Indian Food Segmentation and Calorie Estimation Using CatBoost and Masked Convolutional Neural Networks

Accurate food segmentation and calorie estimation play a pivotal role in effective dietary monitoring and health management. In this paper, after extensive experimentation on a diverse Indian food dataset of 20 classes collected by us, we present two separate innovative models designed to address these tasks with exceptional precision. The first model employs a machine learning approach by incorporating the CatBoost algorithm, while the second model leverages a deep learning technique utilizing the U-Net architecture for high-quality image segmentation. Additionally, we merge the U-Net outcomes with a convolutional neural network (CNN) to enhance the deep learning-based classification. Our models effectively handle categorical features and address imbalanced data, resulting in significantly improved accuracy in food item delineation within images. This advancement enables a more reliable and comprehensive analysis of dietary patterns. We demonstrate the effectiveness and robustness of our proposed models. Comparative evaluations against state-of-the-art methods verify exceptional performance in terms of segmentation accuracy, precision in food classification, and estimation of calorie intake. Notably, our models are specifically tailored for food images and do not take ingredients or any other information into consideration. In comparison CNN model is performs better than CatBoost, achieving an accuracy of 78%.