Unofficial implementation of the DeblurGAN model based on the paper DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks. With this repository, you will be able to implement the model from scratch using your own dataset. The steps to train and test the model, along with the project details, are explained below.
To prepare the dataset, you should first run the file dataset.py, setting the parameters at the beginning of the file as desired. These parameters are:
- dataset_path: specify the paths where the images to be used for training are located, separated by the character
;. - train_split: defines the ratio for splitting images into training and validation sets. For example, if there are 100 images and this parameter is set to 0.8, then 80% of the images will be used for training and the remaining 20% for validation.
The script lists all image paths in the given folders, shuffles them randomly, and then splits them into training and validation sets according to the train_split parameter. Finally, two JSON files will be generated: one containing the training image paths and the other containing the validation image paths. These JSON files will be used during training to load the images. I trained the model on all the scenarios from the TartanAir dataset, which contains image sequences at 480×640 resolution.
To train the model, you can adjust the parameters and hyperparameters as needed. In my experiments, I used the same values as in the original paper, but I implemented a different approach to blur the images. Specifically, I use the OpenCV function GaussianBlur, selecting a kernel size of 5 and a sigma value of 5.0 as well. To start the training, run:
python train.py
When the training starts, a folder is created. The folder name consists of the name provided in the parameters, followed by relevant information about the training parameters, such as batch size, learning rate, number of residual blocks added to the network, and weighting parameters for the loss function. This way, you can always identify the parameter values used for the training. In my case, the name of my training folder is deblurGAN_bs1_lr0.0001_numresblocks9_lambdaG100_lambdaD10/. Five txt files are stored in this folder, containing the values of the generator loss and the discriminator loss and accuracy at each epoch. Additionally, within this folder, a subfolder called checkpoints/ is created to store the model every 10 epochs.
To test the model, specify the checkpoint to be evaluated in the parameters, along with the name of the folder containing the model file. Place the images you want to deblur in the test_images/original/ folder and run:
python test_images.py
The Gaussian-blurred images and the model-deblurred ones are saved in the test_images/blurred/ and test_images/deblurred/ folders, respectively, with the same file name. I’ve uploaded an example image to this repository, which you can also see below. The image corresponds to a screenshot from the video game Assassin’s Creed Valhalla.
The same process used for images can also be applied to videos. You can place some video files in the test_videos/original/ folder, and run
python test_videos.py
This will generate the blurred and deblurred versions of the videos in the test_videos/blurred/ and test_videos/deblurred/ folders, respectively, using the same file names.
You can check the following video to see the result of the model after processing a fragment from the video game Red Dead Redemption: https://www.youtube.com/watch?v=22fwREg7PO4
NOTE: the files in these folders are not real video files, they are just empty placeholders to illustrate the expected structure. Similarly, the files in the checkpoints/ folder are not real model files and are also empty placeholders for demonstration purposes.
