Efficient Expiration Date Recognition in Food Packages for Mobile Applications

Abstract

The manuscript introduces an innovative framework for expiration date recognition aimed at improving accessibility for visually impaired individuals. The study underscores the pivotal role of convolutional neural networks (CNNs) in addressing complex challenges, such as variations in typography and image degradation. The system attained an F1-score of 0.9303 for the detection task and an accuracy of 97.06% for the recognition model, with a total inference time of 63 milliseconds on a single GeForce GTX 1080 GPU. A comparative analysis of quantized models—FP32, FP16, and INT8—emphasizes the trade-offs in inference speed, energy efficiency, and accuracy on mobile devices. The experimental results indicate that the FP16 model operating in CPU mode achieves an optimal equilibrium between precision and energy consumption, underscoring its suitability for resource-constrained environments.