Automatic License Plate Recognition for Indian Roads Using Faster-RCNN

Use of deep learning in Automatic License Plate Recognition has been a topic of common research, in the past few years. Recognition of Indian plates have always been a challenge due to their aberrant nature with different fonts, size of letters, padding and number of lines. In our work, we propose a fully functional, end to end solution for license plate recognition in India, considering all irregularities. Our system uses a series of the state-of-the-art Faster Regional Convolutional Neural Network to create a pipeline that gives an efficient solution to the Indian situation under various scenarios. Since there is no publicly available dataset for Indian License plates, we made a balanced dataset using frames from videos and photos from handheld devices, taking into consideration all the irregularities. Our pipeline produced an overall 88.5% total correctness and 10% partial correctness (greater than 5 characters correct) for Indian plates. Inclusion of a novel heuristics system increased total correctness to 91%. License plate detection had a precision of 94.98% for all types of vehicles. Our pipeline successfully segments over 99% of characters from license plates with a mean average precision of 99.55% and was able to correctly recognise 98.6% of the segmented characters.