A Tiny Object Detection Approach for Maize Cleaning Operations.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Efficient grain cleaning requires real-time impurity detection for dynamic parameter adjustment.
• Traditional methods may lack the precision for identifying diverse impurities in maize.

Purpose of the Study:

• To develop a specialized object detection network for identifying and localizing impurities in maize during cleaning operations.
• To improve the accuracy and efficiency of impurity removal in harvested maize.

Main Methods:

• Utilized EfficientNetB7 as the backbone for a Faster Region Convolutional Neural Network (Faster R-CNN).
• Integrated a cross-stage feature integration mechanism for multi-scale feature learning.
• Developed an adaptive region proposal network (ARPN) for precise detection of small objects.

Main Results:

• The proposed model demonstrated effectiveness in detecting and classifying impurities in maize.
• Ablation experiments validated the contribution of each improved component in the detection network.
• The system achieved accurate localization of tiny impurity objects.

Conclusions:

• The developed object detection network enhances real-time maize cleaning processes.
• The integration of EfficientNetB7 and ARPN significantly improves impurity detection accuracy.
• This technology offers a pathway to reduced grain loss and optimized cleaning strategies.