Content-based image retrieval techniques such as the reverse image search tools offered by Google and TinEye are popular among those wanting to trace the origin of a photo or find similar pictures.
The technology, which has many applications, is far from perfect though, particularly as the volume of visual data hosted in databases rapidly increases, making the retrieval of a specific piece of information increasingly difficult.
With the support of an Australian Research Council Discovery Grant worth $445,000 over three years, Associate Professor Lei Wang from the Faculty of Engineering and Information Sciences and the Visual Information Learning and Analysis research group (Group Leader), aims to improve the accuracy of image retrieval techniques using some innovative methods.
“Content-based image retrieval (CBIR) is theoretically significant because it addresses a central issue in computer vision—how to decide whether two images are similar or not,” A/Prof Wang, whose research interests include machine learning, pattern recognition and computer vision, said.
“Image retrieval and analysis touch more of our everyday life than most would imagine.”
“It’s a piece of technology that’s used in improved logistics, modelling, regulation, urban design, electrified transport, sensor technologies, real-time data and spatial analysis. It’s used in everything from self-driving cars to environmental monitoring and medical diagnosis.”
“This Discovery Project proposes to develop high-order visual representation, a newly emerging powerful representation for CBIR,” A/Prof Wang explained.
The ARC project, titled ‘Learning kernel-based high-order visual representation for image retrieval’, aims to improve the technology’s capacity to classify objects that are similar but have subtle differences, such as different birds, cars, or aircraft.
“That process is known as fine-grained image classification and it’s required to further improve the precision of current image retrieval techniques.”
A/Prof Wang said the project is expected to have benefits for theory and practice.
“Outcomes are expected to include theory development on visual representation, more effective retrieval techniques, improving the computational efficiency in learning and utilising high-order visual representation and developing foreground object focused high-order visual representation to handle the background clutter in image retrieval.”
“Developing advanced retrieval techniques has great potential to contribute to Australia's national interest and provide substantial economic benefit also by significantly automating visual information access and increasing the efficiency and precision of information extraction. This can lower labour costs and improve productivity.”
“Without effective image retrieval techniques, most, if not all, of the stored visual data across our networks will become inaccessible, causing a significant waste of time, money and labour.”
- ASSOCIATE PROFESSOR LEI WANG
To find out more about Associate Professor Lei Wang visit his Scholars profile