RE-YOLO: An apple picking detection algorithm fusing receptive-field attention convolution and efficient multi-scale...

Area of Science:

• Computer Vision
• Robotics
• Agricultural Technology

Background:

• Robotic apple picking requires efficient and accurate detection algorithms.
• Current algorithms struggle with dense fruit distribution and occlusion.
• Deploying high-accuracy models on resource-limited edge devices is challenging.

Purpose of the Study:

• To develop an improved apple detection algorithm (RE-YOLO) for robotic picking.
• To enhance accuracy and efficiency in dense and occluded scenarios.
• To enable deployment on edge devices with limited computational resources.

Main Methods:

• Introduced Receptive-Field Attention Convolution (RFAConv) to the backbone and neck network.
• Developed an EMA_C2f module for uniform spatial semantic feature distribution and improved occlusion discrimination.
• Integrated the Wise Intersection over Union (WIOU) loss function for accelerated optimization and better detection of varied target sizes.

Main Results:

• RE-YOLO demonstrated significant improvements in precision (+2%), recall (+2.1%), mAP@0.5 (+2.7%), and mAP@0.5-0.95 (+3.9%) over YOLOv8.
• Compared to YOLOv5, RE-YOLO showed improvements of +4% in precision, +1.9% in recall, +1.7% in mAP@0.5, and +3% in mAP@0.5-0.95.
• The algorithm effectively addresses challenges of dense fruit distribution and occlusion.

Conclusions:

• RE-YOLO offers an advanced and practical solution for apple detection in robotic picking.
• The proposed improvements enhance model recognition accuracy and computational efficiency.
• The algorithm is suitable for deployment on edge devices, advancing automated harvesting.