Integration of U-Net and FastSAM for Accurate Leaf Image Segmentation in Complex Backgrounds

Abstract

Leaf segmentation plays a crucial role in plant phenotyping and precision agriculture, enabling the monitoring of growth, disease detection, and informed crop management. However, accurate segmentation in natural environments is challenging due to complex backgrounds, overlapping structures, irregular boundaries, and varying illumination. This paper proposes a hybrid six-stage framework that integrates U-Net with the Fast Segment Anything Model (FastSAM) to achieve accurate and efficient leaf segmentation. The pipeline consists of initial U-Net segmentation, largest component filtering, contour extraction with convex hull transformation, bounding box derivation via distance transform, promptable refinement with FastSAM, and final contour selection. The experiments conducted used 633 images from the Pl@ntLeaves database: 333 images for model development with a train/validation split of 266/67 (20% validation), and a held-out test set of 300 images. On the 300-image test set, the proposed framework achieved superior results (Precision = 0.966, Recall = 0.945, Intersection over Union (IoU) = 0.917, Dice = 0.953, HD95 = 27.859), outperforming DeepLabV3 and CLIPSeg. These findings confirm that combining U-Net's fine-grained feature extraction with FastSAM's efficient prompt-based refinement provides a robust and scalable solution for plant phenotyping and precision agriculture, particularly by enhancing boundary accuracy in complex natural scenes.

Plant phenotyping relevance

植物葉の画像セグメンテーション手法を開発し、独立テストデータで既存手法と比較検証しており、葉形態の取得を目的とするフェノタイピング手法が中心である。

Code and data availability

The paper's only paper-specific asset is the Pl@ntLeaves leaf image dataset (with ground-truth masks) used for all experiments; the authors explicitly state it is publicly available at the LIRIS REVERES databases page. No author analysis code, trained model checkpoints, or supplementary data deposits are mentioned.