Deep Learning Techniques for Enhanced Automatic Number Plate Recognition

ANPR (Automatic Number Plate Recognition) is a key technology in providing vehicle recognition for applications such as law enforcement, toll gate, and traffic management. Traditional ANPR approaches depend on image processing and OCR and have difficulty coping with blurry images, occlusions, and various plate formats. In order to overcome the issues faced in conventional systems, this paper presents a state-of-the-art automatic number plate recognition (ANPR) system, employing Efficient Net as the backbone for the license plate detector and a transformerbased OCR model. Unlike existing works, our method adapts Efficient Net with bounding box regression for improved localization and uses a custom domain-constrained Transformer tailored for fixed-length alphanumeric recognition. Its building blocks are attention mechanisms and extensive data augmentation for generalization. Our method obtains 92 % accuracy, which is the state-of-the-art, as it exceeds existing autoregressive models by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3-5 \%$</tex>. This solves challenges such as overfitting on small datasets and lack of adaptability to different conditions. Designed to be highly scalable and reliable, it has been effectively trained on extensive and diverse data sets before being implemented.