Multi-Information Model for Large-Flowered Chrysanthemum Cultivar Recognition and Classification.

Area of Science:

• Horticulture
• Computer Science
• Plant Taxonomy

Background:

• The Chinese large-flowered chrysanthemum (Chrysanthemum × morifolium Ramat.) exhibits significant morphological diversity across numerous cultivars.
• Existing classification systems based on comparative morphology struggle with accurate recognition of many cultivars.
• Automated classification methods are needed to address the challenges posed by this morphological variation.

Purpose of the Study:

• To develop and evaluate a deep learning-based multi-information model for recognizing and classifying large-flowered chrysanthemum cultivars.
• To assess the effectiveness of image features in representing chrysanthemum characteristics, such as flower color and petal type.
• To investigate the impact of training parameters on model performance for improved automated classification.

Main Methods:

• Collected images of 213 large-flowered chrysanthemum cultivars over two years (2018-2019).
• Constructed a non-pre-trained ResNet18 deep learning model using the 2018 dataset.
• Utilized Affinity Propagation (AP) clustering and Principal Component Analysis (PCA) to analyze image features.

Main Results:

• The developed model achieved 70.62% top-5 test accuracy on the 2019 dataset.
• Image features effectively discriminated flower colors (AP clustering) and petal types (PCA).
• A non-pre-trained model demonstrated advantages over ImageNet pre-trained models by not ignoring color information.

Conclusions:

• Deep learning models, particularly non-pre-trained ones, show promise for automated chrysanthemum cultivar classification.
• Image feature analysis provides valuable insights into cultivar discrimination, supporting morphological studies.
• This research provides a foundation for developing robust automated systems for chrysanthemum cultivar identification.