Pakistani traffic-sign recognition using transfer learning

Initially, the traffic-sign recognition was done using the conventional image processing techniques which are sluggish and can cause fatal delays in real-world implementations. Majority of the state-of-the-art detectors are based on a Convolutional Neural Network (CNN) which is a de-facto leader in computer vision research over the past decade. Easy availability of datasets is the main reason for the interest of researchers in CNNs. These datasets are needed to be organized and maintained as the CNN requires colossal amounts of data to work well. Unfortunately, no traffic-sign dataset exists in Pakistan to enable any detection based on the CNN. Therefore, in our work, we have collected and annotated a dataset to help foray into this research area. We propose an approach revolving around the deep learning where a model is pre-trained on the German traffic-sign dataset. This model is then fine-tuned using the Pakistani dataset (of 359 different images) collected across Pakistan. Preprocessing and regularization are used to improve the overall performance of the model. Through results, we show that our fine-tuned model reaches to a training accuracy of nearly 55% outperforming the other related techniques. The results are encouraging as we have achieved a high accuracy keeping in mind the small size of the available Pakistani dataset.