Research on Chinese License Plate Automatic Recognition Based on Transfer Learning

This paper introduces an optimized approach that harnesses the power of transfer learning, specifically the ResNet34 architecture, to transcend the limitations of existing self-developed model in recognizing Chinese license plates. By integrating advanced digital image processing (DIP) techniques with the robustness of the ResNet34 model, that not only refines the accuracy and efficiency of the automatic license plate recognition (ALPR) system but also underscores the pivotal role of DIP in enhancing feature extraction and image quality enhancement. This method leverages a suite of sophisticated, open-source DIP tools, namely the OpenCV framework and Python Imaging Library (PIL), complemented by the versatility of artificial neural networks (ANN). This synergy enables the concurrent generation of training, validation, and test datasets, thereby streamlining the model development pipeline. Innovating further, our deep learning training paradigm operates in parallel, managing data generation, processing, and training simultaneously. This approach not only refines the recognition capabilities of the ALPR system but also optimizes the utilization of computational resources. Fine-tuning the established parameters of the ResNet34 model alongside specific training adjustments has led to a remarkable improvement in accuracy, ranging from 3.12% to 12.50% across the seven characters of the license plate, outperforming our self-developed CNN baseline model. Despite an increase in the model's parameter count from 9,023,463 to 22,919,879, the implementation efficiency has impressively increased by over 80%. The significance of this research lies in its potential to evolve traffic management, security monitoring, and parking systems by providing a more accurate, efficient, and robust solution for Chinese license plate recognition. The integration of DIP with deep learning models, as demonstrated in this study, sets a new resolution for future research and applications in the ALPR domain.