A density map-based method for counting wheat ears.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate wheat ear counting is vital for estimating crop yield and supporting agricultural management.
• Existing methods face challenges in accurately counting dense wheat ears in field conditions.

Purpose of the Study:

• To develop an improved method for rapid and effective wheat ear counting in field environments.
• To enhance the accuracy and efficiency of dense wheat counting using a modified YOLACT framework.

Main Methods:

• The proposed method is based on the YOLACT framework, incorporating a GeM pooling layer within the CBAM module.
• A density map is introduced into the Feature Pyramid Network (FPN) to improve performance in dense scenarios.

Main Results:

• The improved attention mechanism reduced the Root Mean Square Error (RMSE) from 1.75 to 1.57.
• Pixel-level density estimation and the density map mechanism enhanced the coefficient of determination (R²) from 0.9615 to 0.9798.
• The model accurately discerns overlapping count targets, providing more granular information.

Conclusions:

• The developed framework demonstrates significant improvements in wheat ear counting performance, especially in complex backgrounds.
• The findings highlight the practical potential of this approach for intelligent agriculture applications and reliable food supply chain management.