Towards infield, live plant phenotyping using a reduced-parameter CNN

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Growing global population necessitates high-yield crops adaptable to climate change.
• Plant phenotyping is crucial for breeding improved crop varieties.
• Accurate plant image segmentation is a key challenge in phenotyping.

Purpose of the Study:

• To develop compressed deep convolutional neural networks (CNNs) for efficient plant image segmentation.
• To enable the deployment of advanced phenotyping tools on low-cost, resource-limited devices.
• To facilitate infield plant phenotyping for climate-adaptive crop breeding.

Main Methods:

• Applied separable convolutions and Singular Value Decomposition (SVD) to compress very deep CNN models.
• Focused on pixel-wise segmentation of plants into multiple classes, including background.
• Evaluated model performance on public plant and non-plant datasets.

Main Results:

• Achieved up to 95% reduction in model weight matrices using the combined separable convolution and SVD method.
• Maintained high pixel-wise segmentation accuracy despite significant model compression.
• Successfully demonstrated the deployment and functionality of compressed models on a mobile device.

Conclusions:

• Compressed deep CNNs offer a viable solution for deploying advanced plant image segmentation in infield phenotyping.
• The developed methods significantly reduce model size without compromising accuracy, making them suitable for resource-limited devices.
• This advancement supports the development of climate-resilient crops through efficient, accessible phenotyping technologies.