include two other high-performance and lighter SR models: SRGAN (Ledig et al., 2017) and VDSR (Kim et al., 2016), to better evaluate the utility-efficiency trade-off of our compressed models.”
- Hierarchical Neural Architecture Search for Single Image Super-Resolution (2020, methodology): “In this experiment, we compare our method with several hand-crafted SR models, including the Bicubic, SRCNN [7], VDSR [27], DRCN [28], DRRN [29], SelNet [30], CARN [31].”
- Fast, Accurate and Lightweight Super-Resolution with Neural Architecture Search (2019, methodology): “In addition, it dominates DRCN [Kim et al. , 2016b] and MoreMNAS-A [Chu et al. , 2019] over three objectives on four datasets.”
- The Perception-Distortion Tradeoff (2017, methodology): “The evaluated algorithms include: A+ [42], SRCNN [9], SelfEx [13], VDSR [18], Johnson et al .”
- EnhanceNet: Single Image Super-Resolution Through Automated Texture Synthesis (2016, background): “Figure 5 gives an overview of different approaches including the current state of the art by PSNR [25, 26] on the zebra image from Set14 which is particularly well-suited for a visual comparison since it contains both smooth and sharp edges, textured regions as well as repeating patterns.”
- ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks (2018, methodology): “ image restoration problems. Several methods have been proposed to stabilize training a very deep model. For instance, residual path is developed to stabilize the training and improve the performance [18,5,12]. Residual scaling is rst employed by Szegedy et al. [21] and also used in EDSR. For genera”
- Checkmate: Breaking the Memory Wall with Optimal Tensor Rematerialization (2019, background): “Deep learning is rapidly pushing the limits of memory capacity on neural network accelerators as researchers train neural networks on high-resolution images [15, 32, 55], 3D point-clouds [10, 59], and long Natural Language Processing (NLP) sequence data [11, 14, 56].”
- PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models (2020, background): “Traditional supervised super-resolution algorithms train a model (usually, a convolutional neural network, or CNN) to minimize the pixel-wise mean-squared error (MSE) between the generated super-resolved (SR) images and the corresponding ground-truth HR images [14] [7].”
- Image Super-Resolution Using Very Deep Residual Channel Attention Networks (2018, methodology): “Manga109 (4×): YumeiroCooking HR Bicubic SRCNN [1] FSRCNN [2] VDSR [4] PSNR/SSIM 24.”
- Contextual-Based Image Inpainting: Infer, Match, and Translate (2017, background): “[21,22,23] study the problem of image super-resolution, which can be interpreted as image synthesis conditioning on low-resolution image.”
- Pixel Recursive Super Resolution (2017, methodology): “Super resolution models based on MSE regression ( e.g., [5, 16, 18]) fall within this family of pixel independent models, where the outputs of a neural network parameterize a set of Gaussians with fixed bandwidth.”
- Multi-Objective Reinforced Evolution in Mobile Neural Architecture Search (2019, methodology): “In particular, we take SRCNN (Dong et al., 2014), FSRCNN (Dong et al., 2016) and VDSR (Kim et al., 2016), and we apply our models at × 2 scale since it is one of the most commonly used tasks to draw comparisons.”