A hybrid neural network and genetic programming approach to the automatic construction of computer vision systems

Both genetic programming and neural networks are machine learning techniques that have had a wide range of success in the world of computer vision. Recently, neural networks have been able to achieve excellent results on problems that even just ten years ago would have been considered intractable, especially in the area of image classification. Additionally, genetic programming has been shown capable of evolving computer vision programs that are capable of classifying objects in images using conventional computer vision operators. While genetic algorithms have been used to evolve neural network structures and tune the hyperparameters of said networks, this thesis explores an alternative combination of these two techniques. The author asks if integrating trained neural networks with genetic programming, by framing said networks as components for a computer vision system evolver, would increase the final classification accuracy of the evolved classifier. The author also asks that if so, can such a system learn to assemble multiple simple neural networks to solve a complex problem. No claims are made to having discovered a new state of the art method for classification. Instead, the main focus of this research was to learn if it is possible to combine these two techniques in this manner. The results presented from this research indicate that such a combination does improve accuracy compared to a vision system evolved without the use of these networks.