[FFmpeg-devel] [GSoC] proposal idea: Color Constancy using Machine Learning

[Mina]

Hi,

   I'm proposing a Color Constancy filter as a project idea for GSoC and 
was asking for any remarks about that idea.

   Color constancy is the ability of the human visual system that 
ensures the perception of the color of objects to remain relatively 
constant under varying illumination conditions. A color constancy filter 
would basically be applied on images with various degrees of damage to 
restore original colors as close as possible.

This <https://imgur.com/a/hj2ox> is a simple example of a low-statistics 
filter output I've implemented as an academic project along with the 
original damaged image. The filter is an implementation of a 
mathematical equation presented with some parameters that's used to 
introduce variations to the filter.

   Use cases: despite being an end-result filter that helps retrieving 
damaged images original colors it's very helpful as a pre-processing 
utility for many computer vision algorithms. A simple example would be a 
classifier that identifies an object utilizing its color.

Implementation: this is still a research filed where last paper I've 
read was out earlier this year and still more and more papers are being 
published introducing various learning-methods to improve accuracy 
including unsupervised and supervised methods and even new 
learning-based algorithms to improve low-statistics light-computational 
ones. Variations of this filter and their accuracies will also prove to 
be a very appropriate project for GSoC since there is many factors 
affecting the choice of filter such as real-time need, filed of 
application, available dataset, ..etc. So, it's more like a group of 
filters.

   Impact: introduction of machine learning to this filed is quiet 
recent so enhancing FFmpeg with this technology and possibly adding to 
it will prove entrepreneurship. Photography, Computer Vision and many 
other fields that performs fine image processing would make good use of 
this filter.

Best regards,
Mina Sami.