Multi-stage tomato fruit recognition method based on improved YOLOv8.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Accurate, multi-stage tomato identification and localization are crucial for facility agriculture.
• Complex environments pose challenges for conventional detection methods, necessitating advanced solutions.

Purpose of the Study:

• To propose a novel YOLOv8-EA model for improved tomato fruit localization and identification.
• To enhance feature extraction, inference speed, and detection accuracy in challenging agricultural settings.

Main Methods:

• Replaced YOLOv8 backbone with EfficientViT for reduced parameters and enhanced feature extraction.
• Introduced C2f-Faster module by integrating convolutions into the C2f module to optimize inference.
• Modified bounding box loss to SIoU for accelerated convergence and improved detection.
• Incorporated Auxiliary Detection Head (Aux-Head) module to boost network learning capacity.

Main Results:

• YOLOv8-EA achieved 91.4% accuracy, 88.7% recall, and 93.9% average precision at 163.33 FPS on a custom dataset.
• Demonstrated significant improvements over baseline YOLOv8n, with accuracy, recall, and AP enhanced by 10.9%, 11.7%, and 7.2%, respectively.
• Showcased strong generalization on a public dataset with high accuracy and improved detection speed.

Conclusions:

• The YOLOv8-EA model reliably identifies and localizes multi-stage tomatoes in complex environments.
• Provides a robust technical foundation for developing intelligent tomato-picking devices.