Many computer vision systems make use of artificial neural networks to undertake complex image analysis tasks, such as locating objects in a clutter scene. Currently these are trained in advance to undertake a specific task and then deployed. The performance of the deployed network is heavily dependent on the quality of the training step.
The problem with artificial neural networks is that for good training they require large quantities of accurately labelled training data, e.g. images of a scene with the target objects defined and labelled. Creating this data is time consuming and costly, as it must consider all possible environmental influences on the scene (e.g. lighting, perspective, shadows, reflections, etc) and it is therefore not always possible to include a full set of all possible scenarios in the training data.
This project will investigate new ways of training artificial neural networks for image analysis. In particular it will consider how the cost an time of network training can be reduced, and how the system can improve performance even after the training stage by self-learning during service. This PhD has the potential to transform the way artificial intelligence is applied to many situations, making it cost effective and suitable for many new application areas.
The work will be undertaken within the Loughborough University Centre for Intelligent Automation, as part of a multi-disciplinary team of Engineers and Scientist who are focused on opening up new areas of potential for robotics and automation.
- Applicants should have, or expect to achieve, at least a 2:1 Honours degree (or equivalent) - A relevant Master's degree and / or experience in one or more of the following will be an advantage: - Camera-based measurements - Computer/machine vision - Artificial intelligence or machine learning - Automation and robotics