Combating data incompetence in pollen images detection and classification for pollinosis prevention

Area of Science:

• Botany
• Computer Science
• Allergy and Immunology

Background:

• Accurate pollen recognition is vital for preventing and treating pollinosis (hay fever).
• Deep learning models require extensive data, but existing pollen datasets are small.
• This research addresses the challenge of pollen recognition with limited data.

Purpose of the Study:

• To create a novel, open pollen image dataset annotated for both detection and classification.
• To investigate state-of-the-art deep learning approaches for learning from small pollen datasets.
• To develop and evaluate advanced models for pollen detection and classification.

Main Methods:

• A new Bayesian RetinaNet model was developed to incorporate aleatoric uncertainty for pollen detection.
• Convolutional Neural Network (CNN) and Siamese Neural Network classifiers were pre-trained on synthetic pollen images generated by Generative Adversarial Networks (GANs).
• The performance of models was evaluated on a newly created, annotated pollen dataset.

Main Results:

• The Bayesian RetinaNet achieved higher detection precision compared to the baseline RetinaNet.
• A CNN classifier pre-trained on StyleGAN-generated synthetic images demonstrated superior performance for pollen classification.
• The best models achieved 96.3% mean average precision for detection and 97.7% F1 score for classification across 13 pollen species.

Conclusions:

• The developed pollen dataset and deep learning models significantly advance automated pollen recognition.
• Learning from small datasets using GAN-generated synthetic data and Bayesian networks is effective for pollen analysis.
• The findings contribute to improved pollinosis prevention and treatment strategies through enhanced pollen identification.