Overview
- EPSRC
- Cortexica Vision Systems
- The National Archives
- Research Councils UK (RCUK)
- Royal Society
- Royal Academy of Engineering
- Faculty of Engineering, Imperial College London
- American College of Cardiology
Historically, the research of the group had a focus on computational models of the mammalian visual cortex. Specifically, the group wavelet-based models to describe the spatial sensitivity functions of visual neurons. These models are similar in feel to Gabor models for describing simple-cell behaviour in the primate visual cortex, but were optimised for the properties of speed, precision of representation and adaptability. These properties make such models suitable for use in artificial intelligence, particularly for applications in image analysis and computer vision.
Published work containing examples of the use of these models is to be found in:
- Image Denoising, via adaptation of wavelet bases
- Feature detection, using phase-invariant properties
- Direction of visual attention
Please see the BICV website (www.bicv.org) for a discussion on the connections between some of our prior work and modern convolutional networks.
Related Work
Models of the Mammalian Retina (with Hiroshi Momiji)
The mammalian retina performs both spatial and temporal processing of visual data, prior to conveying it to the brain. Dynamical models of individual cells are a powerful, yet tractable way of understanding the relationships between cell physiology, connectivity, and visual function (see Momiji et al, 2006). This has been used to study the behaviour of colour-opponent retinal responses, and its relationship to colour perception (Momiji et al, 2007).
Motion Analysis
Motion computation plays a key part in our perception of visual data. In conjunction with the late Professor MJ Lever, Dr Spyretta Golemati and Professor A Nicolaides, our group has worked on analysing the motion of the carotid artery from dynamic B-mode ultrasound scans. More recently, we have combined wavelet pyramids to the analysis of human motion in interview settings.
Analogue Temporal-Derivative CNNs
Biological computation in the visual cortex involves massively parallel processing of the visual field that is incident on the retina. Performing such computation is demanding, and digital implementation consumes too much power for cheap autonomous systems. With appropriate models of spatial and temporal receptive fields, and with rigorous analogue design, pwoer-efficient mixed mode VLSI implementations are possible. Working with Henry Ip and Dr E Drakakis allowed the development of the temporal derivative CNN (TDCNN). This is a generic VLSI network for spatiotemporal filtering using novel circuitry that employs derivative-diffusion between network nodes.
The Generic Vision Platform
The GVP is a wavelet-based plaform for rapid generation of low-level image features. These features can be generated in real-time, by utilising relative standard computer hardware, and are based on models, developed over many years, of biological visual processe. See the BICV website (www.bicv.org) for more about the techniques that we use today, and a brief discussion on the relationship to convolutional nerual networks.
Collaborators
Prof Zhaoping Li, University College London, Neuroscience
Professor Petrou, Biologically Inspired Computer Vision
Professor Cheung, Imperial College London, FPGA Implementations for Computer Vision
Prof Nick Kingsbury, University of Cambridge, Complex Wavelet Methods
Dr Dragotti, Wavelet representations; Cognitive Systems
Dr Mandic, Cognitive Systems
Professor Guang-Zhong Yang, Imperial College London, Focus-of-attention, Cognitive Systems
Christopher Kennard, Imperial College London, Visual Neuroscience
Research Staff
Research Student Supervision
Engin,Z, Biologically Inspired Low-Level Visual Computation
Ip,M, Analogue, Biologically-Inspired Spatiotemporal Filters
Lim,M, Visual Search in Human and Machine Vision
Roche,A, Neural Substrates of Biological Motion Perception