Date of Graduation

5-2026

Document Type

Thesis

Degree Name

Bachelor of Science in Biological Engineering

Degree Level

Undergraduate

Department

Biological and Agricultural Engineering

Advisor/Mentor

Wang, Dongyi

Committee Member

Koparan, Cengiz

Second Committee Member

Howell, Terry Jr.

Abstract

Inconsistent quality grading in beef production leads to inefficiencies, economic disparity, and consumer mistrust. While USDA meat grading traditionally relies on skilled visual inspectors, these human evaluations suffer from cross-facility variability and subjectivity. This paper introduces the first known application of unsupervised domain adaptation regression for cross-facility beef marbling score prediction—an innovation that improves generalization across diverse environments in the beef supply chain. Utilizing numerical scores ranging from 100-900, the research employed convolutional neural networks (CNNs), including ResNet, VGG, and AlexNet architectures. The study specifically introduced and validated a unified unsupervised domain adaptation regression method using the ResNet-50 architecture to enhance model generalization across diverse environments, accounting for variations in lighting, equipment, and operational practices. Statistical analyses demonstrated that the deep learning approach significantly reduced grading variability compared to human graders, achieving greater consistency and accuracy across facilities. The proposed domain adaptation model notably outperformed conventional CNN approaches, offering a scalable, robust, and practical solution for widespread industry adoption. Beyond automating grading, this work lays a foundation for scalable machine vision systems in livestock and distribution logistics, with implications for robotics, food equity, and next-generation supply chain automation.

Keywords

Beef carcass grading; Computer Vision; Deep Learning; Domain Adaptation; Marbling Score Prediction; USDA Grading

Download

Share

COinS