Irregular License Plate Recognition via Global Information Integration

Irregular license plate recognition remains challenging due to the irregular layouts of characters, such as multi-line and perspective-distorted layouts. Many previous methods are based on different attention mechanisms, which generate attention weights and aggregate the features to obtain character features for recognition. However, we found that attention-based methods suffer from attention deviation and character misidentification of similar glyphs. We infer that the lack of global perception is the main reason for these problems. Hence, we propose to integrate sufficient global information into the network to improve irregular license plate recognition. Firstly, we propose the deformable spatial attention module to integrate global layout information into attention calculations, thus generating more accurate attention. Secondly, we propose the global perception module that integrates global visual information into feature extraction to enhance the completeness of character features, making them more representative, and thereby alleviating character misidentification. Experiments demonstrate that our method achieves state-of-the-art results, with a significant improvement of 6.7% on irregular license plates. Our codes are available at https://github.com/MMM2024/GP_LPR .