In-Field Automatic Identification of Pomegranates Using a Farmer Robot.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Precision agriculture relies on vision-based systems for orchard monitoring.
• Traditional deep learning for fruit detection requires extensive manual image annotation.

Purpose of the Study:

• To develop a semi-supervised deep learning framework for automatic pomegranate detection.
• To reduce the labor-intensive nature of image labeling in agricultural computer vision tasks.

Main Methods:

• A semi-supervised deep learning framework utilizing a multi-stage transfer learning approach.
• Fine-tuning a pre-trained network with controlled fruit images, then extending to field conditions.
• Implementation using the DeepLabv3+ (Resnet18) architecture.

Main Results:

• The framework achieved high accuracy in pomegranate segmentation.
• An F1-score of 86.42% and an IoU of 97.94% were obtained in field tests.
• The method significantly reduced the need for manual image annotation.

Conclusions:

• The proposed semi-supervised framework enables accurate and efficient pomegranate detection in orchards.
• This approach alleviates the burden of manual labeling, making precision agriculture more accessible.
• The multi-stage transfer learning strategy is effective for adapting models to complex field scenarios.