FIDMT-GhostNet: a lightweight density estimation model for wheat ear counting.

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Wheat production is crucial for global food security and economic stability.
• Accurate wheat ear counting is vital for agricultural management, yield prediction, and resource allocation.
• Existing deep learning methods face challenges due to complex backgrounds, dense and small wheat ear targets.

Purpose of the Study:

• To develop an automatic wheat ear positioning and counting method robust to field complexities.
• To improve the accuracy and efficiency of wheat ear counting for better agricultural insights.

Main Methods:

• Proposed FIDMT-GhostNet, a lightweight network using GhostNet for multi-scale feature extraction.
• Integrated FIDMT (focal inverse distance transform maps) for improved accuracy with dense wheat ears.
• Introduced dense upsampling convolution to enhance feature extraction for small wheat ear targets.
• Implemented a local maximum value detection strategy to handle background noise and interference.

Main Results:

• The FIDMT-GhostNet model achieved a wheat ear counting accuracy of 0.9145.
• The model has 8.42 million parameters, indicating efficiency.
• Experiments on WEC, WEDD, and GWHD datasets validated the model's effectiveness.

Conclusions:

• The FIDMT-GhostNet model demonstrates strong performance in automatic wheat ear counting.
• This method offers a significant advancement for precision agriculture and crop monitoring.
• The model's accuracy and efficiency contribute to more reliable yield prediction and management.