Colour index evaluation method for plant segmentation from a soil background

Date

2007

Authors

Golzarian, M.
Desbiolles, J.M.A.
Lee, M.K.

Editors

Advisors

Journal Title

Journal ISSN

Volume Title

Type:

Conference paper

Citation

Precision agriculture '07, 2007, pp.325-332

Statement of Responsibility

Conference Name

6th European Conference on Precision Agriculture (3 Jun 2007 - 6 Jun 2007 : Skiathos, Greece)

Abstract

Computer vision applications in agriculture require reliable techniques of plant segmentation from variable soil backgrounds. In this research, a range of 13 RGB and hue-based colour indices were evaluated for their potential to segment wheat seedlings from a red loamy-sand soil background. To compare the performance of colour indices, a 'buffer' concept separating plant and background pixel distributions was developed. With an algorithm in MATLAB environment, the colour values were extracted from 60 images (840 separate regions) and the indices were computed for three region types of interest, namely green plants, soil background and surface pebble objects. A buffer ratio was used to enable the comparative evaluation of the 13 indices. Under these experimental conditions, the difference between red and green as well as NDI (normalized difference index) were found the best performing plant segmentation parameters against the soil background, while MEGI (modified excessive green index) showed potential for plant-pebble segmentation, followed by normalized g, EGI (excessive green index) and hue. It has also mathematically been proved that EGI is a linear function of g. The buffer ratio approach followed can use separate optimisation procedures for reducing either type I or type II errors, which are the percentages of misclassified plant and background regions respectively, depending upon the application of the study.

School/Discipline

Dissertation Note

Provenance

Description

Access Status

Rights

Copyright status unknown

License

Grant ID

Published Version

Call number

Persistent link to this record