Persian Text Based Traffic sign Detection with Convolutional Neural Network: A New Dataset

Recently, traffic panel detection has attracted both academic and industrial attention. However, there are a few works that studied text based traffic panels. This is because there are many challenges in this kind of traffic panels. To obtain an appropriate accuracy in text recognition in the text based traffic panels, we need to detect the panel. Since there is no public text based traffic panels dataset, we collected a new dataset included the Persian text based traffic panels in the streets of Tehran-Iran for the first time. Our dataset contains two sets of figures. The first set has 9294 pictures and the second set has 3305 pictures. The second dataset is more uniform than the first dataset. Therefore, we exploit the first set as an additional dataset and use the second one as the main dataset. Accordingly, we pretrain the network by the additional dataset and train it by the main dataset. We use the tiny YOLOv3 (You Only Look Once version three) algorithm to pretrain, train, and test the dataset. The algorithm is fast and has low complexity. We use K-fold cross validation method to appraise efficiency of the algorithm. From the results section we could see that Precision is 0.973, Recall is 0.945, and Fmeasure is 0.955.