<p>Driven by economic development, the number of cars continues to grow. The variety and complexity of car license plates in China have brought great pressure to traffic control. A region-based faster convolutional neural network is built to address the urgent need for optimization in license plate positioning systems. The basic network for feature extraction is replaced by a residual network, and a license plate recognition system is designed by combining convolutional block attention mechanism and open-source instruction set architecture technology. The improved method quickly converged the model, with recognition accuracy of 97.3% and 92.6% respectively in the presence of ice and snow and tilt in the images. The recognition system showed that in complex situations such as obstructed or tilted license plate images, the average recognition accuracy was 91.8% and 79%, respectively. Compared with other recognition systems, with a recognition accuracy of 98.25%, it only took 2.03&#xa0;s. Overall, the designed system has feasibility and good performance in license plate localization and recognition, effectively solving the low recognition efficiency and speed of traditional license plate recognition systems.</p>