Design of composite correlation filters for object recognition using multi-objective combinatorial optimization

Correlation filters for object recognition represent an attractive alternative to feature based methods. These filters are usually synthesized as a combination of several training templates. These templates are commonly chosen in an ad-hoc manner by the designer, therefore, there is no guarantee that the best set of templates is chosen. In this work, we propose a new approach for the design of composite correlation filters using a multi-objective evolutionary algorithm in conjunction with a variable length coding technique. Given a vast search space of feasible templates, the algorithm finds a subset that allows the construction of a filter with an optimized performance in terms of several performance metrics. The resultant filter is capable of recognizing geometrically distorted versions of a target in high cluttering and noisy conditions. Computer simulation results obtained with the proposed approach are presented and discussed in terms of several performance metrics. These results are also compared to those obtained with existing correlation filters.