VMamba for plant leaf disease identification: design and experiment.

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Crop diseases threaten agricultural production, reducing yield and quality.
• Machine vision and AI show promise in plant disease recognition.
• Current deep learning models struggle with complex agricultural environments and small datasets, leading to reduced accuracy and longer training times.

Purpose of the Study:

• Introduce VMamba, a visual backbone model, for agricultural plant disease detection.
• Address challenges of computational complexity, accuracy, and small sample sizes in plant disease identification.
• Propose the DDHTLVMamba method combining VMamba with diffusion models and transfer learning for small-sample datasets.

Main Methods:

• Implemented VMamba, a visual backbone model with a selective scanning mechanism.
• Developed the DDHTLVMamba method integrating VMamba, diffusion models, and transfer learning.
• Evaluated VMamba performance on various datasets and training strategies, comparing it with mainstream deep learning architectures.

Main Results:

• VMamba outperformed ResNet50, Vision Transformer, and Swin Transformer in disease recognition accuracy on both large and small datasets.
• VMamba achieved a 3.49% accuracy increase and 80% training time reduction compared to Swin Transformer.
• The DDHTLVMamba method effectively reduced pre-training time on small-sample datasets while maintaining high recognition accuracy.

Conclusions:

• VMamba offers a superior approach for agricultural plant disease detection, enhancing accuracy and efficiency.
• The DDHTLVMamba method provides an effective solution for small-sample agricultural datasets.
• This research contributes to advancing intelligent agricultural disease prevention and control technologies.