A pilot programme of fruit grading system using computer vision

Loading...
Thumbnail Image
Other Title
Authors
Wang, Liang
Author ORCID Profiles (clickable)
Degree
Master of Information Technology
Grantor
Eastern Institute of Technology (EIT)
Date
2021
Supervisors
Erturk, Emre
Dang, Daniel
Type
Masters Dissertation
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
computer vision
object detection
machine learning
deep learning
fruit grading
SageMaker
MXNet
TensorFlow
New Zealand
Citation
Wang, L. (2021). A pilot programme of fruit grading system using computer vision. (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Information Technology). Eastern Institute of Technology (EIT), New Zealand. https://hdl.handle.net/10652/5667
Abstract
RESEARCH QUESTIONS: 1. How can New Zealand’s agriculture industry use computer vision technologies to improve the accuracy of their fruit grading system? 2. What are the similarities and differences in object detection and image classification effectiveness in grading fruits? 3. What factors influence agricultural companies’ decision on embracing computer vision for fruit grading in New Zealand? ABSTRACT: Traditional fruit grading work depends on a large labour force during harvesting time. However, the grading accuracy varies, resulting in difficulties in product quality management. Due to the COVID-19 pandemic, many of New Zealand's farming industries lack seasonal workers from overseas. Hence, they are looking at taking advantage of promising computer vision technologies to avoid reliance on the labour-intensive grading method. This research project designs a prototype fruit grading system using object detection algorithms to automatically sort fruits (e.g., squash), minimising manual intervention during the production process. The proposed system consists of fruit handling and image processing modules. Amazon's machine learning platform SageMaker and Google's machine learning framework TensorFlow are the two main software components in the system. We tested the prototype in a simulated production environment. The result proved that the selected approach could suit farming industries to achieve automation transformation during post-harvesting. Other findings include that object detection has better performance than image classification on identifying defects on fruits. The high cost of setting up a new fruit grading system has hindered agriculture companies from adopting the plan. Further research within this topic could incorporate farming experts to help gain higher dataset accuracy during labelling jobs and choose a proper approach to avoid unnecessary difficulties in manipulating data between different machine learning platforms.
Publisher
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved
Copyright license
Available online at