Convolutional Neural Networks (CNN) are currently being implemented in a wide variety of applications. This subdomain of Artificial Intelligence shows a powerful performance in machine vision applications and may be used to categorise and classify objects, amongst other image processing tasks.
In the Artificial Perception Group of the Centre for Automation and Robotics (CAR) we are interested in identifying and classifying weed species within crop fields, which is a very specific problem, as the system will only need to process images of soil and plants.
The state-of-the-art CNN architectures, unfortunately, are often very large, and require a long time to be properly trained for their specific task. We believe that application specific architectures could reduce the size of the networks drastically, as well as improve the accuracy in the required task. However, designing new architectures demands a lot of time and expertise.
Using Artemisa, we are evaluating and confirming that an evolutionary strategy can find the best possible architecture for a given classification task. In other words, given a fixed dataset, a Genetic Algorithm (GA) is used to search a wide variety of possible architectures to find which one performs best. When balancing both the performance of the CNN, as well as its size, this approach results in the smallest possible network for the given task.