FasterPlateNet: A Faster Deep Neural Network for License Plate Detection and Recognition

License plate recognition is an important technology for vehicle management. However, the existing methods based on traditional image processing or deep learning have limitations in accuracy and efficiency. In this paper, we propose a novel method for license plate detection and recognition, called FasterPlateNet. FasterPlateNet consists of two stages: detection localization and character recognition. In the detection localization stage, it employs a backbone network with stacked C3_Faster modules and a one-stage object detection algorithm with a multi-scale feature fusion pyramid to extract multi-scale features of the license plate, and enhances the resolution and global feature dependency of the feature map by using the content-aware feature reassembly (CARAFE) algorithm and the exposed visual center (EVC) module. In the character recognition stage, it adopts a character recognition network with a convolutional neural network and a connectionist temporal classification loss function, which can directly output the character sequence from the license plate image without character segmentation. The experimental results on the CPPD dataset and its subsets demonstrate that our method achieves 98.8% and 98.5% accuracy in detection localization and recognition respectively. This method has great potential and value in practical applications.