A Deep Learning Model for Classifying Bangladeshi Food and Estimating Calories

Abstract

Accurate food calorie estimation is essential for managing dietary intake, particularly for individuals with diabetes. This project presents an end-to-end system for food classification, weight estimation, and calorie calculation from a single food image without requiring a reference object. The system integrates deep learning techniques including convolutional neural networks (CNNs) for food classification, YOLO and Segment Anything Model (SAM) for food segmentation and MiDaS for depth estimation. The DeshiFoodBD dataset consisting of 19 food classes such as bakorkhani, beguni, and kala bhuna is used for food classification. To evaluate classification performance multiple deep learning models including a basic CNN, EfficientNetB0, ResNet50, MobileNetV2 and InceptionV3 have been compared. The segmented food region is processed using depth maps to estimate its volume which is then converted to weight based on standard nutritional values. Finally, the estimated weight is used to compute calorie content. By leveraging advanced computer vision techniques, this method aims to provide an automated and practical approach for dietary monitoring, enhancing convenience and accuracy in food intake assessment.