Single Image Super-Resolution via a Dual Interactive Implicit Neural Network

Overview

Single image super-resolution (SISR) is a fundamental low-level computer vision problem that aims to recover a high-resolution image from its low-resolution counterpart. There are two main reasons for performing SISR: (i) to enhance the visual quality of an image for human consumption, and (ii) to improve the representation of an image for machine perception. SISR has many practical applications including robotics, remote sensing, satellite imaging, thermal imaging, medical imaging, and much more.

This repository provides source code for our 2023 WACV paper titled "Single Image Super-Resolution via a Dual Interactive Implicit Neural Network." Our dual interactive implicit neural network (DIINN) is capable of producing images of arbitrary resolution, using a single trained model, by capturing the underlying implicit representation of the input image. DIINN achieves state-of-the-art SISR results on extensive experimental evaluations across many settings and datasets.

Citation

If you find this project useful, then please consider citing our work.

@inproceedings{nguyen2023single,
  title={Single Image Super-Resolution via a Dual Interactive Implicit Neural Network},
  author={Nguyen, Quan H. and Beksi, William J},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages={4936--4945},
  year={2023}
}

Model Architecture

Installation

First, clone the project

$ git clone https://github.com/robotic-vision-lab/Dual-Interactive-Implicit-Neural-Network.git
$ cd Dual-Interactive-Implicit-Neural-Network

Then, create a running environment via conda

$ conda create -n myenv python=3.8
$ conda activate myenv

Finally, install the software requirements

$ conda install pytorch torchvision cudatoolkit=11.6 -c pytorch -c conda-forge
$ conda install pytorch-lightning=1.6 -c conda-forge
$ conda install omegaconf rich -c conda-forge

Datasets

The following datasets were used to evaluate this work:

• DIV2K

• Set5

• Set14

• B100

• Urban100

The data folder structure should be the following:

├── data
│   ├── DIV2K
│   │   ├── DIV2K_train_HR
│   │       │── 0001.png
|   |       │── ...
|   |       │── 0900.png
│   └── benchmark
│       ├── Set5
|       |   ├── HR
|       |       ├── ...
|       ├── Set14
|       |   ├── HR
|       |       ├── ...
|       ├── B100
|       |   ├── HR
|       |       ├── ...
|       └── Urban100
|           ├── HR
|               ├── ...
└── ...

Usage

Model Training

To train a model with the default configuration located in configs/default.yaml, run the following command

$ python main.py fit -c configs/default_test.yaml --model=SRLitModule --model.arch=diinn --model.mode=3 --model.init_q=False --trainer.logger=TensorBoardLogger --trainer.logger.save_dir=logs/ --trainer.logger.name=3_0

You may edit configs/default.yaml to best utilize your machine resources. For example, --model.mode=3 and --model.init_q=False is the configuration of our final model (i.e., model (f) in the paper).

Running Benchmarks

To benchmark a trained model with the datasets used in the paper, run

$ python benchmarks.py --ckpt_path=<path_to_checkpoint>

Super-Resolution Example

A low-resolution image can be super-resolved to a desired resolution as follows

$ python demo2.py --lr_path=<path_to_lr_image> --ckpt_path=<path_to_ckpt> --output_size <height> <width>

License