pillow-dat 0.1.7

PIL × DAT - Pillow extension for AI-based image upscaling.

PIL × DAT - Pillow extension for AI-based image upscaling.

---

Installation

For PyPI:

pip install pillow-dat

Get started

from PIL.Image import open

from PIL_DAT.Image import upscale

lumine_image = open(".github/lumine.png")
lumine_image = upscale(lumine_image, 2)

Remark: We strongly advocate for the utilization of DAT light models owing to their streamlined design and outstanding speed performance. However, should you opt for alternative models, please note that *.pth model weights can be accessed via Google Drive.

Example

Input (lumine.png)	DAT light (x2)	Bicubic (x2)

Benchmarks

Speed

Performance benchmarks have been conducted on a computing system equipped with an Intel(R) CORE(TM) i7-9750H CPU @ 2.60GHz processor, accompanied by a 2 × 8 Go at 2667MHz RAM configuration. Below are the recorded results:

In seconds	320 × 320	640 × 640	960 × 960	1280 × 1280
DAT light (x2)	13.7	54.9	127.2	299.3
DAT light (x3)	13.2	56.5	-	-
DAT light (x4)	12.8	56.6	-	-

The results were compared against the renowned OpenCV library, utilizing its EDSR model known for delivering superior image quality.

In seconds	320 × 320	640 × 640	960 × 960	1280 × 1280
EDSR (x2)	25.6	112.9	264.1	472.8
EDSR (x3)	24.3	112.5	-	-
EDSR (x4)	23.6	111.2	-	-

Remark: All speed benchmark results presented here are reproducible. For detailed implementation, please refer to the following files: benchmark_speed_dat_light.py and benchmark_speed_edsr.py.

Quality

DAT light (x2)	EDSR (x2)

Remark: All quality benchmark results presented here are reproducible. For detailed implementation, please refer to the following files: example.py and benchmark_quality_edsr.py.

Contribution

Please install Miniconda.

Please install VSCode extensions:

• Black Formatter
• isort
• Python
• Pylance

To create or update the pillow-dat Python environment:

conda env create --file environment.yml

conda env update --file environment.yml --prune

To install dependencies:

poetry install

To run unit tests:

pytest

Acknowledgement

This library is founded upon the pioneering research paper, "Dual Aggregation Transformer for Image Super-Resolution".

@inproceedings{chen2023dual,
    title={Dual Aggregation Transformer for Image Super-Resolution},
    author={Chen, Zheng and Zhang, Yulun and Gu, Jinjin and Kong, Linghe and Yang, Xiaokang and Yu, Fisher},
    booktitle={ICCV},
    year={2023}
}