Automatically Extracting Features Using Genetic Programming for Low-Quality Fish Image Classification

Fish image classification is an important task in the protection of precious marine resources. However, this task is difficult due to the low-quality images and the high inter-class variations across images. Most existing methods use high-quality images for classification and need domain knowledge. In this paper, we develop a genetic programming (GP) approach to automatically selecting image operators to deal with the low-quality images and extracting effective features from these images for low-quality fish image classification. To achieve this, a new program structure and a new function set are developed. With these designs, the proposed GP approach can evolve solutions that use effective filtering or restoration operators to deal with the input image, select informative regions from the fish image, and extract effective global and/or local features from the fish images. The results show that the proposed approach achieves significantly better performance than 12 benchmark methods, including a state-of-the-art GP approach, on the well-known fish image classification dataset. Further analysis shows the high interpretability of the evolved GP trees and the effectiveness of the employed image filtering or restoration operators.