A lightweight grape detection model in natural environments based on an enhanced YOLOv8 framework

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Automated grape harvesting requires accurate fruit and stem detection.
• Dense fruit arrangements and similar stem/branch colors challenge existing models.
• Current models are often too large and slow for mobile deployment.

Purpose of the Study:

• To develop a lightweight and efficient model for grape detection in automated harvesting.
• To address challenges of dense fruit clusters and similar stem/branch coloration.
• To improve detection speed and reduce computational load for mobile applications.

Main Methods:

• Proposed TiGra-YOLOv8 model based on YOLOv8n.
• Integrated Attentional Scale Fusion (ASF) module for feature extraction.
• Employed Adaptive Training Sample Selection (ATSS) for improved sample quality.
• Utilized Weighted Interpolation of Sequential Evidence for Intersection over Union (Wise-IoU) loss function.
• Reduced model size via channel pruning.

Main Results:

• TiGra-YOLOv8 achieved a 3.33% increase in mAP(0.5) over YOLOv8n.
• Detection speed improved by 7.49% (FPS).
• Parameter count reduced by 52.19%, and computational demand decreased by 51.72%.
• Model size was reduced by 45.76%.

Conclusions:

• TiGra-YOLOv8 enhances detection accuracy for challenging grape targets.
• The model offers significant reductions in parameters and computational requirements.
• This lightweight model is suitable for deployment on mobile devices for automated grape harvesting.