Gradient-guided boundary-aware selective scanning with multi-scale context aggregation for plant lesion segmentation.

Area of Science:

• Computer Vision
• Plant Pathology
• Precision Agriculture

Background:

• Plant lesion segmentation is crucial for early disease diagnosis and management in agriculture.
• Challenges include variations in lesion scale and ambiguous boundaries that blend with healthy tissue.

Purpose of the Study:

• To develop an advanced deep learning model for accurate plant lesion segmentation.
• To address the challenges of scale variation and boundary ambiguity in plant disease detection.

Main Methods:

• Introducing GARDEN (Gradient-guided boundary-Aware Region-Driven Edge-refiNement network).
• Utilizing a Multi-Scale Context Aggregation (MSCA) module for scale-consistent lesion priors.
• Employing a Boundary-aware Selective Scanning (BASS) module with a Gradient-Guided Boundary Predictor (GGBP) for selective long-range refinement.

Main Results:

• GARDEN achieves state-of-the-art performance on plant disease datasets for both overlap and boundary metrics.
• Demonstrates significant improvements in segmenting small lesions and boundary-ambiguous cases.
• Qualitative results show sharper contours and increased robustness to illumination and viewpoint variations.

Conclusions:

• GARDEN offers accurate and reliable plant lesion segmentation by combining scale robustness and boundary precision.
• This method provides a robust solution for automated disease analysis in challenging agricultural conditions.