MultiFuseYOLO: Redefining Wine Grape Variety Recognition through Multisource Information Fusion

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Traditional deep learning models struggle with wine grape variety recognition due to high similarity between varieties.
• Single-feature classification methods are insufficient for accurate differentiation.

Purpose of the Study:

• To develop a multisource information fusion method for enhanced wine grape variety recognition.
• To improve the accuracy and reliability of identifying visually similar grape varieties.

Main Methods:

• Optimization of the YOLOV7 model to create WineYOLO-RAFusion for improved fruit localization and recognition.
• Integration of multisource information fusion into WineYOLO-RAFusion, resulting in the MultiFuseYOLO model.
• Utilizing the SynthDiscrim algorithm as a core component for multisource information fusion.

Main Results:

• MultiFuseYOLO significantly outperformed existing models with precision, recall, and F1 scores of 0.854, 0.815, and 0.833, respectively.
• Precision for distinguishing Chardonnay and Sauvignon Blanc varieties increased substantially, from 0.512 to 0.813 and 0.533 to 0.775, respectively.
• The model demonstrated superior performance in identifying challenging, visually similar wine grape varieties.

Conclusions:

• The MultiFuseYOLO model provides a robust solution for wine grape variety identification.
• Multisource information fusion is critical for achieving high-precision recognition, particularly for similar varieties.
• This approach enhances the reliability of automated grape identification systems in viticulture.