A Model for Classification of Traffic Signs Using Improved Convolutional Neural Network and Image Enhancement

In an advanced driver assistance system (ADAS), recognition of traffic signs is very important for safety driving. Recently, the convolutional neural networks (CNNs) have presented promising results. In this work, we propose a robust model based on VGG network by adding batch normalization operation. Dropout is also used to reduce the overfitting of the model. Due to the imbalance of the dataset, data augmentation is performed. Then, in order to enhance images, Contrast limited adaptive histogram equalization (CLAHE) and normalization are performed. The performance of the model is evaluated on German traffic sign recognition benchmark (GTSRB) dataset using different performance metrics namely confusion matrix, precision, recall. Experiments results show that, the proposed model reaches a state-of-art accuracy of 99.33 % and surpasses the best human performance of 98.84 %. This model can be used for real world system.