Improving license plate recognition via diverse stylistic plate generation

Deep learning-based license plate recognition heavily relies on the quantity and diversity of training images. However, manually collecting diverse license plate images is a highly time-consuming and labor-intensive task. Hence, generating license plate images to supplement training data emerges as a cost-effective solution for recognition enhancement. Previous generation methods treated the style of license plates as discrete features and extracted style features from a referenced image to generate multi-style license plate images. However, these methods are incapable of generating license plates with arbitrary new styles, as the style of the generated images is constrained by the existing images. To solve this problem, we propose to extract style features from randomly sampled vectors in continuous space for diverse stylistic license plate generation. This way, license plates with arbitrary styles can be generated by simply modifying the sampled style vectors. Besides, to provide controllable content information for the generator, we propose an orientation-aware content encoder to extract multi-oriented glyph content features from the input horizontal glyphs. Moreover, a dual-discriminator is proposed to conduct adversarial learning on both color images and image gradients, aiming to mitigate unnatural patterns and ensure realistic generated results. Finally, extensive experiments on multiple public license plate datasets demonstrate that our method can generate realistic and diverse license plates, and the generated images can significantly improve recognition performance by up to 11.5%.