Deep neural network for photographing in the dark (GONGH2U19SF)

Have you ever struggled in taking a good photo in low light? Despite of the dramatic improvements in camera sensors, imaging in darkness is still challenging because in many conditions, there is simply too little light on the scene and the imaging sensors are sensitive to noise in these conditions. This may be compensated by increasing exposure time. However, a problem is that this also introduces motion blur and we need to reduce video capturing frame rate to minimise the number of blurs. Meanwhile, some image de-noising, de-blurring and enhancement methods (e.g. [ii-iii]) have been proposed, but their applicability is limited in dim light conditions and not suitable for the real-time video capture in the extremely low light condition.

In this PhD project, we plan to develop a pipeline for enhancing nearly black images/videos taken in low light. The solution will be based on end-to-end training of a fully-convolutional network [iv]. The network directly operates on raw sensor data and is expected to perform re-lighting, tone mapping, and noise removal. We will start by developing a supervised learning approach which does require some ground truth pairs of dark images and well-lit images for training. In the next stage, we will further investigate an unsupervised learning framework which doses require any corresponding ground truth image pairs. Instead, the training will be based on analysing different un-annotated sets of dark and well-lit images available from the Internet. This is a higher level of intelligence that self-learns without a “supervisor”. The possible unsupervised learning solution could be based on an image-to-image Generative Adversarial Network (GAN) [v]. We will also compare our solution with start-of-the-art methods (e.g. [i]) using both quantitative evaluation and user experiments.

The outcome of this project is expected to provide the powerful applications in smartphone camera development and night-time surveillance.

A start date prior to October 2019 is possible but should be discussed with Dr Gong in the first instance.

For more information on the supervisor for this project, please go here: https://people.uea.ac.uk/en/persons/h-gong
The type of programme: PhD  
The start date of the project: October 2019
The mode of study: full-time
Entry requirement: Acceptable first degree in Computer Science, Mathematics, Physics (or related degree), the standard minimum entry requirement is 2:1