A simplified network topology for fruit detection, counting and mobile-phone deployment

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Fruit detection systems face challenges with complex network topologies, high computational costs, and large parameter counts.
• Environmental variability further complicates accurate and efficient fruit detection, tracking, and counting.

Purpose of the Study:

• To design a simplified network topology for fruit detection, tracking, and counting that addresses existing limitations.
• To develop a model that is robust, fast, accurate, and suitable for mobile deployment.

Main Methods:

• A novel backbone integrating Convolutional Neural Networks (Conv), Maxpooling (Maxpool), feature concatenation, and Spatial Pyramid Pooling (SPPF).
• A modified decoupled head based on YOLOv8 architecture for the detection and segmentation tasks.
• Validation on a diverse dataset including strawberry, jujube, and cherry fruit images.

Main Results:

• The simplified network exhibits significantly fewer parameters: 32.6% lower than YOLOv5n, 127% lower than YOLOv7-tiny, and 50.0% lower than YOLOv8n.
• Achieved a mean Average Precision (mAP@50%) of 82.4%, showing competitive accuracy against mainstream YOLO variants.
• Demonstrated superior speed, being 12.8% faster than YOLOv5n, 17.8% faster than YOLOv7-tiny, and 11.8% faster than YOLOv8n.

Conclusions:

• The simplified network offers a robust, fast, and accurate solution for fruit detection, tracking, and counting.
• Its reduced parameter count and improved efficiency make it highly deployable on mobile devices.
• The proposed architecture effectively overcomes the challenges associated with complex networks and environmental variability in fruit analysis.