<p>Genetic programming (GP) approaches have shown considerable potential in image classification. However, the inherent complexity of deep and wide tree structures in existing GP methods often leads to issues such as overfitting and redundancy. Additionally, most of the traditional GP approaches focus on feature extraction from either the original image or its local regions, overlooking the multi-scale image information that could improve classification performance. To this end, a flattened-tree GP approach with multi-scale feature extraction for image classification is proposed. Specifically, we propose a flattened-tree GP program structure incorporating specialized functions and terminal sets. This structure allows for an efficient combination of features while minimizing redundancy and enhancing feature learning quality. By avoiding the complexity that may be associated with deep and wide tree structures, the proposed approach reduces overfitting and improves model performance. Then, a novel region resampling layer that enables multi-scale feature extraction is suggested in the proposed GP program structure. This layer allows GP to capture global and local features on different scales, enhancing the model’s ability to learn more discriminative feature representations. By integrating multi-scale information, the model could better adapt to varying patterns within images. Finally, we validate the proposed approach on eight diverse datasets compared with the state-of-the-arts. The experimental results show the superiority of our proposed approach in image classification, achieving higher accuracy and better generalization. Visual analysis of optimal GP individuals further highlights the interpretability and effectiveness of learned characteristics with our proposed approach.</p>