Image/video annotation and retrieval by taking advantages of both deep learning and traditional approaches

Project title:

Image/video annotation and retrieval by taking advantages of both deep learning and traditional approaches

Supervisors:

• Professor Guojun Lu (Principal)
• Professor Manzur Murshed
• Associate Professor Shyh Wei Teng
• Dr Suryani Lim (ECR)

Contact person and email address:

• Professor Guojun Lu: guojun.lu@federation.edu.au

Project description:

Current content-based image retrieval systems are mainly based on similarity of low level features. They are not ideal as humans recognise/retrieve images based on high concepts (e.g. objects and their relationships within images) and similar high level concepts may have very different low level features. This project will develop a new image retrieval method that will derive high level concepts and then annotate images automatically. The annotated images can then be organised and retrieved effectively and efficiently using text based retrieval techniques. There are currently two approaches to derive high level concepts. In the traditional approach, image features such as colour, texture and shapes are derived and then the high level concepts are derived from a combination of these features. The other approach uses deep learning to learn high level concepts from images directly. There are advantages and disadvantages in both approaches. We will explore advantages of these two methods to develop an effective and efficient method for image annotation.