Discrimination properties of invariants using the line moments of vectorized contours

In this paper a new approach for image analysis in real time based on the vectorized contours of a scene is presented. Taking advantage of the fast and efficient determination of the line moments, invariants with respect to translation, scaling and rotation are derived. Four different sets of rotational invariants are introduced and their performance is examined on two examples. Moreover, a quality measure for class discrimination of feature sets is presented and investigated. Using this quality measure as a cost function, heuristic search strategies and genetic algorithms are employed for the feature selection. Thus, high dimensional feature spaces are reduced significantly without losing relevant image information. The performance of both the full and the reduced data sets is investigated on a set of noisy patterns and on a set of letters