百科.dev
全部条目在线工具AI 编程趋势榜开源项目技术资讯提交条目
登录
< 返回工具列表
C

CARN-pytorch

> 编程语言
开源

使用级联残差网络实现快速、准确且轻量级的超分辨率 (ECCV 2018)

388 stars0 点赞1 次浏览
访问官网GitHub

工具介绍

使用级联残差网络实现快速、准确且轻量级的超分辨率 (ECCV 2018)

Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network

Namhyuk Ahn, Byungkon Kang, Kyung-Ah Sohn.

European Conference on Computer Vision (ECCV), 2018. [arXiv]

Abstract

In recent years, deep learning methods have been successfully applied to single-image super-resolution tasks. Despite their great performances, deep learning methods cannot be easily applied to real-world applications due to the requirement of heavy computation. In this paper, we address this issue by proposing an accurate and lightweight deep learning model for image super-resolution. In detail, we design an architecture that implements a cascading mechanism upon a residual network. We also present a variant model of the proposed cascading residual network to further improve efficiency. Our extensive experiments show that even with much fewer parameters and operations, our models achieve performance comparable to that of state-of-the-art methods.

FAQs

  1. Can't reproduce PSNR/SSIM as recorded in the paper: See issue#6

Requirements

  • Python 3
  • PyTorch (0.4.0), torchvision
  • Numpy, Scipy
  • Pillow, Scikit-image
  • h5py
  • importlib

Dataset

We use DIV2K dataset for training and Set5, Set14, B100 and Urban100 dataset for the benchmark test. Here are the following steps to prepare datasets.

  1. Download DIV2K and unzip on dataset directory as below:
dataset
└── DIV2K
    ├── DIV2K_train_HR
    ├── DIV2K_train_LR_bicubic
    ├── DIV2K_valid_HR
    └── DIV2K_valid_LR_bicubic
  1. To accelerate training, we first convert training images to h5 format as follow (h5py module has to be installed).
bash
$ cd datasets && python div2h5.py
  1. Other benchmark datasets can be downloaded in Google Drive. Same as DIV2K, please put all the datasets in dataset directory.

Test Pretrained Models

We provide the pretrained models in checkpoint directory. To test CARN on benchmark dataset:

bash
$ python carn/sample.py --model carn \
                        --test_data_dir dataset/ \
                        --scale [2|3|4] \
                        --ckpt_path ./checkpoint/.pth \
                        --sample_dir 

and for CARN-M,

bash
$ python carn/sample.py --model carn_m \
                        --test_data_dir dataset/ \
                        --scale [2|3|4] \
                        --ckpt_path ./checkpoint/.pth \
                        --sample_dir  \
                        --group 4

We provide our results on four benchmark dataset (Set5, Set14, B100 and Urban100). Google Drive

Training Models

Here are our settings to train CARN and CARN-M. Note: We use two GPU to utilize large batch size, but if OOM error arise, please reduce batch size.

…

In the --scale argument, [2, 3, 4] is for single-scale training and 0 for multi-scale learning. --group represents group size of group convolution. The differences from previous version are: 1) we increase batch size and patch size to 64 and 64. 2) Instead of using reduce_upsample argument which replace 3x3 conv of the upsample block to 1x1, we use group convolution as same way to the efficient residual block.

Results

Note: As pointed out in #2, previous Urban100 benchmark dataset was incorrect. The issue is related to the mismatch of the HR image resolution from the original dataset in x2 and x3 scale. We correct this problem, and provided dataset and results are fixed ones.

Citation

@article{ahn2018fast,
  title={Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network},
  author={Ahn, Namhyuk and Kang, Byungkon and Sohn, Kyung-Ah},
  journal={arXiv preprint arXiv:1803.08664},
  year={2018}
}

Issues· 0 开放

查看全部 Issues在 GitHub 打开

暂无开放 Issues,或尚未同步最近议题。

> 标签

Python

暂无评论,来聊聊你的看法吧

> 工具信息

发布日期2026年8月1日
最后更新2026年9月17日
分类编程语言
定价开源

> 相关工具

T
TypeScript
JavaScript 的超集,为前端与全栈提供静态类型
P
Python
通用编程语言,广泛用于 Web、数据与 AI
G
Go
Google 推出的简洁高效系统语言