Evolving texture features by Genetic Programming

Feature extraction is a crucial step for Computer Vision applications. Finding appropriate features for an application often means hand-crafting task specific features with many parameters to tune. A generalisation to other applications or scenarios is in many cases not possible. Instead of engineering features, we describe an approach which uses Genetic Programming to generate features automatically. In addition, we do not predefine the dimension of the feature vector but pursue an iterative approach to generate an appropriate number of features. We present this approach on the problem of texture classification based on co-occurrence matrices. Our results are compared to those obtained by using seven Haralick texture features, as well as results reported in the literature on the same database. Our approach yielded a classification performance of up to 87% which is an improvement of 30% over the Haralick features. We achieved an improvement of 12% over previously reported results while reducing the dimension of the feature vector from 78 to four.