Semantic segmentation for plant phenotyping using advanced deep learning pipelines

Large strides have been made in the field of semantic segmentation which finds its application in extensive areas of research. However, these advancements have not been completely utilized in the field of plant phenotyping. Deriving quantitative plant phenotypes in a non-destructive manner from plant images is a key challenge that strongly relies on the precise segmentation of plant images. In this paper, we propose novel semantic segmentation pipelines for the task to improve the automated phenotyping process. In this work architectures such as U-Net, Attention-Net and Attention-Augmented Net are introduced that are trained on the Arabidopsis Thaliana plant dataset released under the CVPPP14 competition. Dice coefficient is used as the evaluation metric to compare performances of the proposed architectures, and also benchmark them against existing algorithms in literature. Results of semantic segmentation of Rosette plants shows the state-of-the-art results, with attention net achieving a 0.985 dice score that easily outperforms all the other deep learning and image processing techniques proposed earlier for plant segmentation in this domain. Results are exhibited with comparison analysis successfully with these advanced deep learning architectures and can be used as a base for plant phenotyping related applications.