Leaf Segmentation and Counting for Phenotyping of Rosette Plants Using Xception-style U-Net and Watershed Algorithm

AbstractLeaf count is an important plant trait that helps for the analysis of plant growth in various plant phenotyping tasks. Overlapping leaves and nastic leaf movements are the major challenges in the recognition and counting of plant leaves. In this paper, we present a two-stage framework for the segmentation, counting and localization of plant leaves. In the first stage, we have designed an encoder-decoder-based deep neural network, namely Xception-style U-Net to segment plant leaves from the background. In the second stage for counting plant leaves, we use distance transform and watershed algorithm. The performance of the proposed method was tested on a publicly available leaf counting challenge (LCC) 2017 dataset that includes images of rosette plants, namely Arabidopsis thaliana and tobacco. In this work, Xception-style U-Net achieves improved segmentation accuracy on the test dataset with the dice coefficient of 0.9685. Xception-style U-Net, along with the watershed algorithm, achieves an average difference in leaf count (DiC) of 0.26 and absolute difference in leaf count (\(\vert DiC \vert \)) of 1.93, better than existing methods in the literature.KeywordsLeaf segmentationLeaf countingXception-style U-NetSegNetWatershed algorithmPlant phenotyping