A dual-branch model combining convolution and vision transformer for crop disease classification.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Crop disease classification is crucial for food security but challenging due to complex visual features.
• Existing computer vision models often struggle with the intricate textures and shapes of plant diseases.
• The integration of Convolutional Neural Networks (CNN) and Vision Transformers (ViT) offers a promising avenue for improved classification.

Purpose of the Study:

• To develop a lightweight and highly accurate dual-branch model for crop disease classification.
• To effectively combine local feature extraction (CNN) with global feature understanding (ViT).
• To enhance the representation capability of Transformer models for agricultural applications.

Main Methods:

• A dual-branch architecture integrating CNN for local features and ViT for global features.
• Introduction of an Aggregated Local Perceptive Feed Forward Layer (ALP-FFN) to enhance Transformer locality.
• Development of a lightweight Transformer block with ALP-FFN and linear self-attention for reduced complexity.

Main Results:

• Achieved 99.71% accuracy on the PlantVillage dataset with only 4.9M parameters, outperforming state-of-the-art models.
• Attained 98.78% accuracy on the Potato Leaf dataset, surpassing ResNet-18.
• Demonstrated superior performance with significantly reduced parameters and computational cost.

Conclusions:

• The proposed dual-branch CNN-ViT model effectively leverages both local and global features for precise crop disease identification.
• The lightweight design, incorporating ALP-FFN, offers an efficient solution for real-world agricultural applications.
• This approach provides a robust and scalable method for automated crop disease diagnosis.