Semantically Contrastive Learning for Low-light Image Enhancement

Here, we propose an effective semantically contrastive learning paradigm for Low-light image enhancement (namely SCL-LLE). Beyond the existing LLE wisdom, it casts the image enhancement task as multi-task joint learning, where LLE is converted into three constraints of contrastive learning, semantic brightness consistency, and feature preservation for simultaneously ensuring the exposure, texture, and color consistency. SCL-LLE allows the LLE model to learn from unpaired positives (normal-light)/negatives (over/underexposed), and enables it to interact with the scene semantics to regularize the image enhancement network, yet the interaction of high-level semantic knowledge and the low-level signal prior is seldom investigated in previous methods.

Network

• Overall architecture of our proposed SCL-LLE. It includes a low-light image enhancement network, a contrastive learning module and a semantic segmentation module.

Experiment

PyTorch implementation of SCL-LLE

Requirements

• Python 3.7
• PyTorch 1.4.0
• opencv
• torchvision
• numpy
• pillow
• scikit-learn
• tqdm
• matplotlib
• visdom

SCL-LLE does not need special configurations. Just basic environment.

Folder structure

The following shows the basic folder structure.

├── datasets
│   ├── data
│   │   ├── cityscapes
│   │   └── Contrast
|   ├── test_data
│   ├── cityscapes.py
|   └── util.py
├── network # semantic segmentation model
├── lowlight_test.py # low-light image enhancement testing code
├── train.py # training code
├── lowlight_model.py
├── Myloss.py
├── checkpoints
│   ├── best_deeplabv3plus_mobilenet_cityscapes_os16.pth #  A pre-trained semantic segmentation model
│   ├── LLE_model.pth #  A pre-trained SCL-LLE model

Test

• cd SCL-LLE

python lowlight_test.py

The script will process the images in the sub-folders of "test_data" folder and make a new folder "result" in the "datasets". You can find the enhanced images in the "result" folder.

Train

1. cd SCL-LLE
2. download the Cityscapes dataset
3. download the cityscapes training data google drive and contrast training data google drive
4. unzip and put the downloaded "train" folder and "Contrast" folder to "datasets/data/cityscapes/leftImg8bit" folder and "datasets/data" folder
5. download the pre-trained semantic segmentation model and put it to "checkpoints" folder

python train.py

Contact

If you have any question, please contact [email protected]