Leaf Segmentation and 3D Reconstruction of ARAFIDOPSIS Based on MASK R-CNN

摘要

Plant phenotypic analysis is the study of plant growth by measuring its morphological details. The traditional phenotypic analysis relies on human visual inspection and manual measurements. As a result, phenotypic characteristics that can be studied are limited and the analysis is time-consuming. Most of the phenotypic analyses are currently performed using 2D images of plants. The growth of the plant in 3D cannot be assessed. We first took photographs of the plant from above at two different angles to estimate its height. We then used the Mask R-CNN deep learning algorithm to identify and segment the plant leaves for characterization of the features. We used hand-marked outlines of the Arabidopsis leaf images as the model training set. We then tested four images to identify 81.5% of the leaves, with a leaf range correctness of 72.8%. Finally, a stereo algorithm was used to calculate the corresponding points of the leaves in the images taken at the two different angles to characterize the 3D phenotype and to build a 3D model of the plant.