Supplemental Online Material

Neural networks and robotic microneedles enable autonomous extraction of plant metabolites

Plant Physiology (2021)

Hansol Bae, Magnus Paludan, Jan Knoblauch and Kaare H. Jensen

Department of Physics, Technical University of Denmark, DK-2800 Kgs. Lyngby

E-mail for correspondence: khjensen@fysik.dtu.dk

This repository contains scripts for

1. controlling a Sutter Instruments ROE-200 micromanipulator ("micromanipulator/")
2. generating training data for a CNN for glandular trichome detection on micrographs ("generate_training_data/")
3. training (transfer learning) GoogLeNet for glandular trichome detection on micrographs ("train_neural_network/")
4. testing the network on micrographs ("test_neural_network/")

To generate training data, train the network, and test the network follow these steps: Steps 1-4 may be skipped, as the network is already trained (network weights in train_neural_network/neuralNet.mat).

Generating training data (MAY BE SKIPPED):

1. make folder "training_data" in "generate_training_data".
2. make folders "negative" and "positive" in "generate_training_data/training_data"
3. run the MATLAB code "generate_training_data/generate_training_data.m"

Training (transfer learning) GoogLeNet CNN (MAY BE SKIPPED): 4. run the MATLAB code "train_neural_network/train_network.m"

Test the neural network: 5. run the MATLAB code test_neural_network/test_network.m 6. on line 16 the variable "imageNo" may be changed to any image index in "generate_training_data/full_res_images/". Alternatively, a full-path to a micrograph may be inserted in line 17.