Academy of Interdisciplinary Studies Division of Integrative Systems and Design 213 Division of Integrative Systems and Design Deep Learning for Food Nutrition Supervisor: LI Mitch / ISD Student: JAMALBEKOV Sultan / MGMT Course: UROP 1000, Summer This document is a progress report on the “Deep Learning for Food Nutrition” UROP1000 project. Preparation started by the end of the spring semester. The objective was to create a mobile application that can estimate calories and composition of food with a camera. Topics relevant to computer vision were studied: fundamentals of supervised machine learning, OpenCV-python API, the underlying concepts of image formation and image processing, and a new programming language (Swift) to create a mobile app for iPhone. The task was more difficult than expected, so the work on making a functional app for food volume estimation is continued. Deep Learning for Food Nutrition Supervisor: LI Mitch / ISD Student: YAO Ruixin / SBM Course: UROP 1000, Summer This thesis explores the efficacy of various machine learning models for food image recognition, utilizing the Food101 dataset as a benchmark for performance evaluation. I evaluate several state-of-the-art models, including Convolutional Neural Networks (CNNs) and Deep Neural Networks (DNN), to identify their strengths and weaknesses in recognizing a diverse range of food items. The results demonstrate that the choice of model architecture and training strategies can substantially influence recognition accuracy, paving the way for better nutrition recommendations in future development and application. Generative AI in 3D Food Printing Supervisor: LI Mitch / ISD Student: CHENG Ziheng / ISD Course: UROP 1000, Summer This report introduces a deep learning method of food image segmentation for nutritional analysis of daily food photos. This method uses an improved U-Net architecture to divide food images into different parts to accurate nutritional estimates. The model was trained using the FoodSeg103 dataset consisting of food images and masks with notes. The result of this phase is that after 200 periods, the pixel accuracy of the training set is 0.947, the loss value is 0.187, and the F1 score is 0.948. The method has demonstrated good prospects in food nutrition analysis and accuracy and can be used in applications and research in the fields of food science and nutrition.
RkJQdWJsaXNoZXIy NDk5Njg=