Designing Convolutional Neural Networks Using Neuroevolution for Traffic Sign Datasets

Image recognition systems are critical components in numerous applications, often requiring real-time implementations that are both fast and accurate. Convolutional Neural Networks (CNNs) are an emerging tool used to meet these conditions. However, in image recognition, CNNs are often designed to fit general image datasets leading to implementations that may have more layers and nodes than are warranted in particular applications. In this paper, a neuroevolution algorithm is developed to reduce a CNN architecture’s complexity by determining the minimal CNN structure and hyperparameters needed to fit a traffic sign dataset. A neuroevolution algorithm is employed to tune the CNN’s parameters and topology to enable a more efficient parameter space search. Results show that despite reducing complexity, the system still maintains high recognition accuracy compared to popular CNNs, such as AlexNet, VGGNet 16, VGGNet 19, GoogleNet, ResNet 50, and ResNet 101.