Research and Implementation of a License Plate Detection Algorithm for Complex Environments

To address the problems of distant recognition size discrepancy and unclear recognition due to light intensity transformation in license plate detection under high level camera, an improved YOLOv5s license plate detection optimization algorithm is proposed. The optimization algorithm is improved based on YOLOv5s:MobileNetV2 is introduced as the feature extraction network to make the model more lightweight; the CBAM module is added to introduce the attention mechanism to make the model pay more attention to the license plate area; at the same time, the regression loss of four vertices is added to improve the original loss function to improve the network extraction ability and speed up the extraction speed to further improve the detection ability of license plate characters . Three sets of experiments are observed, including the comparison of the original YOLOv5s and the optimized network model, the comparison of the optimized model with and without the attention mechanism, the current mainstream algorithms and the performance of the optimized model on CCPD. The experimental results show that the optimized model achieves 96.58% mAP on CCPD dataset and the number of parameters of the model is only 6.71M, and the overall performance is also better than that of Faster Region-CNN and SSD, which can better meet the accuracy and real-time requirements of license plate detection.