Abstract
License plate recognition technology is widely applied traffic management, parking monitoring, and electronic toll collection, among other fields. However, in complex scenarios, such as bright light, fog, rain, snow, and nighttime, there is an urgent need to improve the accuracy of license plate recognition and system robustness. To cope with the difficult problem of license plate recognition in complex scenarios, this study proposes a license plate recognition method based on CSCM-YOLOv8 and CSM-LPRNet. The CPA-Enhancer preprocessing module is used to optimize the input feature representation, and the upsampling quality is improved by the perceptual feature reorganization capability of the CARAFE upsampling module. The SEAM is embedded for adaptive weight allocation, thus enhancing the capability to extract key features. The SEAM is combined with the lightweight C2fMLLABlock convolution module to efficiently aggregate features, thereby maintaining the feature representation capability while reducing the computational cost. The experimental results show that on the dataset used in this study, the CSCM-YOLOv8 network achieves 98.9% accuracy in license plate detection, whereas mAP@0.50-0.95 reaches 58.0%. Compared with the original YOLOv8, the accuracy and mAP@0.50-0.95 are improved by 3.1% and 3.9%, respectively. Moreover, CSM-LPRNet achieves a recognition accuracy of 98.56% in character recognition, which is a 7.0% improvement over that of the original LPRNet. The remarkable performance of this method in complex environments provides an efficient and reliable solution for license plate recognition in intelligent transportation systems.