FastQAFPN-YOLOv8s-Based Method for Rapid and Lightweight Detection of Walnut Unseparated Material.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Walnuts are valuable crops, but efficient automated sorting of shells and kernels is challenging.
• Existing object detection models may not meet the speed and size requirements for real-time automated sorting systems.

Purpose of the Study:

• To develop a lightweight and fast object detection network for precise identification of unsorted walnuts.
• To enhance the efficiency of automated walnut processing through rapid and accurate shell and kernel detection.

Main Methods:

• Proposed a novel FastQAFPN-YOLOv8s network utilizing lightweight Pconv operators and a FasterNextBlock backbone.
• Integrated an ECIoU loss function for faster network regression and prediction frame adjustment.
• Employed a QAFPN feature fusion network with a Rep-PAN structure for efficient feature extraction and fusion.

Main Results:

• The FastQAFPN-YOLOv8s model achieved a Mean Average Precision (mAP) of 94.5% at 52.1 Frames Per Second (FPS).
• Demonstrated significant reductions in parameters (45.5%), training time (32.7%), and model size (12.3 MB) compared to the original YOLOv8s.
• Outperformed YOLOv7 and YOLOv6 in model size reduction and FPS, while maintaining comparable or improved accuracy.

Conclusions:

• The FastQAFPN-YOLOv8s network offers an effective solution for reducing model size and inference time in walnut sorting.
• The proposed lightweight design balances performance and speed, making it suitable for real-time automated agricultural applications.
• This advancement contributes to more efficient and accurate automated processing in the nut industry.