A lightweight Yunnan Xiaomila detection and pose estimation based on improved YOLOv8

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Yunnan Xiaomila peppers present detection challenges due to low contrast with complex backgrounds and small target sizes.
• Existing target detection models struggle with subtle feature information and color gradients, impacting accuracy.

Purpose of the Study:

• To develop an improved PAE-YOLO model for accurate and efficient detection of Yunnan Xiaomila peppers in challenging environments.
• To enhance feature extraction and inference speed while maintaining a lightweight model architecture.

Main Methods:

• Integrated the EMA attention mechanism with the C2f module in YOLOv8 for improved feature expression.
• Incorporated DCNv3 deformable convolution in the backbone and head networks for adaptive feature capture.
• Utilized depth camera data for posture estimation and occlusion analysis.

Main Results:

• Achieved an average mean accuracy (mAP) of 88.8%, a 1.3% improvement over the original model.
• Obtained the best F1 score (83.2) among compared networks, with reduced model size (5.7MB) and GFLOPs (7.6G).
• Demonstrated over 85% correct orientation estimation for Xiaomila targets, with an average error angle of 15.91°.

Conclusions:

• The improved PAE-YOLO model effectively detects Yunnan Xiaomila peppers with high accuracy and low computational complexity.
• The model shows robust performance in estimating target posture, even under occlusion.
• This advancement contributes to more precise agricultural automation and crop monitoring.