Easy domain adaptation method for filling the species gap in deep learning-based fruit detection.

Area of Science:

• Computer Vision
• Machine Learning
• Horticulture Technology

Background:

• Deep learning models for fruit detection are crucial in modern agriculture.
• Current models require extensive manual labeling for new fruit species, which is time-consuming and labor-intensive.
• A need exists for methods that transfer existing models to new domains without manual data annotation.

Purpose of the Study:

• To propose a domain adaptation method for transferring fruit detection models between different species without manual labeling.
• To address the challenge of species-specific model performance in deep learning-based fruit detection.
• To reduce the cost and effort associated with creating new training datasets for diverse fruit types.

Main Methods:

• Utilized CycleGAN to transform source domain fruit images (labeled) into the target domain (unlabeled).
• Implemented a pseudo-labeling process to automatically annotate target domain images.
• Employed a pseudo-label self-learning approach to enhance the accuracy of generated labels.
• Evaluated the method using a labeled orange dataset (source) and unlabeled apple and tomato datasets (target).

Main Results:

• Achieved high performance in fruit detection on target domains without manual labeling.
• Mean average precision reached 87.5% for apple detection and 76.9% for tomato detection.
• Demonstrated the effectiveness of CycleGAN, pseudo-labeling, and self-learning in domain adaptation for fruit detection.

Conclusions:

• The proposed domain adaptation method successfully transfers fruit detection models to new species without manual labeling.
• This approach effectively bridges the species gap in deep learning-based fruit detection systems.
• The method offers a practical solution for efficient and scalable fruit detection in diverse horticultural applications.