NTU automatic stomata image detection in python

• install miniconda

• install visual code studio

• open miniconda, create conda virtual environment stomaenv

conda create --name stomaenv python=3.10.0 jupyter pandas
# after the first line has done
conda activate stomaenv
conda install conda-forge::ultralytics
conda install anaconda::dill
conda install anaconda::ipykernel
##optional: when you have GPU
#conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
#conda install nvidia/label/cuda-12.1.0::cuda

• open NTU_detect_v8.ipynb with visual code studio, select virtual environment stomaenv
• run all the cell, select the folder where you store your stomata images, results would be generated in the folder runs

Manual image labelling & comparison with detection in R

preparation

please clone this repository to your local directory.

1. open Stomata.RProj

2. open run.R in folder “src”

3. run set_up.R to install necessary packages.

4. create two folders named data and results from line 6-7.

5. create sub-folder(s) under directory “data” (data/sub-folder)

   For example: folder data/T16L600 contains all .xml files from experiment T16L600.

   !!! .xml files can’t directly put in data, your need to put them under a folder inside data!!!

read ground truth (manual labelling)

6. run read_xml.R from line 11-13 to read the files and generate basic results for each batch
7. run stat_analysis.R line 18-20 to read the files and generate statistics for each stomata rows.
8. run summarize_and_merge.R line 23-25 to generate summarize statistics for each picture.

Quality control of NTU automatic stomata image detection

To tackle the issues of same detected position of a stomata have different class from the NTU pipeline.

9. Remove the duplicated.

1. run line 29 to create the folder “result/Ntu”
2. then placed the no ground truth’s pictures in “result/Ntu”
3. run clean_ntu.R from line 33-35 to remove the duplicated detected stomata that has lower confidence.
4. merge.RDS will be the output removing the duplicated coordinates.

10. Compare the ground truth with the detection

For those picture that is both manually labelled and detected by the NTU

1. run check_ntu.R from line 38-40 to validate the detected data with ground truth

• note that only picture with the same picture name as ground truth when the detected stomata center is 3.2 microm to the ground truth.

unit and names of stomata object

Stomata

Pull request

Want to contribute to the code? Very welcome!

1. copy the script that you want to modified in ‘src/test’
2. modified the code
3. test if there is error
4. move it to the original folder and overwrite
5. make a pull request

❓ Issues/Problems/Errors

Please raise your question in Issues and set a request in Project!

Guideline for reporting error

1. Make sure your code is up to date (pull from github)

2. Describe the issue/problem/error clearly.

3. identify the error: Is it an error? (Do you received error message?)

• If yes, Which code, which line, which error
• If it is inside a loop, can you identify the index inside the loop?
• see here more for debug

4. Interpretation of the message: Google the error/ Chatgpt
5. Reproduce the error: make sure it is reproducible
6. Post on github/slack for help