A novel lightweight hybrid CNN-ViT for maize leaf disease classification.

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Maize is a critical global crop facing significant yield losses due to plant diseases.
• Manual disease diagnosis is labor-intensive, error-prone, and lacks scalability.
• Automated methods using deep learning show potential but face limitations with existing architectures like CNNs and ViTs.

Purpose of the Study:

• To develop a lightweight, hybrid deep learning model for accurate and efficient maize disease diagnosis.
• To overcome the limitations of Convolutional Neural Networks (CNNs) in capturing global context and Vision Transformers (ViTs) in data and computational requirements.

Main Methods:

• Proposed a novel hybrid model inspired by Mixture-of-Experts (MoE) architectures, integrating CNN and ViT components.
• The model adaptively balances local and global feature extraction based on input image characteristics.
• Evaluated the model on a new, real-world dataset comprising full maize plant images.

Main Results:

• Achieved an exceptionally high classification accuracy of 99.90% on the maize disease dataset.
• Significantly outperformed established state-of-the-art models including MobileViT, PiT, EdgeNeXt, and DeiT.
• Demonstrated the effectiveness of lightweight hybrid architectures for plant disease diagnosis.

Conclusions:

• Lightweight hybrid deep learning models offer a viable solution for high-performance, automated plant disease diagnosis.
• The developed model shows great promise for practical deployment in agricultural settings to improve crop management.
• This approach addresses the need for efficient and accurate diagnostic tools to safeguard global food security.