Cross Stage Partial Dilated Convolution Network for License Plate Recognition

License plate recognition (LPR) is a crucial task in traffic management, but traditional methods face limitations in accuracy and speed that are mutually constraining. In this paper, we propose an efficient license plate recognition system that achieves high recognition accuracy while ensuring real-time recognition. In order to achieve high detection accuracy while minimizing computational effort in the license plate detection stage, we propose a lightweight cross stage partial dilated convolution (CSPDC) network. Firstly, we propose a lightweight downsampling design that reduces the computational effort of downsampling while retaining important feature information. Secondly, we introduce a lightweight feature extraction network that reduces computational effort and parameter count while maintaining the network’s feature extraction capability. Finally, to prevent a decrease in detection performance after lightweight processing, we propose a cross stage partial dilated block that expands the receptive field of feature extraction to enhance the network’s learning capability. Experimental results on the CCPD dataset demonstrate that our proposed system achieves a tradeoff between computational effort and accuracy, with an ACC of 99.3% and a detection speed of 89 FPS. We further deployed and tested our algorithm on the Huawei M6 tablet, and the test results shows the practical value of our proposed method.