Abstract:
License plates provide the unique signature used for vehicle identification
worldwide. With pervasiveness of cameras, accurate recognition of plates
has becoming an increasingly popular machine-vision problem and has
gained a lot of attention due to its wide applications. Contemporary license
plate recognition systems work well in controlled environments, with
images of good character visibility, but they are not robust and do not
perform well in real world scenarios. This work presents a deep neural
network approach towards automatic license plate recognition where we
formulate it as an object detection problem i.e. each character is treated
as an object. The proposed approach is robust and performs under various
vision challenges (occlusion, light changes, size variation) owing to both
architectural design and the problem formulation of targeting generic
printed character recognition. Our method is evaluated on both publicly
available, academic and commercial datasets. Furthermore, we also release
the largest dataset of license plates for the research community. The
proposed ALPR system competitively outperforms the state-of-the-art
systems with outstanding recognition results on all license plate datasets.
