Leaf Segmentation Using Modified YOLOv8-Seg Models

Area of Science:

• Computer Vision
• Plant Phenotyping
• Agricultural Technology

Background:

• Automated plant leaf segmentation is crucial for plant classification, growth monitoring, and precision agriculture.
• Existing computer-vision models require enhancement for improved segmentation accuracy, especially for small or overlapping leaves.

Purpose of the Study:

• To improve automated plant leaf segmentation using computer-vision techniques.
• To evaluate the efficacy of integrating Ghost and Bidirectional Feature Pyramid Network (BiFPN) modules into the YOLOv8-seg model.

Main Methods:

• The YOLOv8-seg model was employed as the baseline for leaf segmentation.
• Two modified YOLOv8-seg architectures were proposed, incorporating Ghost modules for efficient feature generation and BiFPN modules for multi-scale feature fusion.
• Experiments were conducted on five datasets from the Computer Vision Problems in Plant Phenotyping (CVPPP) Leaf Segmentation Challenge.

Main Results:

• The standard YOLOv8-seg model demonstrated good performance on the leaf segmentation task.
• The integration of Ghost and BiFPN modules significantly enhanced segmentation performance.
• The proposed modified YOLOv8-seg approach achieved a top score of 86.4% Dice on the CVPPP Leaf Segmentation Challenge datasets.

Conclusions:

• The enhanced YOLOv8-seg models, incorporating Ghost and BiFPN modules, offer superior performance for plant leaf segmentation.
• This technology holds significant potential for advancing precision agriculture and plant phenotyping research.
• The findings suggest that architectural modifications can substantially improve the accuracy of computer-vision-based agricultural applications.