MRSliceNet: Multi-Scale Recursive Slice and Context Fusion Network for Instance Segmentation of Leaves from Plant...

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Biology

Background:

• Plant phenotyping is crucial for understanding genotype-environment interactions, impacting crop breeding and precision agriculture.
• Traditional leaf measurement is labor-intensive and destructive.
• Modern 3D sensing (e.g., LiDAR) offers high-resolution data but struggles with automated leaf segmentation due to occlusion and geometric complexity.

Purpose of the Study:

• To develop an automated deep learning framework for precise 3D leaf segmentation from LiDAR point clouds.
• To overcome challenges in segmentation caused by occlusion, similar shapes, and varying point density.

Main Methods:

• Proposed MRSliceNet, an end-to-end deep learning framework inspired by human visual cognition.
• Integrated Multi-scale Recursive Slicing Module (MRSM) for local feature extraction.
• Incorporated Context Fusion Module (CFM) for integrating local and global features using attention.
• Utilized Instance-Aware Clustering Head (IACH) for accurate instance separation.

Main Results:

• Achieved state-of-the-art performance on two challenging datasets.
• Reported Average Precision (AP) of 55.04%/53.78%, AP50 of 65.37%/64.00%, and AP25 of 74.68%/73.45% on Dataset A and B.
• Demonstrated reliable instance identification with clear boundaries.

Conclusions:

• MRSliceNet provides an effective solution for automated plant phenotyping using 3D data.
• The framework has been successfully implemented in real-world agricultural research pipelines.
• Advances in deep learning offer significant potential for high-throughput plant analysis.