ULeaf-Net: Leaf Segmentation Network Based on U-shaped Symmetric Encoder-Decoder Architecture

The study of plant phenotypes can help improve crop yields in response to the planet's food resources scarcity. Traditional approaches to plant phenotyping are destructive and require the empirical judgment of experts. To improve efficiency and accuracy, researchers begin to explore the feasibility of the determination of plant phenotypic parameters automatically, which is achieved by precisely segmenting the plant leaf profile. Some deep-learning-based leaf segmentation methods proposed in recent years did not ideally work because they are limited by the quality and size of the dataset and the reasonableness of the network architecture itself. Therefore, a leaf segmentation network based on U-shaped symmetric encoder-decoder architecture called ULeaf-Net is proposed. It replaces the traditional same-layer feature fusion with cross-layer feature fusion and introduces a robust feature extraction structure BasicBlock, and it also uses a patch learning method to expand the dataset size in the training stage for better training of the network. Finally, we compare the leaf segmentation results of ULeaf-Net with UNet. ULeaf-Net has an excellent leaf segmentation capability both in terms of evaluation metrics and intuitively.