Classification Strategies for Image Classification in Genetic Programming

This paper describes an approach to the use of genetic programming for multi-class image recognition problems. In this approach, the terminal set is constructed with image pixel statistics, the function set consists of arithmetic and conditional operators, and the fitness function is based on classification accuracy in the training set. Rather than using fixed static thresholds as boundaries to distinguish between different classes, this approach introduces two dynamic methods of classification, namely centred dynamic range selection and slotted dynamic range selection, based on the returned value of an evolved genetic program where the boundaries between different classes can be dynamically determined during the evolutionary process. The two dynamic methods are applied to five image datasets of classification problems of increasing difficulty and are compared with the commonly used static range selection method. The results suggest that, while the static boundary selection method works well on relatively easy binary or tertiary image classification problems with class labels arranged in the natural order, the two dynamic range selection methods outperform the static method for more difficult, multiple class problems.