Automatic detection and counting of wheat spike based on DMseg-Count

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Accurate wheat spike counting is vital for yield prediction and variety evaluation.
• Challenges in wheat spike image analysis include poor lighting, occlusion, and overlap, reducing counting accuracy.
• Existing methods struggle with complex field conditions.

Purpose of the Study:

• To improve the accuracy of automatic wheat spike counting in challenging environments.
• To enhance wheat yield prediction and variety evaluation through advanced image analysis.
• To develop a robust deep learning model for wheat spike detection and counting.

Main Methods:

• An improved wheat spike counting model, DMseg-Count, was developed based on the DM-Count model.
• A local segmentation branch was introduced to extract local contextual information.
• A point multiplication mechanism fused global and local contextual information, optimized by a constructed loss function.

Main Results:

• The DMseg-Count model achieved a Mean Absolute Error (MAE) of 5.79 and Root Mean Square Error (RMSE) of 7.54.
• These results represent significant improvements over the standard DM-Count model.
• The model demonstrated superior performance in detecting wheat spikes under difficult imaging conditions.

Conclusions:

• The proposed DMseg-Count model effectively detects and counts wheat spikes in complex field environments.
• This provides a novel approach for automated wheat spike counting and yield prediction.
• The model exhibits enhanced computer vision capabilities for agricultural applications.