Real-Time Automatic License Plate Recognition through Deep Multi-Task Networks

With the increasing number of cameras available in the cities, video traffic analysis can provide useful insights for the transportation segment. One of such analysis is the Automatic License Plate Recognition (ALPR). Previous approaches divided this task into several cascaded subtasks, i.e., vehicle location, license plate detection, character segmentation and optical character recognition. However, since each task has its own accuracy, the error propagation between each subtask is detrimental to the final accuracy. Therefore, focusing on the reduction of error propagation, we propose a technique that is able to perform ALPR using only two deep networks, the first performs license plate detection (LPD) and the second performs license plate recognition (LPR). The latter does not execute explicit character segmentation, which reduces significantly the error propagation. As these deep networks need a large number of samples to converge, we develop new data augmentation techniques that allow them to reach their full potential as well as a new dataset to train and evaluate ALPR approaches. According to experimental results, our approach is able to achieve state-of-the-art results in the SSIG-SegPlate dataset, reaching improvements up to 1.4 percentage point when compared to the best baseline. Furthermore, the approach is also able to perform in real time even in scenarios where many plates are present at the same frame, reaching significantly higher frame rates when compared with previously proposed approaches.