CATransU-Net: Cross-attention TransU-Net for field rice pest detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Agricultural Technology

Background:

• Accurate detection of rice pests is crucial for effective field pest control.
• Existing models like U-Net excel at local features, while Transformers handle long-range dependencies.
• Integrating these architectures offers potential for improved pest detection.

Purpose of the Study:

• To develop an advanced deep learning model for accurate paddy pest detection.
• To combine the strengths of U-Net and Transformer architectures for enhanced feature extraction.
• To improve the precision and efficiency of automated rice pest identification systems.

Main Methods:

• A novel Cross-Attention TransU-Net (CATransU-Net) model was designed, integrating U-Net and Transformer.
• The architecture features an encoder-decoder structure with a dual Transformer-attention module (DTA) and cross-attention skip-connection (CASC).
• Dilated residual Inception (DRI) was used in the encoder for multiscale feature extraction, DTA for nonlocal interactions, and CASC for multi-resolution representation.

Main Results:

• CATransU-Net demonstrated superior performance in rice pest extraction on IP102 and AgriPest datasets.
• The model achieved a precision of 93.51%, outperforming other methods by approximately 2% and U-Net by 9.36%.
• The enhanced feature representation through DRI, DTA, and CASC contributed to high-resolution insect image generation.

Conclusions:

• The proposed CATransU-Net model effectively addresses the challenge of accurate rice pest detection in field conditions.
• The integration of attention mechanisms and multiscale feature extraction significantly enhances detection performance.
• CATransU-Net shows strong potential for application in practical field rice pest detection systems.