인조 데이터를 활용한 차량 번호판 인식 시스템

In deep learning, training data can determine the performance of the network and increase the performance of the network from a diverse training data. However, gathering and labeling training data requires significant time and personnel. Likewise, various types of license plates and text labeling are required for deep learning based license plate character recognition. In this paper, we developed an synthetic license plate data generation program to solve the problem of time-consuming data collection and labeling. The proposed license plate recognition system is an image taken in a drone environment, which makes it difficult to recognize the license plate character. In order to recognize the license plate character in the adverse environment, we applied the methods such as scale transform, brightness conversion, rotation, background addition, noise and blur addition to the synthetic license plate to generate the training data in the harsh environment. To evaluate the training results of the synthetic license plate data, we collected 1,000 real license plates and constructed a test set. As a result, the total character recognition accuracy was 85% and the individual character recognition accuracy was 94% using 1.5M synthetic license plate data without the real license plate data. After that, 88, 96% accuracy was recorded through the post-processing algorithm.