DBNet with Integrated CBAM: OCR Recognition Optimization for Complex Food Packaging

Accurate OCR for expiration dates on food packaging is crucial for quality control and consumer safety. However, challenges such as dot-matrix printing, complex backgrounds, and printing defects hinder recognition accuracy. This paper proposes an enhanced OCR framework integrating DBNet with CBAM for text detection and a ResNet-based model with CTC loss for text recognition. The CBAM module improves feature extraction, while the fully convolutional recognition model enhances accuracy and efficiency. Experiments on Unitail and a custom expiration date dataset show that DBNet + CBAM achieves an F1 score of 91.05, outperforming TextSnake, SwinTextSpotter and EAST. The ResNet + CBAM + CTC model reaches about 98 % accuracy, reducing inference time by over $50 \%$ compared to CRNN. Additionally, deployment on NVIDIA Jetson Nano with TensorRT enables real-time processing with low computational cost. The proposed approach provides a highly accurate and efficient OCR solution for automated expiration date recognition in industrial applications.