DeepImageDeblurring

Deep Convolutional approaches for Gaussian and Motion Blur removal.

Project Report

Keras implementation of the models and techniques discussed in the report above. These implementations has been developed by Eleonora Mancini and Marcello Simonati as a project work for the Deep Learning course, part of the Master's Degree in Artificial Intelligence in University of Bologna (A.A. 2019-2020).

Our network takes blurry image crops as an input and produce the corresponding sharp estimate, as in the example:

How to run

Prerequisites:

• NVIDIA GPU + CUDA CuDNN (CPU untested)
• Tensorflow 2.10 or superior
• Numpy
• Pandas
• Matplotlib

To run the scripts just clone the this repository on your workstation and execute them. Executing a script with the -h argument will pop up the help, showing you how to use them.

You can dowload our networks pretrained on CIFAR10 and REDS by clicking on the following link:

Models and Weights

Data

You can run the models using the CIFAR10 dataset, which will download automatically on request, or by setting a custom folder for video data. The data folder should have the following structure:

data
│
└───train
│   │
│   └───train_blur
│   |   │   001
│   |   │   002
│   |   │   ...
│   └───train_sharp
│       │   001
│       │   002
│       │   ...
└───val
│   │
│   └───val_blur
│   |   │   001
│   |   │   002
│   |   │   ...
│   └───val_sharp
│       │   001
│       │   002
│       │   ...
└───test
    │
    └───test_blur
    |   │   001
    |   │   002
    |   │   ...
    └───test_sharp
        │   001
        │   002
        │   ...

Where inside the 001, 002, ... folders you should find a sequence of images belonging to the same video, numbered accordingly in order to have subsequent frames next to each other.

This folder structure is the same adopted in the REDS Dataset which is the dataset used for our work.

Model Creation

You can use the create_model.py script to create a model directory and it will serialize the network to a .json file inside it. This directory will be the working directory for all other computations.

Model Training

You can use the train_model.py script to train the model inside the specified directory. The end weights will be saved inside the model directory along with the training history in .csv format.

Model Testing

You can test the trained model using the test_model.py script. You can also test our pretrained model by downloading the files linked above. Testing has two modes:

• evaluate: will run an evaluation on the entire test set
• batch: will generate the provided number of batches of images predicted by the network and compared with its blurred and sharp version, in a folder inside the model directory.