MultiPath ViT OCR: A Lightweight Visual Transformer-based License Plate Optical Character Recognition

Because of the natural conditions of license plate images, the Optical Character Recognition (OCR) of these images is generally a challenging problem. OCR systems are utilized in edge devices with limited computation power. Despite the considerable progress of deep neural networks, state-of-the-art models are not always an excellent solution to this problem. Most models have many parameters, and in practice, they need many resources to train, maintain and implement on edge devices. We propose a lightweight model based on Visual Transformer architecture and achieve competitive results against traditional CRNN models. Due to the lack of a rich and large-scale dataset for Persian license plates, we gathered and annotated 1.3M images of license plates in various natural conditions from different points of view and different cameras. We call this dataset as LicenseNet. Our proposed model achieves 77.25% accuracy against CNN models with 75.18% accuracy and embedded OCR models in cameras with 60.37% accuracy on the LicenseNet test set. Furthermore, we achieved better accuracy with 3.21 times fewer training parameters than previously proposed models.