Leaf vein segmentation with self-supervision

The leaf vein, often viewed as the fingerprint of the leaf, is an important characteristic used to identify plant species. Leaf vein segmentation aims to extract the vein features and obtain the vein architectures from leaf images. Unlike the general semantic or instance segmentation focusing on the block and object level, the leaf vein segmentation is fine and it focuses on the internal details inside the mesophyll, whose color is indistinguishable, making this task difficult. To tackle this problem, we utilize the particularities of leaf veins, namely continuity and branching, and propose a Co nfidence R efining V e in Network (CoRE-Net) to segment leaf veins by handling the intersections, breakpoints, and blurred boundaries of veins to enhance segment prediction. Moreover, the proposed network only needs a few labeled samples to warm start, and then the whole network can converge without any annotated labels. Meanwhile, we collect and release the first leaf vein dataset, Leaf Vein Dataset 2021 (LVD2021). The proposed framework achieves mean Intersection over Union at 71.02% and mean Dice at 79.76% on LVD2021, which outperforms its counterparts in different settings, demonstrating the effectiveness of our framework. • A novel framework CoRE-Net to improve predict results using vein’s particularities. • CoRE-Net converges with only a few labeled samples in a self-supervision manner. • We collect and release the first pixel-wise annotation vein dataset, namely LVD2021.