GCNet: A Deep Learning Framework for Enhanced Grape Cluster Segmentation and Yield Estimation Incorporating Occluded...

Area of Science:

• Agricultural Imaging
• Computer Vision
• Machine Learning

Background:

• Object segmentation algorithms for grape yield estimation are limited by their inability to count occluded grapes.
• Hidden grapes due to cluster compactness or canopy interference lead to inaccurate yield predictions.

Purpose of the Study:

• To develop a novel framework, Grape Counting Network (GCNet), for accurate grape cluster segmentation and yield estimation.
• To address the challenge of counting occluded grapes in indoor agricultural imaging.

Main Methods:

• Proposed the Grape Counting Network (GCNet), integrating deep learning with correction factors for occlusion adjustments.
• Developed GrapeSet, a new dataset with indoor grape cluster imagery and ground truth data.
• Utilized advanced object segmentation techniques for improved grape analytics.

Main Results:

• GCNet achieved a R² of 0.96 and reduced mean absolute error (MAE) by 10% compared to previous methods.
• Demonstrated enhanced segmentation accuracy under challenging conditions like foliage and cluster compactness.
• Successfully addressed the limitation of counting only visible grapes.

Conclusions:

• GCNet sets new standards in agricultural indoor imaging analysis for grape yield estimation.
• The framework provides accurate object segmentation for grape analytics, aiding early yield prediction for harvesters.
• Encourages future research into grape feature analysis for yield estimation.