A deep reinforcement learning approach to character segmentation of license plate images

Automated license plate recognition (ALPR) has been applied to identify vehicles by their license plates and is critical in several important transportation applications. In order to achieve the recognition accuracy levels typically required in the market, it is necessary to obtain properly segmented characters. A standard method, projection-based segmentation, is challenged by substantial variation across the plate in the regions surrounding the characters. In this paper a reinforcement learning (RL) method is adapted to create a segmentation agent that can find appropriate segmentation paths that avoid characters, traversing from the top to the bottom of a cropped license plate image. Then a hybrid approach is proposed, leveraging the speed and simplicity of the projection-based segmentation technique along with the power of the RL method. The results of our experiments show significant improvement over the histogram projection currently used for character segmentation.