Intro

A small project to estimate the motion blur of a camera using a convolutional neural network, to be later used to calibrate a deconvolution algorithm.

There are two parts to this project :
- the Image Processing Part with the deconvolution algorithms
- the Deep Learning part that aims to estimate the blur

See the wiki for some visual insights.

The blur is for the moment linear. Currently, we only try to estimate the angle, later we'll switch to the length.

Requirements

The project is coded using Keras with a Tensorflow backend. We use python3. You can install the dependencies using pip install requirements.txt.
We currently use opencv but will most likely change to lighter library in the future.

Implementation

• The linear kernel is obtained by integrating a line over one pixel so as to take into account discretization effects.
• The deconvolution is a Wiener filter. We plan to add a TV deconvolution.
• Data and training: this project has been implemented on my personal computer that does not possess a GPU and has limited RAM. Therefore the training data consisted of one image linearly blurred at random angles (but constant intensity blur). This of course yields less than optimal performances on real life data but provides notheless a proof of concept.

Usage

• To see visual results over randomly sampled angles: Run python Main.py -l 1 -p models/AngleWeights5
• Test scripts: go to tests and run the scripts to see what the linear blue does, how the kernel is computed and how deconvolution works with a given kernel.