dark-images-dataset-mini-2
Data repository for LUMOS project.
Website link: https://abhishek-choudharys.github.io/dark-images-dataset-mini-2/
(The code will be published soon)
About the parent data:
The parent data folder has been acquired from Exclusively Dark Dataset.
Official github repo link https://github.com/cs-chan/Exclusively-Dark-Image-Dataset
The original dataset contained images captured in low light for 12 different classes.
Original class labels: [‘Bicycle’, ‘Cup’, ‘People’, ‘Cat’, ‘Car’, ‘Boat’, ‘Motorbike’, ‘Dog’, ‘Chair’, ‘Bottle’, ‘Table’, ‘Bus’]
Each class had over 500 images in the parent dataset.
This ensures that the dataset contains images from a wide range of domains, and does not restrict itself to a special category or images.
Mini-Dataset:
As needed, the new mini-dataset has been created by randomly selecting 50 images from each class from the parent dataset. This ensures that images from all classes are included to make the dataset as varied as possible. All the images could not be used because of the unavailability of powerful computational resources.
This equals to a total of 600 images with no class labels.
The code used to create and test the dataset has been included in preprocess_push.ipynb.
The testing was done on MBLLEN model.
Link https://github.com/Lvfeifan/MBLLEN
Data Augmentation:
Data augmentation was also performed to increase the size and scope of the dataset. The process involved adding synthetic noise and blur to original 600 images, inorder to test the performance of enhancement models on a wider range of data and distortions. The added effects are as follows:
- Gaussian Blur (identifiable by the suffix ‘_B’).
- Motion Blur (identifiable by the suffix ‘_M’).
- Poisson Noise (identifiable by the suffix ‘_P’).
The addition of synthetic data increase the size of the dataset to a total of (600x4=)2400 images.
Resizing strategy:
To optimize performance, the images had to be resized to a much practical range, so the results from all enhancement models could be generated in lesser time using modest computing power.
The images were resized to x% of their original size with x being:
- 100 if the images had height or width less than 400 pixels.
- 60 if the images had height or width between 400 to 800 pixels.
- 40 if the images had height or width between 800 to 1500 pixels.
- 20 if the images had height or width more than 1500 pixels.
Processed Data:
Each of the 600 images have been passed through the following models to generate enhanced results. The images have been scaled to x% of their original size, while maintaining aspect ratio, to account for modest computational resources.
- MBLLEN_data contains image files enhanced by the MBLLEN model (https://github.com/Lvfeifan/MBLLEN). The code can be found in MBLLEN.ipynb.
- GLADNet_data contains image files enhanced by the GLADNet model (https://github.com/weichen582/GLADNet). The code can be found in GLADNet.ipynb.
- RetinexNet_data contains image files enhanced by the RetinexNet model (https://github.com/weichen582/RetinexNet). The code can be found in RetinexNet.ipynb.
- CE_data contains image files enhanced by classical image processing pipelie. The code can be found in CE_generator.ipnyb.
- PPRP_data contains image files enhanced by Pure Pixel Ratio Prior approach. The code can be found in PPRP_generator.ipnyb.
- PCA_data contains image files enhanced by Principal Component Analysis approach. The code can be found in PCA_generator.ipnyb.
Sample Results:
Models Used:
Comments:
- Added model images
- Added MIT License
TODO:
- Change resizing strategy. </s>
- add noise code</s>
- add noisy data
- add LICENSE
- add Atmospheric scattering of light model
- add PCA Model
- add graphs
Key:
- src = ‘data_aug’
- src1 = ‘MBLLEN_data’
- src2 = ‘GLADNet_data’
- src3 = ‘RetinexNet_data’
- src4 = ‘CE_data’
- src5 = ‘PPRP_data’
- src6 = ‘PCA_data’
Scores:
- score1 = ‘BRISQUE’
- score2 = ‘Color Scores’
- score3 = ‘PIQE’
- score4 = ‘NIQE’
Website: https://abhishek-choudharys.github.io/dark-images-dataset-mini-2/