在三維花卉影像中自動偵測維管束和花冠輪廓

Abstract

花冠裂片形狀的變化對於研究人員是一項有趣的主題，花冠三維影像完整地保留下花朵型態資訊，對於後續花瓣裂片的差異性研究提供了最完整狀態的花瓣資料。現階段標記花朵三維形狀特徵的主流方法為Landmarks，然而標記的過程需要高度的專注且十分仰賴研究人員的判斷。本研究提出一種自動偵測方法，目的在於從三維花冠影像中偵測花瓣的一級維管束與邊緣輪廓，並且檢視花瓣形狀是否與不同傳粉者存在關聯性；本研究中，共收集了130個樣本，這些樣本來自於28種不同的岩桐屬物種；偵測流程分別為使用高斯海森矩陣與戴克斯特拉演算法進行一級維管束位置的偵測，以及利用向量強度與平滑化描繪花瓣輪廓；一級維管束與花瓣輪廓的組合作為花朵的形狀特徵，用以描述花瓣的形狀多樣性；本實驗測試結果的成功率為86.54%，兩種不同性質的花朵特徵被提出用以量化和瞭解花瓣形狀與傳粉者的關聯性，花瓣裂片的形狀顯著地影響傳粉者的選擇，蜂鳥傳粉的花朵品系，發展出細長的花冠筒藉以過濾蜂鳥之外的傳粉者，而蜜蜂傳粉的花系，花冠開口方向向上藉此迎合蜜蜂；為了檢測本研究所提出的自動化方法是否適用於其他新鮮合瓣花品系，使用了三種岩桐屬之外的樣本進行測試。與現有方法進行比較，本研究提出一種新穎的方法用於辨識花朵三維形狀特徵。

Corolla shape variation is an interesting theme for researchers. Three-dimensional (3D) images of corollas retain genuine information of the corollas and are optimum for studying corolla shape variation. However, labeling landmarks, which define the shape of a corolla, in a 3D image is labor-intensive. This study proposed a method to automatically detect the first-order veins and corolla contours in 3D corolla images. This study also examines the relationship between corolla shapes and pollinators for genus Sinningia. In the study, 3D images of 130 specimens from 28 species of the genus Sinningia were collected. Procedures were then developed to detect the first-order veins of the corollas using Hessian of Gaussian and Dijkstra’s algorithm and to detect the corolla contours using vector harmony and smoothing. The diversity of corolla shape was represented by means of first-order veins and corolla contour. The successful detection rate reached 86.54%. Two traits, contour-vein ratio and corolla angle, were defined and quantified from the first-order veins and corolla contours to realize the relationship between corolla shapes and pollinators. A choice of pollinators was certainly affected by the shape of a flower. Ornithophily species developed corollas with straight and narrow tubes to sift pollinators other than hummingbirds. Melittophily species developed corollas with a lobe that bends upward to cater to bees. The proposed method was also tested on species other than subtribe Ligeriinae (e.g., A. miresa, S. saxorum, and D. tamiana) to show the applicability to fresh sympetalous flowers. This study brought a new method for identifying flower features in 3D space compared with the existing approaches.