End-to-End Nutritional Label Analysis of Packaged Beverages Using YOLOv8, Optical Character Recognition, and Support Vector Machine for Health Risk Assessment

Abstract

The increasing consumption of packaged beverages high in sugar, sodium, and fat has raised concerns about their contribution to non-communicable diseases. Although nutritional information is available on product packaging, consumers often face difficulties in locating, reading, and interpreting it accurately, especially under varying lighting conditions and on complex package surfaces. This study aims to develop and evaluate an integrated nutritional label analysis system for packaged beverages by combining object detection, OCR-based text extraction, structured feature construction, health status classification, and potential disease risk mapping. The proposed system uses YOLOv8 for label detection, applies preprocessing and postprocessing to improve Tesseract v5 OCR results, constructs structured numerical features, balances training data using SMOTE, normalizes features with the z-score method, and classifies products into safe and unsafe categories using an SVM model. The dataset consisted of 2,414 augmented images derived from 1,207 selected nutritional label images. The results show that YOLOv8 achieved an mAP@0.5 of 0.969 for nutritional label detection. In the OCR stage, preprocessing and postprocessing reduced recognition errors across different image conditions, with WER and CER reaching zero in the ideal- and low-light evaluation samples. The optimized SVM-RBF model achieved a test accuracy of 0.9959 in classifying safe and unsafe products. These findings indicate that the proposed system provides an interpretable computational framework that contributes to computer vision and health informatics by transforming visual nutritional labels into structured health-related information to support nutritional literacy and healthier consumption decisions.