International Research Journal of Engineering and Technology (IRJET) Volume: 12 Issue: 11 | Nov 2025
www.irjet.net
e-ISSN: 2395-0056 p-ISSN: 2395-0072
DeepNutriNet: A Dual-Stream Learning Framework for Automated Food Recognition and Caloric Assessment Prof. Rani Prakash1, Suman Biradar2 1Professor, Master of Computer Application VTU’s CPGS, Kalaburagi, India 2Student, Master of Computer Application VTU’s CPGS, Kalaburagi, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract -Recognizing food and estimating calories is
is still difficult because of intra-class variability, dish similarities, inconsistent portion sizes, and various presentations/styles across dishes from various cuisines [5]. Deep learning approaches, specifically convolutional neural networks (CNNs) have demonstrated great success in image classification tasks; object detection systems, such as YOLO (You Only Look Once) and Faster R-CNN, have become standard for detecting multiple items in a single photograph [6][7]. This paper describes a dualstream framework, DeepNutriNet, that combines detection and classification with a knowledge base that supports nutrition information allowing for automation of food recognition/ Calorie estimation with greater accuracy and practicality of use.
part of fostering healthier habits and avoiding diet-related diseases such as obesity and diabetes. Calorie tracking methods of the past are still prone to error and inconvenient, justifying the push for automated approaches. We introduce DeepNutriNet, a dual-stream deep learning framework for food recognition and calorie estimation from images. The framework employs a CNN-based detection model that identifies a set of food items that may or may not be overlapping with a classification stream for resolving classification uncertainty caused by visual similarity. We leverage a nutrition knowledge base as a means for converting the end product of recognition into an estimated calorie total at each item and at the meal level. Experiments conducted on publicly available food image datasets with images featuring varying numbers of food items demonstrated high recognition accuracy and favorable calorie estimates. Collectively, these results speak to the ability of this framework to be a viable means of dietary monitoring and precision nutrition intervention.
2. PROBLEM STATEMENT Accurate dietary intake monitoring is still a poor area of healthcare and nutrition management. Conventional approaches using manual food diaries and apps often suffer from user-specific bias, underreporting, and incorrect estimates that lead to faulty calorie information. Image-based food recognition has some good ideas, but it faces many practical challenges, including variation across food types, overlapping items, inconsistent portion size, and visual similarities between dishes. While deep learning approaches offer strong performance for image recognition, they can often take either a detection approach or classification approach in isolation and limit their application to spectrum of daily use, including potentially simultaneous detection and calorie estimation.
Key Words: Food recognition, Calorie estimation, Deep learning, CNN, Dual-stream framework, Nutritional analysis, Precision nutrition, Dietary monitoring.
1. INTRODUCTION Food is essential to human survival and well-being, as it provides the energy and nutrients necessary to support physical performance and mental well-being. However, modern dietary habits that involve eating high calorie and processed foods have contributed to the global epidemic of diseases related to lifestyle factors including obesity, cardiovascular disease, and type-2 diabetes [1]. According to the World Health Organization, poor dietary habits are still one of the main risk factors for premature death worldwide. [2]. As a result, there has been an increase interest in intelligent systems that can track food consumption, analyze nutritional value, and promote healthier eating habits. Traditional methods for tracking calories, pocket food diaries, nutritional databases, or recording portion sizes, are often tedious, inaccurate, and dependent upon user compliance. [3].These limitations highlight the need for automated, reliable, and nonintrusive solutions. Through advances in computer vision and deep learning, it is possible to accurately identify foods from images which provides the opportunity for dietary monitoring applications [4]. Still, food recognition
© 2025, IRJET
|
Impact Factor value: 8.315
3. OBJECTIVES The main goal of this work is to build an intelligent structured framework for automatic food recognition and caloric estimation using deep learning techniques. The framework will use a YOLO-based object detection model to localize and differentiate multiple food items and a CNN-based classification model, including transfer learning, for more accurate recognition. A pizza foodimage dataset from on-line Kaggle will be used for training and validation that will transfer well to food with a somewhat wider variety of cuisines and portion sizes. Furthermore, the study's purpose is to build a userfriendly application that allows the user to take a picture of food, detect food in real-time, obtain a caloric estimate,
|
ISO 9001:2008 Certified Journal
|
Page 772