Date: Wednesday 2nd March 2016
Speaker: Professor James Orwell (Security Theme Leader, Visual Surveillance Research Group)
Chair: Dr Sarah Barman
Signal processing algorithms have been devised to address a number of distinct challenges, operating on diverse modalities of input data. In the security domain, some typical examples include face recognition, and intruder detection. There are diverse ways of measuring the performance of these algorithms, which makes comparisons between problems, methods and datasets more difficult to undertake. This also makes it less straightforward to obtain an intuitive understanding of algorithm performance. In short, signal processing is currently not well served by its performance metrics. It is argued that Information Theory can provide both the guiding principle and the working details for an approach that allows straightforward comparisons and provides an intuitive grasp. It is proposed that, for signal processing problems, `the proportion of uncertainty removed’ is the appropriate metric to use. The details are provided for recognition, detection and prediction problems, and these are shown to exhibit the necessary properties.