Enhancing Turkish traffic sign recognition: A comparison of training step numbers, lighting conditions and image sizes

As the number of vehicles on the roads increases, traffic signs are becoming more and more important every passing day. Despite the fact that traffic signs are simple and easy to understand, in congested traffic drivers may miss them. It would be a big help if a system could assist the driver with traffic signs. A traffic sign recognition system needs to be implemented to achieve this. Three different steps are proposed in this thesis for the Turkish TSR system. Firstly, as step one, models are trained using selective search method and transfer learning is utilized in model training. Models are developed using the pre-trained MobileNetV2 model. Secondly, as step two, models are trained using the selective search method once more, however this time instead of using the transfer learning, the models are trained by developing a new custom CNN. Lastly, as step three, the Faster R-CNN Inception V2 COCO model is utilized in model training. For training purposes, indigenous dataset is created containing 54 distinct classes and 10842 Turkish traffic sign images. The training process of the models are carried out twice with different training step numbers. Then, these models are used to detect Turkish traffic sign images taken both daytime and nighttime with large size and small size images. The results of step one indicate that the loss value for the model which is trained with 25 epochs is 7%, with a test accuracy of 98.7% and the loss value for the model which is trained with 50 epochs is 4%, with a test accuracy of 99%. The results of step two show that the loss value for the model which is trained with 25 epochs is 8.9%, with a test accuracy of 98.8% and the loss value for the model which is trained with 50 epochs is 9.5%, with a test accuracy of 99.1%. The results of step three demonstrates that the Faster R-CNN model's average precision is 67.2% and average recall is 78.3% when trained with 51,217 steps; on the other hand, the average precision increases to 76% and average recall increases to 82.8% when trained with 200,000 steps. When the total test time for all three steps are compared, the detection and recognition time for the models are approximately 51, 63, and 7 seconds, respectively.