Reusing Extracted Knowledge in Genetic Programming to Solve Complex Texture Image Classification Problems

Abstract

Transfer learning is a process to transfer knowledge learned in one or more source tasks to a related but more complex, unseen target task, in an effort to facilitate learning in the target task. Genetic programming GP is an evolutionary approach to generating computer programs for solving a given problem automatically. Transfer learning in GP has been investigated in complex Boolean and symbolic regression problems, but not much in image classification. In this paper, we propose a novel approach to use transfer learning in GP for image classification problems. Specifically, the proposed novel approach extends an existing state-of-the-art GP method by incorporating the ability to extract useful knowledge from simpler problems of a domain and reuse the extracted knowledge to solve complex problems of the domain. The proposed system has been compared with the baseline system i.e., GP without using transfer learning on multi-class texture classification problems from three widely-used texture datasets with different rotations and different levels of noise. The experimental results showed that the ability to reuse the extracted knowledge in the proposed GP method helps achieve better classification accuracy than the baseline GP method.