Detection of Anomalous Grapevine Berries Using Variational Autoencoders

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Grapevine cultivation is economically significant, necessitating crop monitoring for high-quality yields.
• Identifying unhealthy berries (damaged or diseased) is crucial for quality control before harvest.
• Manual inspection of grapevines is labor-intensive and costly, driving the need for automated solutions.

Purpose of the Study:

• To develop an automated, image-based machine learning approach for detecting damaged grapevine berries.
• To provide farmers with a tool that highlights anomalous areas on plants, reducing the need for manual monitoring.
• To compare the performance of a novel method against existing approaches for berry damage detection.

Main Methods:

• Training a fully convolutional variational autoencoder (FC-VAE) using only images of healthy grapevine berries.
• Utilizing a feature perceptual loss function during model training.
• Identifying damaged berries as image areas deviating from the healthy berry model.
• Generating heatmaps to visualize the detection results for farmer decision support.

Main Results:

• The proposed FC-VAE method effectively identifies deviations indicative of damaged berries.
• Visualizations (heatmaps) clearly highlight anomalous areas, aiding in targeted inspection.
• The developed approach demonstrated superior performance compared to a convolutional autoencoder (CAE) previously used for similar tasks.

Conclusions:

• The automated image-based machine learning approach offers an efficient and cost-effective solution for detecting damaged grapevine berries.
• This technology can significantly assist farmers in quality control and crop management.
• The FC-VAE with feature perceptual loss presents a promising advancement in automated agricultural monitoring.