Genetic Programming for Image Classification: A New Program Representation With Flexible Feature Reuse

Extracting effective features from images is crucial for image classification, but it is challenging due to high variations across images. Genetic programming (GP) has become a promising machine-learning approach to feature learning in image classification. The representation of existing GP-based image classification methods is usually the tree-based structure. These methods typically learn useful image features according to the output of the GP program's root node. However, they are not flexible enough in feature learning since the features produced by internal nodes of the GP program have seldom been directly used. In this article, we propose a new image classification approach using GP with a new program structure, which can flexibly reuse features generated from different nodes, including internal nodes of the GP program. The new method can automatically learn various informative image features based on the new function set and terminal set for effective and efficient image classification. Furthermore, instead of relying on a predefined classification algorithm, the proposed approach can automatically select a suitable classification algorithm based on the learned features and conduct classification simultaneously in a single evolved GP program for an image classification task. The experimental results on 12 benchmark datasets of varying difficulty suggest that the new approach achieves better performance than many state-of-the-art methods. Further analysis demonstrates the effectiveness and efficiency of the flexible feature reuse in the proposed approach. The analysis of evolved GP programs/solutions shows their potentially high interpretability.