STOMATA SEGMENTATION: A LIGHTWEIGHT HIERARCHICAL FOURIER PRE-PROCESSING APPROACH

Submitter: Clint McElroy

Permalink

https://hdl.handle.net/2142/114940 Copy

Description

Title

STOMATA SEGMENTATION: A LIGHTWEIGHT HIERARCHICAL FOURIER PRE-PROCESSING APPROACH

Contributor(s)

Koleva, Liana

Issue Date

2022-05

Keyword(s)

plant phenotyping; computer vision; machine learning

Abstract

The importance of agricultural research soars as the climate crisis intensifies, leading to unpredictable effects on crop growth and food security [4]. Current methods used by agricultural researchers for plant phenotyping purposes are prohibitively expensive and time-consuming [5] and often involve harvesting, which is destructive to samples. Technical development can increase phenotypic throughput [5] while allowing plants to be studied in their natural environment [8], and demand is consequently high. Machine learning has been applied to computer vision and signal processing problems – including phenomics – to achieve results matching and exceeding those accomplished by humans. Current state-of-the-art plant phenotyping methods based on computer vision use deep learning techniques [8], which are computationally expensive and require large datasets. Such methods are promising but unsustainable, as the energy crisis shows no signs of slowing. Motives to develop computationally efficient software range from the financial to the environmental to the practical [9]. Attempts are being made to simplify and lighten models in various contexts [1, 3], but such work has not been extensively explored with regards to semantic segmentation in the context of plant phenomics. Given the periodic nature of cell distribution in grass leaves, exploring the frequency domain is a possible approach to extract useful information and improve the results. This paper pursues a novel hierarchical approach towards stomata segmentation, utilizing the Fourier transform of iimicroscopic leaf surface input images through frequency filters that distinguish cells based on their systematic spatial repetition. Taking a hierarchical approach to model development rather than relying solely on deep learning allows for a greater level of efficiency in the specific domain of interest.

Type of Resource

text

Language

en

Handle URL

https://hdl.handle.net/2142/114940

Owning Collections