Learning Features for Data-Efficient Image Classification Via Improved Decomposition-Based Multi-Objective Genetic Programming

Data-efficient image classification is critical in computer vision with applications across various domains. Although deep convolutional neural networks are successful for image classification, they often require large datasets and high computational resources, which are unsuitable for data-efficient classification tasks. Genetic programming (GP), on the other hand, offers an interpretable, flexible, and efficient alternative to learning features for image classification, particularly when dealing with insufficient training instances. Many multi-objective GP methods control model bloat by limiting tree size or reducing the number of features, but they rarely include objectives specifically aimed at improving generalization. As a result, the evolved models can still overfit the training data. To address these challenges, we propose an improved decomposition-based multi-objective genetic programming (IDMOGP) approach to feature learning in data-efficient image classification. IDMOGP maximizes the classification accuracy with a regularization term and simultaneously minimizes the number of learned features. To reduce overfitting in data-efficient image classification, a novel regularization method is proposed based on Rademacher complexity to improve generalization. In addition, an adaptive global replacement strategy is designed to balance convergence and diversity during evolution. IDMOGP provides a set of trade-off solutions to data-efficient image classification. Experimental results on five different datasets show that IDMOGP achieves superior hypervolume compared to traditional dominance-based multi-objective GP, and higher classification accuracy than non-GP methods. Further analysis demonstrates the effectiveness of the introduced strategies and shows the good interpretability of the IDMOGP approach.