Abstract: To solve the problem that the rectangular detection box outputted from a single deep convolutional neural network (CNN) cannot deal with the non-front license plate location well under unconstrained scenarios, a solution based on cascaded CNNs is proposed. The solution cascades the target detection and target classification CNN network, uses the detection network to obtain the region of interest, and then uses the lightweight classification network to transform the license plate vertex detection problem into a regression problem. Firstly, the YOLOv3 network is used for locating license plates roughly to obtain all the license plate candidate areas in an image. Then, an improved MobileNetV3 lightweight CNN is adopted to locate the license plate vertices in the candidate areas to achieve the precise location of license plate areas. Finally, the license plate is corrected and projected into the rectangular box through perspective transformation. Experiment results show that the proposed cascaded CNNs can effectively solve the problem that a single CNN object detection network can only output a rectangular detection box and is not suitable for unconstrained license plate detection. It has a good application potential. 

Key words: unconstrained license plate location, convolutional neural network, YOLOv3, MobileNetV3