Co-Evolutionary Dual-Tree Genetic Programming for Fine-Grained Flower Image Classification

Fine-grained flower image classification (FGFIC) is challenging due to high similarities among species and variations within species, especially with limited training data. Existing genetic programming (GP) methods address this task using a single GP tree but show limited performance in flower localization. To overcome this limitation, this paper introduces Co-evolutionary Dual Tree Genetic Programming (CEDTGP) for FGFIC, employing a dual-tree structure that separates flower localization from feature learning for classification, thereby improving both localization and classification accuracy. A novel co-evolutionary strategy further optimizes the interaction between the dual trees, enhancing overall performance. Furthermore, a new distance measurement is proposed to refine the assessment of individuals with improved classification performance. Evaluations on various flower datasets with different numbers of classes demonstrate that CEDTGP outperforms popular GP and convolutional neural network (CNN) methods. Further analysis highlights the effectiveness of the distance measurement and the co-evolutionary strategy, and it also demonstrates the precision of the regions detected by the proposed approach.