Flexible Region Detection-based Genetic Programming for Fish Classification With Low-Quality Images

Fish image classification is a very important part of intelligent aquaculture development. However, fish image classification is challenging due to large intra-class variation and high inter-class similarity, especially when the quality of the images is low. To solve these challenges, a flexible region detection-based genetic programming approach with a new terminal set, FGPN, is proposed for fish classification tasks with low-quality images. The proposed approach FGPN could automatically choose suitable terminals, detect key regions, extract multiple types of feature, as well as combine global features and local features to address the classification tasks. Compared with seven benchmark methods, including two GP-based methods and five CNN-based methods, the proposed approach FGPN achieves significantly better performance in most comparisons on three real-world datasets.