One-Class Learning Weed Plants Detection on Multispectral Images

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216305%3A26220%2F22%3APU146892" target="_blank" >RIV/00216305:26220/22:PU146892 - isvavai.cz</a>

Nalezeny alternativní kódy

RIV/62156489:43210/22:43922567

Výsledek na webu

<a href="https://ieeexplore.ieee.org/document/9943391" target="_blank" >https://ieeexplore.ieee.org/document/9943391</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1109/ICUMT57764.2022.9943391" target="_blank" >10.1109/ICUMT57764.2022.9943391</a>

Jazyk výsledku

angličtina

Název v původním jazyce

One-Class Learning Weed Plants Detection on Multispectral Images

Popis výsledku v původním jazyce

Modern precision agriculture methods focus on efficient crop care procedures with targeted chemical applications. As current computer vision algorithms allow us to distinguish different plant species, spot-spraying systems for precise herbicide application only on weed plants are being developed and used in praxis. Many such systems rely on deep learning algorithms trained on large datasets containing all possible plant species. While the crop plants keep the same visual characteristics all around the world, the species composition of weed plants can differ significantly, leading to lower detection accuracy for weed species that are not represented in the training dataset. In this work, we apply the PatchCore and PaDiM, two state-of-the-art anomaly detection algorithms, to a multispectral dataset of corn plant images in a one-class learning paradigm. The best performing algorithm achieved AUROC 94.2 despite the high visual heterogeneity and scarcity of the input data. Our results suggest it is possible to train weed-detection algorithms on a limited dataset in a one-class learning setting to transform the species classification into an anomaly detection problem.

Název v anglickém jazyce

One-Class Learning Weed Plants Detection on Multispectral Images

Popis výsledku anglicky

Modern precision agriculture methods focus on efficient crop care procedures with targeted chemical applications. As current computer vision algorithms allow us to distinguish different plant species, spot-spraying systems for precise herbicide application only on weed plants are being developed and used in praxis. Many such systems rely on deep learning algorithms trained on large datasets containing all possible plant species. While the crop plants keep the same visual characteristics all around the world, the species composition of weed plants can differ significantly, leading to lower detection accuracy for weed species that are not represented in the training dataset. In this work, we apply the PatchCore and PaDiM, two state-of-the-art anomaly detection algorithms, to a multispectral dataset of corn plant images in a one-class learning paradigm. The best performing algorithm achieved AUROC 94.2 despite the high visual heterogeneity and scarcity of the input data. Our results suggest it is possible to train weed-detection algorithms on a limited dataset in a one-class learning setting to transform the species classification into an anomaly detection problem.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

20204 - Robotics and automatic control

Projekt

<a href="/cs/project/FW03010273" target="_blank" >FW03010273: Defektoskopie lakovaných dílů s pomocí automatické adaptace neuronových sítí</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2022

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

2022 14th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT)

ISBN

979-8-3503-9866-3

ISSN

2157-023X

e-ISSN

—

Počet stran výsledku

4

Strana od-do

76-79

Název nakladatele

IEEE

Místo vydání

Valencia, Spain

Místo konání akce

Valencia, Spain

Datum konání akce

11. 10. 2022

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—