Special CISTO (Code 606) Seminar 
10:30am, Monday, July 24th
Building 28, Room W176 

Pragmatic Evolutionary Methods for Object Detection and Image Understanding and a Novel Adaptive Method of Signal Analysis 

Dr. Daniel Howard
QinetiQ PLC (QinetiQ Inc USA) (formerly the Defence Research Agency of the United Kingdom) 

Over the years QinetiQ has developed a number of methods for detection of military targets in reconnaissance imagery.  Many of these are based on the evolutionary computation paradigm.  We have used this paradigm to design intelligent partitions of the data; to take into account context; to develop fast detectors that can help photographic interpreters scourge through imagery abiding by the NATO SUPLAN HOTEL requirement (en force for exploitation of RECCE imagery in a certain time window).  Later we used the same paradigm and its power to enforce a search strategy/architecture to help interpreters develop methods of context-sensitive analysis (image understanding) to intelligently reduce false alarm rates.  The latest move has been for a system that interacts with users.  In the second half of this talk I will present a very useful novel method (patented in the USA) of signal processing.  This allows for: (a) interpolation; (b) approximation; and (c) quasi-interpolation (peak sharpening) of signals.  At worst, it is a very useful generalization and, at best, it is a completely novel functional method of functional recovery.  It has two principal current advantages: (a) it can handle non-uniformly distributed input data with ease to provide a spectral approximation/ interpolation which is infinitely differentiable (derivates are also available) and (b) it contains a parameter $\sigma$ which can be adapted with, for example, a genetic algorithm to suit a particular application.  I will demonstrate the method in the following settings: (1) removing the baseline drift in equipment such as mass spectrometers or any instrumentation that has hard to characterize noise; (2) a novel method of image compression; (3) a novel method of solution of non-self-adjoint differential equations (e.g. convection diffusion, CFD, etc) that can rival weighted residual methods (e.g. finite differences or finite elements) while circumventing the problems that such methods have with the non-self adjointness (lack of ellipticity) of the differential equations.  The new method of signal processing should find general application in science, engineering, and computer graphics (e.g. as an alternative to NURBS: nonuniform rational B splines). 

PROFESSIONAL BIOGRAPHY

After obtaining his MS and PhD degrees in Computational Fluid Dynamics at University College Swansea in the UK, Dr. Howard worked at the Rutherford Appleton Laboratory (a supercomputing facility in the UK).  He was then elected a Rolls-Royce Research Fellow of Pembroke College, Oxford University and also a Research Fellow in the Numerical Analysis Group at the Oxford University Computing Laboratory.  Following completion of this period, Daniel spent five years in the oil industry developing practical systems for engineers before rejoining a research laboratory, an agency of the UK Ministry of Defence known as DERA, the Defence Evaluation and Research Agency of the United Kingdom, based at Malvern (where the Morgan car is made). In 2002, Daniel was elected a Company Fellow of DERA and in 2003 DERA was privatised and is known today as QinetiQ PLC. QinetiQ now has US offices and a number of subsidiaries in the USA (e.g. Westar).  Daniel has been with DERA/QinetiQ for 10 years where he has led a specialist team that has made contributions to Genetic Programming theory and practice; machine vision; traffic modelling; genomics; mammography screening (building a taxonomy of mammograms and with Laszlo Tabar, the grandfather of mammography in central Sweden and winner of the Gold Medal from the American Society of Breast Imaging) and other important research topics.