Effective deep neural networks for license plate detection and recognition

Recently, automatic license plate recognition (ALPR) has drawn much of attention from researchers due to the impressive performance of deep learning (DL) techniques. While a large number of methods for ALPR have been investigated, there are a few attempts emphasizing efficient yet accurate models for facilitating the deployment on traditional CPU boxes or low resource devices. In the present work, we propose lightweight and effective deep convolutional neuron networks to address the problems of license plate detection and recognition. Differing from traditional DL methods, the proposed models discard the use of max-pooling modules, belong to a single-phase object detector, and consist of alternating convolutional layers and Inception residual networks. Different strategies for character prediction are also studied and deeply discussed. To have more insight of convolutional features, we provide a sequence of visual snapshots or slices of feature maps learned by the network at different layers for a given input image. These data are useful to explain the model's behavior for the task of character recognition. Extensive experiments conducted on two public datasets, CCPD and AOLP, showed a promising improvement of LP detection and recognition accuracy. Interestingly, the full system can work on low-resource CPU machines with a real-time speed.