Evolving optimal feature extraction using multi-objective genetic programming
Yang Zhang, Peter I. Rockett · 2005
In this paper we describe a generic methodology to create an "optimal" feature extraction pre-processing stage for pattern classification. Our aim is to map the input data into a new, one-dimensional feature space in which separability is maximized under a simple thresholding classification. We have used multi-objective genetic programming with Pareto strength-based ranking to bias the selection procedure. The methodology is applied to the edge detection problem in image processing; we make quantitative comparison with the pre-processing stages of the well-known Canny edge detector using synthetic and real-world edge data and conclude that the performance of our evolutionary-based method is much superior to the Canny algorithm based on the criterion of minimum Bayes risk.