Classification of Commercial Beverages Using Convolutional Neural Networks and Hyperparameter Optimization

Image recognition on edge devices demands high precision and low computational cost. The objective of this work is to propose a new approach for commercial beverage classification using convolutional neural networks and hyperparameter optimization (HPO). The methodology comprised three main steps: structuring a hybrid dataset of Brazilian products; systematically exploring input resolutions and width multipliers in MobileNetV1 and V2 architectures; and evaluating the impact of Int8 quantization. The results indicate that the optimal configuration (MobileNetV2 160x160) achieved an 87,08\% accuracy and high architectural resilience. The optimized model ensures computational lightness without significant predictive degradation, enabling future real-time systems for glycemic control in mobile health (mHealth).