Single-stage automatic license plate recognition with contrastive denoising learning

Automatic License Plate Recognition (ALPR) aims to extract license plate strings directly from full vehicle images. Traditional fully single-stage ALPR methods typically rely on exhaustively annotated datasets with character-level bounding boxes, framing recognition as an object detection task. In contrast, we propose CiT2, a cropping-free, end-to-end ALPR approach based on a transformer framework that requires only license plate bounding boxes and string annotations. CiT2 jointly predicts both the bounding box and the license plate string, integrating four key components: auxiliary prediction, one-to-one matching, joint bounding box and string prediction, and a contrastive denoising (CDN) learning task. Following DETR, auxiliary prediction and one-to-one matching utilize fixed-length positional queries and the Hungarian algorithm to align predictions with ground truth. By grouping multiple tokens into a single prediction, the model effectively predicts both the bounding box and the license plate string simultaneously. Informed by CDN, we inject noise into both bounding box and string labels during training, enabling the model to learn denoising as an auxiliary task. Additionally, a Spatial Transformer Network (STN) is seamlessly integrated into the architecture via interpolated positional encodings to improve accuracy while maintaining efficiency. CiT2 achieves 97.8% accuracy on CCPD2018, 68.8% on CCPD2020, and 84.0% on CLPD. The implementation is publicly available at https://github.com/gb16001/CiT2.