Multitree Genetic Programming for Learning Color and Multiscale Features in Image Classification

Data-efficient image classification, which focuses on achieving accurate classification performance with limited labeled data, has garnered significant attention. Genetic programming (GP) has achieved impressive progress in image classification, particularly in scenarios involving small amounts of labeled data. GP research typically focuses on designing tree-based model representations to learn useful image features for classification. However, most GP methods are proposed for gray-scale images and ignore the color features. Furthermore, the existing GP methods typically learn features on a single scale/resolution, restricting potential accuracy enhancements. To address these issues, this article proposes a new multitree GP representation for image feature learning and classification. In each individual, three trees are included to extract discriminative features from the red, green, and blue channels of the image. With the new image resizing layer in the tree representation, the proposed approach can achieve multiscale feature extraction, i.e., flexibly learning fine-grained details and coarse-grained structures in the image, improving the classification performance. In addition, since a limitation of GP is premature convergence due to a decline in population diversity, this article develops a hybrid parent selection method consisting of tournament and lexicase selection to increase population diversity, find the best individual, and improve classification accuracy. The experiments on six image classification datasets indicate that the proposed approach outperforms state-of-the-art neural network-based and GP-based methods in almost all comparisons. Further analyses demonstrate the effectiveness of each component and the potentially high interpretability of the proposed approach.