Food Image Calorie Estimation Using Deep Learning Based Pre-Trained Image Detection Architecture and Machine Learning Algorithms

Abstract

Food-image calorie estimation is an emerging application of computer vision and machine learning that can support automated dietary assessment. This study examined the use of a pre-trained deep-learning object-detection architecture and machine-learning regression models to estimate food volume and calorie content from food images. The ECUST food image dataset was used because it includes food images with calibration objects and related volume and nutritional information. A Faster R-CNN model with an Inception V2 feature extractor was applied to detect food items and calibration objects. Information extracted from the detection process, including food dimensions, calibration-object dimensions, foreground-pixel measures, density and energy values, was used to construct a dataset for volume and calorie estimation. Three machine-learning models were compared: Decision Tree, Random Forest and Support Vector Machine. The dataset was cleaned, transformed, normalised and divided into training and testing subsets before model fitting and evaluation. The object-detection stage identified food items with confidence scores ranging from 51% to 98% across selected food classes. The regression results showed that the Random Forest model performed best among the three models for both volume and calorie estimation, followed by the Decision Tree model, whereas the Support Vector Machine model showed the weakest overall performance. The Random Forest model achieved a higher correlation with actual calorie values than the Decision Tree and Support Vector Machine models. The findings suggest that combining deep-learning-based food detection with ensemble machine-learning regression can support food-image calorie estimation. However, further validation using stronger detection metrics, larger datasets and real-world food images is needed before practical deployment.