Deep learning approach to automatically recognise license number plates

Automatic Number Plate Recognition (ANPR) has become an important aspect in our daily life because of unlimited increase of vehicles and transportation system. This makes it more and more difficult to fully manage and monitor by humans. Due to the diversity of license plates formats, varying scales and sizes, different angles, illuminations, this is quite a challenging problem in the area of computer vision. In this paper, we have proposed methods for automatic detection of a license plate from an image, which is followed by plate correction, or in other words, plate rectification. Thereafter, character recognition methodology has been applied to identify characters from the number plate. Convolutional Neural Networks (CNNs) based approach is used to locate corner points of license plate image, after that plate rectification done using perspective transformation. The CNNs models for locating corner points are neural networks for regression. Here, the loss function is based on an average sum of Euclidean distance between predicted corner points and actual corner points, the loss function is also known as the mean squared error function. The results show that our CNNs models are able to accurately predict corner points from number plate. Furthermore, an optical character recognition (OCR) model is used to identify characters from the plate. The developed methodology shows excellent results on the Chinese City Parking Dataset (CCPD).