DiSCount: computer vision for automated quantification of Striga seed germination.

Abstract

Background Plant parasitic weeds belonging to the genus Striga are a major threat for food production in Sub-Saharan Africa and Southeast Asia. The parasite's life cycle starts with the induction of seed germination by host plant-derived signals, followed by parasite attachment, infection, outgrowth, flowering, reproduction, seed set and dispersal. Given the small seed size of the parasite ( Striga seed germination. Results Here, we introduce DiSCount ( Di gital S triga Count er): a computer vision tool for automated quantification of total and germinated Striga seed numbers in standard glass fibre filter assays. We developed the software using a machine learning approach trained with a dataset of 98 manually annotated images. Then, we validated and tested the model against a total dataset of 188 manually counted images. The results showed that DiSCount has an average error of 3.38 percentage points per image compared to the manually counted dataset. Most importantly, DiSCount achieves a 100 to 3000-fold speed increase in image analysis when compared to manual analysis, with an inference time of approximately 3 s per image on a single CPU and 0.1 s on a GPU. Conclusions DiSCount is accurate and efficient in quantifying total and germinated Striga seeds in a standardized germination assay. This automated computer vision tool enables for high-throughput, large-scale screening of chemical compound libraries and biological control agents of this devastating parasitic weed. The complete software and manual are hosted at https://gitlab.com/lodewijk-track32/discount_paper and the archived version is available at Zenodo with the DOI 10.5281/zenodo.3627138. The dataset used for testing is available at Zenodo with the DOI 10.5281/zenodo.3403956.

Plant phenotyping relevance

Striga種子の発芽状態を画像から自動定量するコンピュータビジョン手法の開発・検証が研究の中心であり、植物状態の表現型取得に該当する。

Code and data availability

The paper's DiSCount software (code, manual, trained YOLOv3 model) is hosted on GitLab and archived on Zenodo; the manually counted test image dataset and the installation-test input dataset are also publicly available on Zenodo. The pytorch-yolo-v3 repository is a third-party code base, not a paper-specific asset.