Deep learning is a research field with great application potential. However, training a deep learning model requires a large amount of paired data, which is time-consuming and expensive for real-world applications. In light of this problem, researchers have conducted methods to overcome the lack of training data, including data augmentation, data enhancement, transfer learning, and semi-supervised learning (SSL). Based on this concept, in this study, we conduct a self-training method to fine-tune a YOLOv4 license plate detection framework that requires only a limited amount of training datasets. The framework requires only a small amount of labeled data to complete the training process. Furthermore, this study utilizes Tesseract optical character recognition (OCR) on the detected license plates to achieve better performance. We propose a high-performance license plate recognition system for detecting different tilting angles and complex backgrounds.