CSNet: A Count-Supervised Network via Multiscale MLP-Mixer for Wheat Ear Counting.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Wheat yield is critical for global food security, making accurate ear counting essential for breeding and estimation.
• Current automated wheat ear counting methods require position-level annotations, which are labor-intensive and hinder deep learning adoption in agriculture.

Purpose of the Study:

• To develop an automated wheat ear counting method that utilizes count-level supervision instead of position-level annotation.
• To reduce the annotation cost and labor associated with training deep learning models for agricultural counting tasks.

Main Methods:

• Propose a count-supervised multiscale perceptive wheat counting network (CSNet).
• CSNet employs MLP-Mixer for a multiscale perception module with a global receptive field to learn attention maps for small targets without location information.
• Train and evaluate the network on a public wheat head detection dataset.

Main Results:

• The proposed CSNet, trained with count-supervised learning, outperforms existing position-supervised methods.
• CSNet achieves lower mean absolute error (MAE) and root mean square error (RMSE) compared to traditional methods.
• Demonstrates the effectiveness of using quantity information for accurate wheat ear counting.

Conclusions:

• The count-supervised strategy significantly reduces labeling costs while maintaining or improving counting accuracy.
• CSNet shows great potential for advancing automated counting in agriculture, particularly for wheat breeding and yield estimation.
• The developed method offers a practical solution for applying deep learning in agriculture by minimizing annotation requirements.