NanoCNN: A Parameters Efficient Network for Traffic Sign Recognition

Traffic Sign Recognition (TSR) plays a vital role in intelligent transportation systems for the prevention of road accidents. Recent Convolutional Neural Network (CNN) based TSR studies have exhibited adequate accuracy but these techniques usually require a huge number of training parameters. Therefore, real-time implementation of these CNN networks demands higher computational and memory resources leading to uneconomical solutions. In this study, NanoCNN is proposed for traffic sign recognition. NanoCNN is based on a novel CNN structure that effectively captures global features with parameters efficient layers' arrangement without compromising on classification performance. The proposed TSR network achieves a classification accuracy of 99.18% on the German Traffic Sign Recognition Benchmark (GTSRB) dataset with 0.12 million parameters. NanoCNN outperforms existing benchmark TSR schemes by training 10 times less number of parameters. The proposed TSR framework is the most lightweight as per the best knowledge of the authors. The exceptional performance of NanoCNN with a drastically reduced number of parameters makes it a preferred option for real-time TSR solutions.