Towards low-cost image-based plant phenotyping using reduced-parameter CNN

Atanbori, John, Chen, Feng, French, Andrew P. and Pridmore, Tony P. (2018) Towards low-cost image-based plant phenotyping using reduced-parameter CNN. In: CVPPP 2018: Workshop on Computer Vision Problems in Plant Phenotyping, 6 Sept 2018, Newcastle upon Tyne, UK.

[img]
Preview
PDF - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Download (1MB) | Preview

Abstract

Segmentation is the core of most plant phenotyping applications. Current state-of-the-art plant phenotyping applications rely on deep Convolutional Neural Networks (CNNs). However, these networks have many layers and parameters, increasing training and test times. Phenotyping applications relying on these deep CNNs are also often difficult if not impossible to deploy on limited-resource devices. We present our work which investigates parameter reduction in deep neural networks, a first step to moving plant phenotyping applications in-field and on low-cost devices with limited resources. We re-architect four baseline deep neural networks (creating what we term "Lite CNNs") by reducing their parameters whilst making them deeper to avoid the problem of overfitting. We achieve state-of-the-art, comparable performance on our "Lite" CNNs versus the baselines. We also introduce a simple global hyper-parameter (alpha) that provides an efficient trade-off between parameter-size and accuracy.

Item Type: Conference or Workshop Item (Paper)
Additional Information: Workshop is held at 29th British Machine Vision Conference, Northumbria University, UK, 4-6 September 2018.
Schools/Departments: University of Nottingham, UK > Faculty of Science > School of Biosciences
University of Nottingham, UK > Faculty of Science > School of Computer Science
Related URLs:
URLURL Type
https://www.plant-phenotyping.org/CVPPP2018UNSPECIFIED
Depositing User: Eprints, Support
Date Deposited: 13 Sep 2018 07:46
Last Modified: 13 Sep 2018 07:46
URI: https://eprints.nottingham.ac.uk/id/eprint/54696

Actions (Archive Staff Only)

Edit View Edit View