Improved YOLOv4 recognition algorithm for pitaya based on coordinate attention and combinational convolution.

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Accurate pitaya fruit recognition is crucial for developing automated picking systems.
• The complex spatial arrangement of pitaya fruits and branches presents a significant challenge for existing recognition methods.

Purpose of the Study:

• To propose an improved YOLOv4-based pitaya recognition method to address the challenges of fruit detection in natural environments.
• To enhance the accuracy and efficiency of pitaya recognition for automatic harvesting applications.

Main Methods:

• Utilized GhostNet as the backbone network for YOLOv4 to reduce model parameters and computational cost.
• Integrated coordinate attention mechanisms to improve the extraction of fine-grained target features.
• Designed an improved combinational convolution module and incorporated the Ghost Module into the YOLO Head to optimize feature extraction and processing speed.

Main Results:

• The improved YOLOv4 model achieved a pitaya recognition accuracy of 99.23% on a dataset of 8,800 images.
• Performance metrics included Precision (95.10%), Recall (98%), F1-score (98.94%), and Average Precision (AP).
• The model demonstrated a detection speed of 37.2 frames per second with a model weight size of 59.4MB.

Conclusions:

• The enhanced YOLOv4 model effectively meets the accuracy and speed requirements for pitaya fruit recognition in natural settings.
• This method provides robust technical support for the rapid and precise operation of autonomous fruit-picking robots.
• The proposed approach offers a viable solution for advancing agricultural automation in pitaya cultivation.