EfficientNet-B7 for Real-Time Optical Character Recognition in Diverse Text Environments

This paper presents an Optical Character Recognition (OCR) system based on the EfficientNet-B7 model, designed to improve text recognition accuracy and efficiency. Using the OCR dataset, which includes a wide variety of printed and handwritten text images with different fonts, sizes, orientations, and noise levels, the model leverages EfficientNet-B7’s compound scaling to balance computational complexity with performance. By employing transfer learning from ImageNet pre-trained weights, the system accelerates convergence while maintaining high generalization across unseen data. Advanced data augmentation techniques, such as random rotations, brightness adjustments, and cropping, are applied during preprocessing to enhance the model’s robustness. The system achieves a validation accuracy of 93.22%, significantly outperforming traditional OCR methods. Additionally, hyperparameter tuning using metaheuristic optimization further refines the model, ensuring optimal performance in terms of accuracy and loss reduction. The system’s scalability and efficiency make it suitable for real-time OCR applications in industries like healthcare, finance, and education, where precise text recognition is essential. Future research will focus on extending the model to support multi-lingual OCR and incorporating attention mechanisms to handle more complex text layouts and structures.