A Novel Genetically Optimized Convolutional Neural Network for Traffic Sign Recognition: A New Benchmark on Belgium and Chinese Traffic Sign Datasets

Abstract

Traffic signs are a key constituent of the road network and prove to be very useful for warning and guiding the drivers. In intelligent transport systems, traffic sign recognition (TSR) is indispensable for autonomous driving. However, due to the complex outdoor environment, real-time recognition of traffic signs is much more challenging in comparison with many other pattern recognition tasks. Convolutional neural networks (CNNs) have an exceptional capability of recognizing patterns and are one of the most popular deep learning techniques. Finding the optimal configuration of a CNN for a task is a major challenge and is an active area of research. Genetic algorithm (GA) is a meta-heuristic approach well-known for its optimization power. In this paper, we propose a novel deep learning technique based on the concept of domain transfer learning for the recognition of traffic signs. This technique utilizes a newly proposed variant of the GA for finding the optimal values of the number of epochs and the learning rate parameter for each layer of the pre-trained CNN model (VGG-16). To examine the effectiveness of our technique, we apply it to the following two benchmark datasets of TSR: Belgium Traffic Sign Classification (BTSC) dataset and Chinese Traffic Sign Dataset (TT100K). The results indicate that our model outperforms all the existing approaches applied to these datasets and gives a new benchmark of the recognition accuracies of 99.16% for the BTSC and 96.28% for the TT100K datasets, thus establishing the robustness of our model.