Wide Activation for Efficient and Accurate Image Super-Resolution

Abstract

In this report we demonstrate that with same parameters and computational budgets, models with wider features before ReLU activation have significantly better performance for single image super-resolution (SISR). The resulted SR residual network has a slim identity mapping pathway with wider (\(2\times\) to \(4\times\)) channels before activation in each residual block. To further widen activation (\(6\times\) to \(9\times\)) without computational overhead, we introduce linear low-rank convolution into SR networks and achieve even better accuracy-efficiency tradeoffs. In addition, compared with batch normalization or no normalization, we find training with weight normalization leads to better accuracy for deep super-resolution networks. Our proposed SR network \textit{WDSR} achieves better results on large-scale DIV2K image super-resolution benchmark in terms of PSNR with same or lower computational complexity. Based on WDSR, our method also won 1st places in NTIRE 2018 Challenge on Single Image Super-Resolution in all three realistic tracks. Experiments and ablation studies support the importance of wide activation for image super-resolution. Code is released at: https://github.com/JiahuiYu/wdsr_ntire2018

arXiv commenttech report and factsheet

References in corpus (3)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
• Flattened Convolutional Neural Networks for Feedforward Acceleration

Cited by in corpus (47)

• PhyCRNet: Physics-informed Convolutional-Recurrent Network for Solving Spatiotemporal PDEs
• Super Resolution Convolutional Neural Network Models for Enhancing Resolution of Rock Micro-CT Images
• Unfolding the Alternating Optimization for Blind Super Resolution
• AutoSlim: Towards One-Shot Architecture Search for Channel Numbers
• Boosting Resolution and Recovering Texture of micro-CT Images with Deep Learning
• Lightweight Image Super-Resolution with Adaptive Weighted Learning Network
• Learning Detail-Structure Alternative Optimization for Blind Super-Resolution
• UltraSR: Spatial Encoding is a Missing Key for Implicit Image Function-based Arbitrary-Scale Super-Resolution
• Bayesian Image Super-Resolution with Deep Modeling of Image Statistics
• Iterative Network for Image Super-Resolution
• Generative Adversarial Networks for Image and Video Synthesis: Algorithms and Applications
• Multi-modal Deep Guided Filtering for Comprehensible Medical Image Processing
• HINet: Half Instance Normalization Network for Image Restoration
• Kernel Based Progressive Distillation for Adder Neural Networks
• Bridging Component Learning with Degradation Modelling for Blind Image Super-Resolution
• Spatial-Spectral Residual Network for Hyperspectral Image Super-Resolution
• AIM 2019 Challenge on Constrained Super-Resolution: Methods and Results
• Universally Slimmable Networks and Improved Training Techniques
• Lightweight Single-Image Super-Resolution Network with Attentive Auxiliary Feature Learning
• CoCoPIE: Making Mobile AI Sweet As PIE --Compression-Compilation Co-Design Goes a Long Way
• Rank-One Network: An Effective Framework for Image Restoration
• NTIRE 2020 Challenge on Perceptual Extreme Super-Resolution: Methods and Results
• 3D Detection and Characterisation of ALMA Sources through Deep Learning
• Multi-scale deep neural networks for real image super-resolution
• Deep network series for large-scale high-dynamic range imaging
• AIM 2020 Challenge on Rendering Realistic Bokeh
• AIM 2020 Challenge on Video Extreme Super-Resolution: Methods and Results
• Wider Channel Attention Network for Remote Sensing Image Super-resolution
• Achieving Real-Time LiDAR 3D Object Detection on a Mobile Device
• Boosting the Performance of Video Compression Artifact Reduction with Reference Frame Proposals and Frequency Domain Information
• NTIRE 2021 Challenge on Quality Enhancement of Compressed Video: Methods and Results
• Achieving on-Mobile Real-Time Super-Resolution with Neural Architecture and Pruning Search
• Pedestrian Detection in Low-Light Conditions: A Comprehensive Survey
• Adaptive Loss Function for Super Resolution Neural Networks Using Convex Optimization Techniques
• Localized Super Resolution for Foreground Images using U-Net and MR-CNN
• VSpSR: Explorable Super-Resolution via Variational Sparse Representation
• FC${}^2$N: Fully Channel-Concatenated Network for Single Image Super-Resolution
• Deep Adaptive Inference Networks for Single Image Super-Resolution
• A Lightweight Solution of Industrial Computed Tomography with Convolutional Neural Network
• Towards Real-Time DNN Inference on Mobile Platforms with Model Pruning and Compiler Optimization
• A Learning-from-noise Dilated Wide Activation Network for denoising Arterial Spin Labeling (ASL) Perfusion Images
• Weak Texture Information Map Guided Image Super-resolution with Deep Residual Networks
• NTIRE 2020 Challenge on Video Quality Mapping: Methods and Results
• ERNet Family: Hardware-Oriented CNN Models for Computational Imaging Using Block-Based Inference
• Adaptive Densely Connected Super-Resolution Reconstruction
• Back-Projection Pipeline
• Image restoration quality assessment based on regional differential information entropy