Modern systems of automatic license plate recognition are mostly based on deep neural networks and trained on image data. If data distribution changes, the systems need re-training. For example, new license plate formats result in changing the distribution. Deep neural networks are known to require a huge amount of data for training. The introduction of a new plate shape and sequence of numbers and letters requires the collecting a new training sample, which is impossible due to the lack of the necessary amount of real world examples. The solution is to generate images of the new license plate format based on the old format images while maintaining photorealism. This allows to effectively train systems for both detection and recognition of new license plates and adapt them to real world data in advance. The paper presents a fully automatic approach of photorealistic generation for the new format of Russian license plates. The approach is a sequential algorithm based on deep neural networks, computer vision, projective geometry, and style transfer techniques. It has been tested for the new format of Russian vehicles license plates. The license plates generated by our approach are detected and recognized with better performance as the corresponding old license plates. The approach shows the generalization to the real world data.
展开▼
