Cassava disease detection using a lightweight modified soft attention network.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Cassava crops are vulnerable to viral infections impacting yield and quality.
• Manual disease identification is time-consuming and requires expert knowledge.
• There's a need for efficient, lightweight disease detection for edge devices.

Purpose of the Study:

• To introduce CDDNet, an efficient and lightweight framework for early cassava disease detection.
• To leverage MobileNetV3Small for optimized feature extraction.
• To enhance disease region prioritization using a modified soft attention module.

Main Methods:

• Utilized MobileNetV3Small as a backbone for feature extraction.
• Implemented a modified soft attention mechanism to focus on diseased plant areas.
• Validated features at early and intermediate stages of disease development.

Main Results:

• Achieved high accuracies: 98.95% (CID), 97.03% (CPDM), and 98.25% (CPCD).
• CDDNet surpasses state-of-the-art methods in accuracy, parameter count, and FPS.
• Demonstrated superior performance for real-time cassava disease detection.

Conclusions:

• Lightweight and efficient techniques are vital for real-time cassava disease management.
• Modified soft attention significantly improves model performance in disease detection.
• CDDNet offers a practical solution for early-stage cassava disease classification.