Metric learning for image-based flower cultivars identification.

Area of Science:

• Plant science
• Computer vision
• Machine learning

Background:

• Deep learning in plant phenotyping shows promise for plant breeding.
• Traditional deep learning struggles with flower cultivar identification due to limited sample data for numerous cultivars.
• Metric learning offers a solution for accurate identification with insufficient data.

Purpose of the Study:

• To develop an efficient metric learning method for flower cultivar identification.
• To address the challenge of limited sample data in identifying a large number of flower cultivars.
• To improve recognition accuracy and feature interpretability in flower identification tasks.

Main Methods:

• Implemented a metric learning approach incorporating center loss for sample dispersion and compactness.
• Utilized ResNet18, ResNet50, and DenseNet121 for feature extraction.
• Employed joint supervision of center loss and L2-softmax loss for model training.

Main Results:

• Achieved high test accuracy rates of 91.88%, 97.34%, and 99.82% on three distinct datasets.
• Verified the method's effectiveness through T-distributed stochastic neighbor embedding (T-SNE) visualization of feature distributions.
• Demonstrated improved inter-class sample dispersion and intra-class sample compactness.

Conclusions:

• The proposed metric learning method provides an efficient solution for flower cultivar identification with high recognition rates.
• The extracted features are interpretable, offering new insights for identification tasks with limited data.
• This research holds significant reference value for advancing flower cultivar identification.