AHCMM-Net-NGO: A Progressive Hybrid CNN–Mamba Framework with Adaptive Attention Refinement for Robust License Plate Recognition

Automatic License Plate Recognition (LPR) is an essential technology in modern intelligent transportation systems, facilitating the identification of vehicles without manual intervention. It supports a wide range of applications, including traffic monitoring, electronic toll payment, parking automation, secure access management, and law enforcement operations, thereby improving transportation efficiency, security, and operational effectiveness. However, real-world license plate images are frequently affected by motion blur, noise, illumination variations, adverse weather conditions, and low resolution, which significantly degrade recognition performance. To address these challenges, this study proposes an Adaptive Hybrid CNN–Mamba–Multi-Head Attention Network with Northern Goshawk Optimization (AHCMM-Net-NGO) for robust LPR. The proposed framework combines license plate detection, progressive image restoration and enhancement, hierarchical multi-scale feature learning, efficient contextual modeling using Vision Mamba, adaptive attention refinement, and automatic hyperparameter optimization within a unified end-to-end architecture. The framework was evaluated on the UFPR-ALPR dataset containing 4500 fully annotated vehicle images captured under real-world driving conditions. Experimental outcomes demonstrate superior recognition performance, achieving 98.96% accuracy, 98.89% precision, 98.81% recall, and a 98.85% F1-score. Comprehensive experimental evaluations, including ablation studies, hyperparameter sensitivity analysis, cross-validation, and comparative performance analysis, further demonstrate the efficacy, robustness, and generalization capability of the proposed framework. Overall, the proposed AHCMM-Net-NGO framework provides an accurate, reliable, and robust solution for license plate recognition under challenging imaging conditions and demonstrates strong potential for intelligent transportation systems, although practical deployment should consider the computational requirements of the integrated deep learning framework.