Evaluation of Convolutional Neural Network Based Classification and Feature Detection Supporting Autonomous Robotic Harvesting
Date of Award
2023
Degree Name
M.S.in Electrical and Computer Engineering
Department
Department of Electrical and Computer Engineering
Advisor/Chair
Raúl Ordóñez
Abstract
Advances in autonomy have created many new opportunities for machine learning applications. Fruit and berry harvesting is one area where intensive human labor continues to be the only way to harvest these crops at scale. To alleviate the high costs and physical demands of human labor, an autonomous system is designed with the purpose of mechanically harvesting strawberries. A Convolutional Neural Network (CNN) provides a decision-making component of the system, aiding in the guidance of a robotic harvesting arm. The CNN is trained on preprocessed strawberry image datasets and determines the stem location in an image to provide situational awareness to the harvesting system. Various overfitting reduction techniques are assessed in CNN accuracy improvement, including batch normalization, regularization, and dropout. Batch normalization and dropout techniques increased overall performance slightly in most models, where regularization decreased performance. The most significant performance increase is attributed to opening the acceptance threshold from one to three stem zones. The highest performing CNN model trained on real-world imagery reached 60% at the single zone threshold and 92% at the three zone threshold. This model was trained using a limited dataset of less than five thousand images spread across eight classifications. A CNN model achieving this accuracy with a small dataset suggests that gathering more data will further improve accuracy. The ideas presented in this work can be expanded and applied to other agricultural or zone-based CNN solutions.
Keywords
Agriculture, Computer Engineering, Department of Computer Science, Robotics, Convolutional Neural Network, Deep Learning, Machine Learning, Data Overfitting Reduction Techniques, Autonomy, Robotic Harvesting, Image Processing, Agriculture
Rights Statement
Copyright © 2023, author
Recommended Citation
Green, Allison Michelle, "Evaluation of Convolutional Neural Network Based Classification and Feature Detection Supporting Autonomous Robotic Harvesting" (2023). Graduate Theses and Dissertations. 7205.
https://ecommons.udayton.edu/graduate_theses/7205
