Traffic Sign Recognition with Binarized Multi-Scale Neural Networks

Computer vision has grown tremendously thanks to advances in convolutional neural networks (CNNs), the field of traffic sign recognition and intelligent transportation performs better because of that. However, as traffic sign recognition systems are often deployed in computing and storage constrained environments, it is critical for the network design to be accurate and real-time in these instances. Whether it is training or running an ordinary neural networks framework, it requires one or many high-power Graphic Processing Units (GPUs) and considerable memory resources. To solve this problem, this project creatively propose a 2-stage binarized multi-scale neural network framework, called B-MNN, extends Pierre Sermanet and Yann LeCun's Convolutional Networks. When training a B-MNN, we intentionally restrict the weights and activations of the model to +1 or -1. The mathematical calculation (addition, subtraction, multiplication and division) of the binarization network can be implemented using bit-wise bit operations, which is convenient for hardware to implement the algorithm, faster and more energy efficient. B-MNN approaches state-of-the-art results on the GTSRB traffic sign dataset implemented with the Keras artificial neural network library. We achieved 91.34% accuracy. Although it is slightly less accurate than models based on CNNs, the same ordinary neural network structure is an order of magnitude larger than ours. It is an important direction to realize the neural network into the realistic scene in the future.