Related Experiment Video
Updated: Nov 12, 2025
05:03
Author Spotlight: Advancing Stomatal Research with Automated Aperture Measurement
Published on: February 9, 2024
1.9K
A generalised approach for high-throughput instance segmentation of stomata in microscope images
Hiranya Jayakody1, Paul Petrie2,3, Hugo Jan de Boer4
1School of Mechanical and Manufacturing Engineering, UNSW, Sydney, Australia. hiranya.jayakody@unsw.edu.au.
Plant Methods
|March 22, 2021
Summary
This study introduces a new automated algorithm for stomata detection in plant images, improving accuracy and boundary identification. The robust method works across various conditions, aiding plant science research.
Area of Science:
- Plant Science
- Computer Vision
- Bioinformatics
Background:
- Stomata analysis from microscope imagery is crucial for understanding plant physiology and environmental responses.
- Existing automated stomata detection methods lack generalizability due to sensitivity to image variations and only provide bounding boxes, limiting morphological studies.
- There is a need for a robust, automated algorithm capable of detecting stomata and their boundaries across diverse plant species, imaging techniques, and magnifications.
Purpose of the Study:
- To develop a fully automated and robust algorithm for stomata detection and boundary identification in microscope imagery.
- To overcome the limitations of existing methods by ensuring general applicability across different plant species, sample collection methods, imaging techniques, and magnification levels.
- To provide plant scientists with a tool that eliminates the need for dataset-specific retraining.
Main Methods:
- Image pre-processing to mitigate color space biases from varied sample preparation and imaging.
- Application of Mask R-CNN with a feature pyramid network for accurate stomata boundary estimation at multiple scales.
- Implementation of a statistical filter to reduce false positives generated by the Mask R-CNN.
Main Results:
- The algorithm achieved high performance metrics: 95.10% precision, 83.34% recall, and 88.61% F-score on diverse datasets.
- Demonstrated generalizability by successfully testing on 7 unseen microscope datasets.
- Achieved an Intersection over Union (IoU) score of 0.70 for boundary estimation, a 7% improvement over bounding-box methods.
Conclusions:
- The developed algorithm exhibits robust performance across varied microscope image datasets, proving its generalizability.
- This automated stomata detection and boundary identification method significantly aids plant scientists by removing the need for re-labeling and retraining for new datasets.
- The open-source code is readily deployable in environments like Google Colab and TensorFlow.
