Leaf Segmentation by Functional Modeling

The use of Unmanned Aerial Vehicles (UAVs) is a recent trend in field based plant phenotyping data collection. However, UAVs often provide low spatial resolution images when flying at high altitudes. This can be an issue when extracting individual leaves from these images. Leaf segmentation is even more challenging because of densely overlapping leaves. Segmentation of leaf instances in the UAV images can be used to measure various phenotypic traits such as leaf length, maximum leaf width, and leaf area index. Successful leaf segmentation accurately detects leaf edges. Popular deep neural network approaches have loss functions that do not consider the spatial accuracy of the segmentation near an object's edge. This paper proposes a shape-based leaf segmentation method that segments leaves using continuous functions and produces precise contours for the leaf edges. Experimental results prove the feasibility of the method and demonstrate better performance than the Mask R-CNN.