CTRNet: a lightweight and efficient deep learning model for field maize whorl identification

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate maize whorl detection is difficult due to small target size, leaf occlusion, complex backgrounds, and variable lighting.
• Existing methods struggle with these challenges, impacting precision agriculture applications.

Purpose of the Study:

• To develop an efficient and robust maize whorl detection network.
• To enhance feature representation for small targets under occlusion and varying illumination.

Main Methods:

• Proposed the Contextual and Texture-enhanced Representation Network (CTRNet).
• Integrated multi-scale contextual interaction (MSCIM), dual-channel fine-grained feature enhancement (DCFEM), and gated adaptive information fusion (GAIFusion).
• Incorporated an adaptive channel-edge attention mechanism (ACEA) to handle background and illumination variations.

Main Results:

• CTRNet improved mAP@0.5 from 81.6% to 84.7% in complex field scenarios.
• Achieved significant enhancements in detection robustness for small targets under occlusion and varying lighting.
• The network is efficient and lightweight, with only 2.38 million parameters.

Conclusions:

• CTRNet offers an effective solution for maize whorl detection in challenging environments.
• The proposed network contributes to precision monitoring and targeted pesticide application in maize.
• This research advances automated agricultural management systems.