Enhanced Disease Segmentation in Pear Leaves via Edge-Aware Multi-Scale Attention Network

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate segmentation of pear leaf diseases is crucial for effective agricultural management.
• Challenges include variations in disease appearance, lighting, and progression.

Purpose of the Study:

• To develop an advanced deep learning model for precise pear leaf disease segmentation.
• To address the limitations of existing methods in handling complex visual variations.

Main Methods:

• Proposed EBMA-Net, an edge-aware multi-scale network.
• Introduced a Multi-Dimensional Joint Attention Module (MDJA) for multi-scale lesion analysis.
• Incorporated an Edge Feature Extraction Branch (EFFB) to focus on disease boundaries.

Main Results:

• EBMA-Net achieved a Mean Intersection over Union (MIoU) of 86.25%.
• Achieved Mean Pixel Accuracy (MPA) of 91.68% and Dice coefficient of 92.43%.
• Demonstrated superior performance compared to existing models on a custom pear leaf disease dataset.

Conclusions:

• EBMA-Net effectively segments pear leaf diseases under complex conditions.
• The model's architecture enhances diagnostic precision and agricultural disease management.
• Highlights the potential of edge-aware and multi-scale networks in plant pathology.