A spontaneous keypoints connection algorithm for leafy plants skeletonization and phenotypes extraction

Zhen WangXiangnan HeYuting WangChenxue YangXian Li^*

• Agricultural Information Institute, Chinese Academy of Agricultural Sciences, Haidian, China

Leaves' phenotypes are closely related to growth status of plants. Skeletonization can simplify leaves into central skeleton structures, enabling efficient and precise geometric calculations of phenotypes based on keypoints and skeleton. Deep learning-based skeletonization methods require extensive manual labeling, long-term training, and are constrained by pre-annotated keypoints. In this study, a spontaneous keypoints connection skeletonization approach was developed for leafy plants without labeling and training, including random point generation and keypoints connection. For plants with random leaf morphology, a threshold for angle difference between any three consecutive adjacent points was determined, and keypoints in circular search areas were adaptively identified one by one to accurately skeletonize leaves. For plants with regular leaf morphology, the trajectory pattern was fitted by minimizing curvature. The performance of the proposed approach was validated using vertical and front view images of leafy orchids, representing random and regular morphological cases, respectively. The average curvature error of skeleton fitting was 0.12 and the average leaf recall rate was 92%. Five phenotypic parameters of orchids were accurately extracted from the skeleton. Additionally, generalization capability of the algorithm was validated on a maize dataset. This approach demonstrated effective skeletonization results for both randomly distributed and regularly shaped leafy plants.