\[CVPR 2024\] Look-Up Table Compression for Efficient Image Restoration

Yinglong Li, Jiacheng Li, Zhiwei Xiong

CVPR Paper

Usage

Updating! Any questions, please contact me at any time.

Dataset

task	training dataset	testing dataset
super-resolution	DIV2K	Set5, Set14, BSDS100, Urban100, Manga109
denoising	DIV2K	Set12, BSD68
deblocking	DIV2K	Classic5, LIVE1
deblurring	GoPro	GoPro test set

Pretrained Models

Some pretrained LUTs and their compressed version can be download here.

Step 1: Training LUT network

Let's take the SPF-LUT for x4 sr as an example.

cd sr
python 1_train_model.py --model SPF_LUT_net --scale 4 --modes sdy --expDir ../models/spf_lut_x4 --trainDir ../data/DIV2K --valDir ../data/SRBenchmark

The trained LUT network will be available under the ../models/spf_lut_x4 directory.

Step 2: Transferring LUT network into compressed LUTs

python .\2_compress_lut_from_net.py --model SPF_LUT_net --scale 4 --modes sdy --expDir ../models/spf_lut_x4 --lutName spf_lut_x4 --cd xyzt --dw 2 --si 5

The compressed LUTs will be available under the ../models/spf_lut_x4 directory. --cd: The number of compressed dimensions; --dw: Diagonal width; --si: Sampling interval of non-diagonal subsampling.

Step 3: Fine-tuning compressed LUTs

python 3_finetune_lut.py --model SPF_LUT_DFC --scale 4 --modes sdy --expDir ../models/spf_lut_x4  --trainDir ../data/DIV2K --valDir ../data/SRBenchmark --load_lutName spf_lut_x4 --cd xyzt --dw 2 --si 5

The finetuned compressed LUTs will be available under the ../models/spf_lut_x4 directory.

Step 4: Test compressed LUTs

python .\4_test_SPF-LUT_DFC.py --scale 4 --modes sdy --expDir ../models/spf_lut_x4 --testDir ../data/SRBenchmark --lutName weight --cd xyzt --dw 2 --si 5

Contact

If you have any questions, feel free to contact me any time by e-mail yllee@mail.ustc.edu.cn

Citation

If you found our implementation useful, please consider citing our paper:

@InProceedings{Li_2024_CVPR, 
	author = {Li, Yinglong and Li, Jiacheng and Xiong, Zhiwei}, 
	title = {Look-Up Table Compression for Efficient Image Restoration}, 
	booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, 
	month = {June}, year = {2024}, pages = {26016-26025} 
}

Acknowledgement

This work is based on the following works, thank the authors a lot.

SR-LUT

MuLUT