Edge Detector Evolution using Multidimensional Multiobjective Genetic Programming

In this paper we report the evolution of a feature extraction stage for edge detection using multidimensional multiobjective genetic programming. We have employed training and validation data produced using a realistic model of the imaging physics to evolve an n-to-m mapping which projects the pixel intensities of an n × n image patch into an m-dimensional decision space. The (near-)optimal value of m is also simultaneously determined during evolution. A conventional Fisher linear discriminant is then used to classify edge patterns. On the independent validation set, the suggested edge detector is shown to give performance superior to both the well-known conventional Canny detector and to earlier multiobjective genetic programming results which projected the pattern vector into a one-dimensional decision space. In addition, the superiority of the new detector is also demonstrated on a hand-labeled set of real images.