Deep learning for image-based large-flowered chrysanthemum cultivar recognition.

Area of Science:

• Horticultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Cultivar recognition is crucial for flower production, research, and commerce.
• Chinese large-flowered chrysanthemum (Chrysanthemum × morifolium Ramat.) possesses high ornamental value and cultural significance.
• Challenges in chrysanthemum cultivar recognition include complex capitulum structure, diverse floret types, and numerous cultivars.

Purpose of the Study:

• To explore the application of deep learning methods for chrysanthemum cultivar recognition.
• To develop and evaluate deep learning models for accurate identification of Chrysanthemum × morifolium cultivars.

Main Methods:

• Proposed deep learning models utilizing VGG16 and ResNet50 architectures.
• Collected and utilized two datasets: Dataset A (14,000 images, 103 cultivars) for training and Dataset B (197 images, different years) for generalization testing.
• Employed gradient-weighted class activation mapping (Grad-CAM) and feature clustering for model interpretability.

Main Results:

• Achieved high calibration accuracy (Top-5 rate >98%) on Dataset A.
• Demonstrated strong model generalization performance (Top-5 rate >78%) on Dataset B.
• Identified key visual features influencing model decision-making, such as inflorescence edge, disc floret areas, color, and shape.

Conclusions:

• Deep learning represents a breakthrough in horticultural science for cultivar recognition, offering superior performance and speed.
• The models effectively recognize chrysanthemum cultivars, with key factors like inflorescence characteristics influencing decisions.
• The findings provide a foundation for automated and efficient chrysanthemum cultivar identification systems.