Synergetic image recognition with applications to pose estimation

Synergetics is the study of systems in which individual sub-systems act co-operatively. In particular, we are interested as to how this co-operation on a microscopic scale can lead to the formation of macroscopic spatial structures or patterns. Synergetic pattern recognition is a non-conventional form of pattern recognition which is modelled on these synergetic pattern formation systems. Indeed the paradigm of synergetic pattern recognition states that pattern recognition is a type of pattern formation. This thesis is split into two parts. In Part A we introduce synergetic pattern recognition. A review of the current range of synergetic pattern recognition systems leads to a recognition of two major weaknesses in the current crop of synergetic pattern recognition algorithms, so we introduce a number of generalised pattern formation models which extend the capabilities of synergetic pattern recognition. We also investigate the concept of synergetic learning whereby learning is considered as a type of pattern formation. During a review of the current approach to this task we recognise a number of problems and solve them with an important observation concerning the dependence between variables. The new learning algorithm which results is a significant improvement over the current approach. In Part B of this work, we discuss the possible application of synergetic pattern recognition to the task of view-based pose estimation, which is the challenge of estimating the angles at which an image has been rotated, without reference to a full model of the object. Synergetic pattern recognition has not been used to solve this problem previously, so following a review of the view-based pose estimation literature, we introduce two new approaches to pose estimation based on synergetic pattern recognition. Comparison of the results found by the various techniques on a standard dataset reveal a number of common issues, foremost among which is the need to make a compromise between the precision of a system's estimate and the amount of time taken to produce the estimate. This observation leads to the most exciting finding of this work. We describe a new approach, not just to view-based pose estimation, but to the general field of pattern analysis, which we call explicit inversion. The reason for the name is that we have replaced the problem of numerically inverting a high-dimensional, unknown equation with analytically inverting a known, low-dimensional equation. Comparison with other approaches shows that this new approach yields comparable accuracy to current algorithms but with a dramatic reduction in calculation times. For datasets used in this dissertation, the increase in speed was between one to two orders of magnitude. We then apply our algorithms to the real-world task of tracking the pose of an aircraft. While a number of attempts have been made previously, this is the only approach which shows promise of being able to track aircraft pose in real-time.