ConvTransNet-S: A CNN-Transformer Hybrid Disease Recognition Model for Complex Field Environments.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Crop disease identification faces challenges with low accuracy and high model complexity in real-world field conditions.
• Existing models struggle to balance detailed feature extraction with global context understanding, especially under environmental variations like lighting and occlusion.

Purpose of the Study:

• To develop a novel hybrid model, ConvTransNet-S, for accurate and efficient crop disease identification.
• To improve the robustness and reduce the complexity of crop disease identification models in complex field environments.

Main Methods:

• Proposed ConvTransNet-S, a hybrid model integrating Convolutional Neural Networks (CNNs) and transformers.
• Introduced Local Perception Unit (LPU) for fine-grained details and Lightweight Multi-Head Self-Attention (LMHSA) for global dependencies.
• Employed Inverted Residual Feed-Forward Network (IRFFN) for optimized feature propagation and phased architecture for multi-scale feature fusion.

Main Results:

• Achieved 98.85% accuracy on the PlantVillage dataset with 25.14 million parameters.
• On a self-built in-field dataset, ConvTransNet-S reached 88.53% accuracy, outperforming EfficientNetV2, Vision Transformer, and Swin Transformer.
• Demonstrated significant improvements (up to 14.22%) in complex background conditions and reduced parameter count by 46.8% compared to other models.

Conclusions:

• ConvTransNet-S effectively balances local texture perception and global context modeling, overcoming limitations of standalone CNNs or transformers.
• The model's multi-scale feature mechanism successfully distinguishes disease from background features in complex agricultural scenarios.
• ConvTransNet-S offers a promising technical approach for disease diagnosis, showing substantial application value for intelligent agricultural management.