Enhancing Bus Number Detection Efficiency with Transfer Learning and Fine-Tuned Optical Character Recognition

This research investigates two crucial aspects of computer vision and object recognition, with implications for real-world applications, specifically in bus detection and bus number recognition. The first aspect focuses on enhancing the efficiency and accuracy of bus detection using Transfer Learning with Frozen Layers (TLFL), particularly emphasizing the YOLOv5n model. TLFL selectively adapts a pre-trained neural network, preserving low-level features while fine-tuning domain-specific components, significantly improving training speed and performance. It outperforms the baseline YOLOv5n model with fewer training epochs, highlighting the importance of balancing existing knowledge and domain-specific adaptation. The second aspect explores Optical Character Recognition (OCR) models for bus number recognition. Customization and fine-tuning, as seen in the transformation of Paddle OCR to Fine-Tuning Paddle OCR, substantially increase prediction confidence and reduce Character Error Rate (CER). This fine-tuning process enhances OCR adaptability and precision, especially in challenging scenarios like accurate bus number recognition. A comparative analysis of OCR models shows the superiority of Fine-Tuning Paddle OCR, offering a balance between accuracy and efficiency, while other models present varying trade-offs between precision and speed.