High-precision object detection network for automate pear picking.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Increasing agricultural output and labor shortages necessitate agricultural intelligence.
• Existing deep learning object detection methods struggle with background clutter and fruit occlusion in pear orchards.
• Current methods lack the precision required for complex automated pear picking tasks.

Purpose of the Study:

• To propose a high-precision object detection network for automated pear picking.
• To enhance detection accuracy by addressing background redundancy and fruit occlusion.
• To develop a system meeting the demands of complex automated agricultural tasks.

Main Methods:

• Developed High-level deformation-perception Network with multi-object search NMS (HDMNet), based on YOLOv8.
• Incorporated a high-level semantic-focused attention mechanism to filter background information.
• Utilized a deformation-perception feature pyramid network for improved detection of distant and small fruits.
• Implemented multi-object search non-maximum suppression for effective multi-pear detection.

Main Results:

• HDMNet achieved a mean Average Precision (mAP) of 75.7% and mAP50 of 93.6%.
• The network demonstrates high efficiency with 73.0 Frames Per Second (FPS) and low computational cost (41.1 GFLOPs).
• HDMNet has a low parameter count (12.9 M), outperforming state-of-the-art methods in real-time detection, precision, and positioning.

Conclusions:

• HDMNet offers a significant advancement in object detection for automated pear harvesting.
• The network's design effectively handles challenges like background noise and fruit occlusion.
• HDMNet provides a computationally efficient and highly accurate solution for agricultural robotics.