We have all taken photos from our smartphone cameras in low light – of people, places or food. And then inevitably we turn to filters in order to increase the brightness and add context to the picture. How often has it turned out the way we wanted it to?
Thanks to a group of researchers from Intel and the University of Illinois Urbana-Champaign, we now have another approach to turn up the brightness in a low light photo, with stunning accuracy. As you might have guessed by now, deep learning, and more specifically computer vision and pattern recognition, is at the core of this approach.
The above example was shared by the researchers on their GitHub page. The results are truly mesmerizing.
The algorithm behind this technique has trained the model on how an image, taken in awfully poor light, should be brightened and coloured. In order to build the said model however, the researchers first curated a dataset of 5,094 raw short-exposure low-light images, with the same number of corresponding long-exposure reference images as well.
At the core of the algorithm is a convolutional neural network that operates directly on raw sensor data and replaces much of the traditional image processing pipelines, which have previously tended to perform poorly on such data. The results have been promising so far.
You can read their official research paper here and access the GitHub library here. Check out the below video illustrating this technique:
The advantages of this deep learning algorithm are obvious. This will not only help professional photographers in their work but will make data scientists into photographers as well! Instead of spending big bucks on expensive software and working on minute details, you now have the option to just leverage this model and get to work on low level images.
It’s the perfect algorithm for data scientists getting started in the image processing, computer vision and patter recognition fields.
Lorem ipsum dolor sit amet, consectetur adipiscing elit,