A precise berry counting method for in-cluster grapes to guide berry thinning

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Manual berry counting in table grape production is labor-intensive, slow, and error-prone.
• Accurate berry counts are crucial for effective cluster management and yield optimization.
• Current manual methods present a bottleneck in efficient and precise vineyard management.

Purpose of the Study:

• To develop an automated method for accurate berry counting in table grapes.
• To introduce a dual-branch network (MVDNet) for simultaneous berry counting and bunch segmentation.
• To create a computationally efficient model suitable for resource-constrained edge devices.

Main Methods:

• Proposed a dual-branch network (MVDNet) with UIB modules, multi-scale feature fusion, and SimAM attention.
• MVDNet performs density map regression for counting and bunch segmentation.
• A post-processing algorithm refines per-cluster berry counts.

Main Results:

• MVDNet achieved a Mean Absolute Error (MAE) of 7.7 and Root Mean Square Error (RMSE) of 12.6.
• The model attained a Mean Intersection Over Union (MIoU) of 0.90 for bunch segmentation.
• The post-processing algorithm achieved a coefficient of determination (R²) of 0.886 for per-cluster counting.
• The model has only 3.372 million parameters, indicating low computational resource consumption.

Conclusions:

• The proposed automated berry counting method is robust, efficient, and practical.
• This technology facilitates precise vineyard management, enhancing grape quality and productivity.
• The MVDNet model is suitable for deployment on edge devices in agricultural settings.