Is a single image sufficient for evolving edge features by genetic programming?

Abstract

Typically, a single natural image is not sufficient to train a program to extract edge features in edge detection when only training images and their ground truth are provided. However, a single training image might be considered as proper training data when domain knowledge, such as used in Gaussian-based edge detection, is provided. In this paper, we employ Genetic Programming (GP) to automatically evolve Gaussian-based edge detectors to extract edge features based on training data consisting of a single image only. The results show that a single image with a high proportion of true edge points can be used to train edge detectors which are not significantly different from rotation invariant surround suppression. When the programs separately evolved from eight single images are considered as weak classifiers, the combinations of these programs perform better than rotation invariant surround suppression.