Research and Implementation of Millet Ear Detection Method Based on Lightweight YOLOv5

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Millet ear detection is challenging due to dense growth, small size, and occlusion in field conditions.
• Existing target detection models require high computational power, limiting real-time deployment on mobile devices.

Purpose of the Study:

• To develop a lightweight and efficient real-time target detection method for millet ears.
• To improve the accuracy and robustness of millet ear detection in complex agricultural environments.

Main Methods:

• The YOLOv5s model was enhanced by replacing its backbone with the MobilenetV3 lightweight network.
• A micro-scale detection layer was added for improved detection of small and occluded targets.
• The Merge-NMS technique was employed in post-processing to mitigate boundary blur effects.

Main Results:

• The improved model achieved an AP value of 97.78% and an F1-score of 94.20%.
• Model size was reduced to 7.56 MB (53.28% of standard YOLOv5s), with improved detection speed.
• The model demonstrated strong robustness and generalization, outperforming classical models on Jetson Nano.

Conclusions:

• The lightweight YOLOv5 model effectively addresses challenges of dense distribution and occlusion in millet ear detection.
• The developed millet detection system, deployed on Jetson Nano, meets the requirements for intelligent agricultural machinery.