A comparative study of multi-objective evolutionary trace transform methods for robust feature extraction

Abstract

Recently, Evolutionary Trace Transform (ETT) has been developed to extract efficient features (called triple features) for invariant image identification using multi-objective evolutionary algorithms. This paper compares two methods of Evolutionary Trace Transform (method I and II) evolved through similar objectives by minimizing the within class variance (Sw) and maximizing the between-class variance (Sb) of image features. However, each solution on the Pareto front of method I represents one triple features (i.e. 1D) to be combined with another solution to construct 2D feature space, whereas each solution on the Pareto front of method II represents a complete pair of triple features (i.e. 2D). Experimental results show that both methods are able to produce stable and consistent features. Moreover, method II has denser solutions distributed in the convex region of the Pareto front than in method I. Nevertheless, method II takes longer time to evolve than method I. Although the Trace transforms are evolved offline on one set of low resolution (64 ×64) images, they can be applied to extract features from various standard 256×256 images.