Home

Awesome

DSLR-release

This repository is a Pytorch implementation of the paper "DSLR: Deep Stacked Laplacian Restorer for Low-light Image Enhancement"

Seokjae Lim and Wonjun Kim
IEEE Transactions on Multimedia

When using this code in your research, please cite the following paper:

Seokjae Lim and Wonjun Kim, "DSLR: Deep Stacked Laplacian Restorer for Low-light Image Enhancement," IEEE Transactions on Multimedia DOI:10.1109/TMM.2020.3039361.

Experimental results with state-of-the art methods on the MIT-Adobe FiveK dataset

example1 Test samples from the MIT-Adobe FiveK dataset and corresponding enhancement results by previous methods and the proposed DSLR. (a) Origianl input. (b) CLAHE. (c) LDR. (d) LIME. (e) HDRNet. (f) DR. (g) DPE. (h) UPE. (i) DSLR (proposed). (j) Ground truth

Experimental results with state-of-the art methods on the MIT-Adobe FiveK dataset

example2 More examples of enhancement results on the MIT-Adobe FiveK dataset. (a) Origianl input. (b) CLAHE. (c) LDR. (d) LIME. (e) HDRNet. (f) DR. (g) DPE. (h) UPE. (i) DSLR (proposed). (j) Ground truth

Experimental results with state-of-the art methods on our own dataset

example3 Test samples from our own dataset and corresponding enhancement results by previous methods and the proposed DSLR. (a) Origianl input. (b) CLAHE. (c) LDR. (d) LIME. (e) HDRNet. (f) DR. (g) DPE. (h) UPE. (i) DSLR (proposed).

Experimental results with state-of-the art methods on our own dataset

example4 More examples of enhancement results on our own dataset. (a) Origianl input. (b) CLAHE. (c) LDR. (d) LIME. (e) HDRNet. (f) DR. (g) DPE. (h) UPE. (i) DSLR (proposed).

Requirements

Pretrained models

You can download pretrained DSLR model

Note

  1. you should place the weights in the ./data/model/
  2. Dataset is also placed in the ./data directory (i.e., ./data/training_dataset)
  3. Testset is placed in the ./data/test/input directory
  4. test results are saved in the ./data/result/

Training

python main.py n

Testing

python test.py t