A wavelet-based frequency-domain approach for accurate multi-crop disease detection.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate crop disease diagnosis is vital for effective management and reduced pesticide use.
• In-field lesion detection is challenging due to variable appearance, scale, lighting, and shadows.
• Existing methods struggle to balance high accuracy with lightweight inference for real-time applications.

Purpose of the Study:

• To develop a lightweight yet accurate crop disease recognition model for in-field applications.
• To introduce novel modules for enhanced feature extraction and multi-scale context aggregation.
• To improve the efficiency and deployment friendliness of crop disease detection systems.

Main Methods:

• Proposed WGA-YOLO, a lightweight YOLO variant incorporating Wavelet Channel Recalibration (WCR) for multi-resolution feature fusion.
• Introduced PS-C2f module with Pinwheel-shaped convolutions for capturing fine lesion details.
• Replaced SPPF with Dynamic Group Attention Pooling (DGAP) for efficient multi-scale context aggregation.

Main Results:

• WGA-YOLO achieved 3.02% and 2.85% higher accuracy than YOLOv8n on the PlantDoc_boost dataset.
• Reduced model parameters by ~0.18M and FLOPs by ~0.3G compared to YOLOv8n.
• Demonstrated superior inference efficiency and deployment friendliness.

Conclusions:

• WGA-YOLO effectively addresses the challenges of in-field crop disease detection.
• The proposed WCR and PS-C2f modules enhance feature representation and detail capture.
• WGA-YOLO offers a promising solution for real-time, accurate, and efficient crop disease recognition.