Deep learning-based approach for extracting inflorescence morphology features in cut chrysanthemum

Abstract

Accurate identification of floral morphological traits, such as flower type and the diameters of ligulate and bisexual flowers, is essential for the quality evaluation and varietal improvement of cut chrysanthemums ( Chrysanthemum morifolium Ramat.). Traditional manual or rule-based image processing methods are inefficient and struggle with complex floral structures. To address these limitations, we developed a lightweight deep learning and machine learning pipeline for automated trait extraction in over 30 chrysanthemum cultivars. A ShuffleNet V2 model achieved 95.24% accuracy in flower type classification, with lightweight characteristics (1.26M parameters, 0.15 GFLOPs), fast inference time (14.78 ms per image), and 67.65 FPS. Ligulate and bisexual flowers were segmented using an optimized U-Net achieving a reduction of over 95% in parameters, achieving an average Dice similarity coefficient (DSC) of 0.934. For diameter estimation, mean squared errors (MSE) of 6.605 mm (ligulate) and 2.034 mm (bisexual) were obtained, with coefficients of determination (R2) approaching 0.98. Fine-grained classifications—Single-petals vs. Repeating-petals and Incurve vs. Honeycomb—were achieved using geometric and texture features with F1-scores above 0.87. These results demonstrate a scalable and efficient solution for floral trait analysis, supporting high-throughput phenotyping in ornamental horticulture.

Plant phenotyping relevance

キクの花器官形態を画像から自動抽出・推定する深層学習パイプラインを開発しており、植物表現型取得が研究の中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.