An Analytical Study of Efficient CNNs Tuning and Scaling for Traffic Signs Recognition

Deep learning-based traffic sign recognition has been a very active area of research in autonomous driving since the appearance of Convolutional Neural Networks (CNN) as a substitute for classical machine learning algorithms. However, a good traffic sign recognition system (TSR) should inclusively fulfill accuracy, and response time compromise to be palatable in self-driving applications. Besides, the considerable computational load remains a burden to the adaptation and the design of CNN architectures for real-time applications. This paper aims to investigate the relationship between accuracy, efficiency, and computational complexity for the classification of traffic signs. MobileNetV2 and EfficientNet architectures were evaluated as they are specifically designed to be computationally efficient. When most of the contributed work in the literature focuses on accuracy, we rather focus on the choice of the most efficient model (best accuracy/model complexity ratio). The results support the intuitive idea that performance remains proportional to network size up to a given level beyond which it saturates.