A new genetic programming approach to fine-grained flower image classification

Fine-grained flower image classification (FGFIC) is a challenging task in computer vision due to intra-class variations and inter-class similarities. This paper proposes a new genetic programming-based approach to FGFIC, including a new program structure, a new function set, and a new terminal set. The proposed approach could automatically enhance the images to highlight the flowers, localize the flower and detect discriminative flower regions, and effectively extract and combine the flower's global, local, and/or color features for classification. The regions detected by this method are based on the flowers in each image and thus contain discriminative information about flowers. The RGB image input improves the performance of effective region detection and feature extraction. The experimental results on datasets with different numbers of classes and varying difficulty show that the proposed approach has achieved significantly better performance in most comparisons. Further analysis demonstrates the potential high interpretability of the evolved programs, and improvements in computational cost and searching efficiency compared with other Genetic Programming methods.