Instance Segmentation and Berry Counting of Table Grape before Thinning Based on AS-SwinT

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Grape berry thinning is crucial for high-quality table grapes but faces labor shortages due to an aging population.
• Automated berry thinning requires precise machine vision systems for counting and locating berries.
• Existing methods struggle with the repetitive and labor-intensive nature of manual thinning.

Purpose of the Study:

• To develop an intelligent berry-thinning machine to reduce manual labor.
• To propose a novel instance segmentation and berry counting method for grape bunches.
• To enhance the precision and efficiency of automated table grape management.

Main Methods:

• Utilized Swin Transformer as a backbone for feature extraction in the AS-SwinT model.
• Implemented adaptive feature fusion in the neck network to improve detection of small and occluded berries.
• Optimized anchor scales and employed Soft-NMS for accurate counting of densely packed berries.

Main Results:

• The AS-SwinT model achieved superior performance metrics (e.g., 65.7 AP) compared to Mask R-CNN variants.
• Achieved high precision in berry counting with RMSE of 7.13 and R² of 0.95.
• Demonstrated significant advantages over existing models in berry counting estimation.

Conclusions:

• The proposed AS-SwinT method offers a robust solution for automated berry counting in table grapes.
• This technology can facilitate the development of intelligent berry-thinning machines, addressing agricultural labor challenges.
• The model's high accuracy and efficiency support precision agriculture and sustainable grape production.