Computer vision based corn kernel quality evaluation: Traditional versus machine learning

Li, Xing

Permalink

https://hdl.handle.net/2142/101131 Copy

Description

Title

Computer vision based corn kernel quality evaluation: Traditional versus machine learning

Author(s)

Li, Xing

Issue Date

2018-03-27

Director of Research (if dissertation) or Advisor (if thesis)

Ahuja, Narendra

Department of Study

Electrical & Computer Eng

Discipline

Electrical Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

Computer Vision, Corn Kernel Quality, Machine Learning, Fast R-CNN, RPN, Faster R-CNN, FPN, Retinanet.

Abstract

Corn kernel quality evaluation is a trivial task for experienced farmers and agriculture researchers, but it becomes tricky if we try to develop a computer vision based automatic solution. In this thesis, we present two approaches for this problem, briefly introduce the data sets corresponding to each method and compare the accuracy between them. We attack the corn kernel quality evaluation problem by two different methods: (1) Evaluate the quality based on the percentage of good corn kernels within the scope by a “percentage” classifier trained with multi-class support vector machine (SVM).(2) Evaluate the quality by a good corn kernel detector trained with multiple state-of-the-art detectors, specifically Faster R-CNN and Retinanet. We collected two databases for both methods separately: (1) Images of many corn kernel batches containing different percentages of good corn kernels vs.foreign matter randomly placed on a flat surface were taken as both training and testing data for multi-class SVM. (2) Reuse the images taken for the SVM data set and add bounding box annotations to each image following the Microsoft COCO fashion. Our experiments show that multi-class SVM reaches a rank-1 accuracy of 78%, while the deep learning detectors achieved96% precision. While the multi-class SVM approach shows good classification results, deep learning models provide more precise detection results.Unfortunately, previous works are all based on lab environments and there is no benchmark available in this field. Therefore, we consider our work as a baseline for corn kernel quality evaluation.

Graduation Semester

2018-05

Type of Resource

text

Permalink

http://hdl.handle.net/2142/101131

Copyright and License Information

Copyright 2018 Xing Li

Owning Collections