Recipient Organization
OREGON STATE UNIVERSITY
(N/A)
CORVALLIS,OR 97331
Performing Department
(N/A)
Non Technical Summary
In the United States, 30-40% of the food supply is wasted annually. Food waste is partly due to inaccurate food quality and shelf-life assessments. Traditional assessment methods mainly rely on destructive testing or subjective visual inspection. This Seed Grant aims to develop more robust and generalizable machine learning models that will be essential to enhance detection accuracy across diverse applications. The specific aims are: (i) Develop deep learning models to provide a generalized, robust prediction of food shelf life; (ii) Create a user-friendly software application that uses deep learning to assess food remaining shelf life. The expected outcomes include a rapid, non-destructive, and mobile imaging method for shelf-life prediction and a database oflabeled food images. This approach will enable users to make informed decisions about food quality without the need for specialized equipment, thereby reducing food waste and optimizing food utilization.
Animal Health Component
50%
Research Effort Categories
Basic
0%
Applied
50%
Developmental
50%
Goals / Objectives
In the United States, 30-40% of the food supply is wasted annually. Accurately predicting food shelf life is essential for reducing food waste, enabling packinghouses to make data-driven decisions on distribution, on-shelf rotation, and consumption. While current analysis methods provide valuable insights, they are often subjective, costly, time-intensive, and sometimes require laboratory settings, creating obstacles for practical application. The overall goal of this project is to develop a mobile imaging and deep learning-based approach for non-destructive rapid assessment of food shelf-life prediction.
Project Methods
1. Curate a food image database. Unripe, undamaged food samples will be purchased from local supermarkets during their harvest seasons. Samples will be selected based on standard hand firmness scales to confirm they are unripe. Images will be obtained for each selected food type under various imaging conditions, including lighting, background, and camera variations. Firmness will be measured as the indicator of ripeness using a TA.XTplus texture analyzer (Texture Technologies, USA). The firmness will be used as the ground truth to label the images.2. Train and validate deep learning models. We propose a novel multi-domain adaptation framework for food quality assessment. The workflow will include dataset split, model training, and model testing. Images will be randomly shuffled and split into a training set, a validation set, and a test set. Machine learning models will be trained using the PyTorch library in Python. We will optimize training hyperparameters such as initial learning rate, mini-batch size, and epoch numbers. Transfer learning with public databases will also be performed for better model convergence. After the training is completed, the performance of deep learning models will be evaluated using the test set. The evaluation measures will include R2 and RMSE.3. Develop a smartphone application to access food shelf life. We will develop an Android application featuring user-friendly interfaces that facilitate automatic image processing. The existing PyTorch Mobile framework will be used to deploy the deep learning models obtained in this project. The performance will be assessed if the app accurately estimates the remaining shelf life.