StyleGAN3-T: an alias-free generative framework for synthetic plant disease image augmentation and recognition

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate plant disease detection is crucial for sustainable agriculture.
• Deep learning models face challenges with small, imbalanced datasets, limiting generalization.
• Existing generative models may produce artifacts or lack spatial consistency.

Purpose of the Study:

• To develop a novel alias-free generative framework (StyleGAN3-T) for high-fidelity synthetic plant disease images.
• To integrate StyleGAN3-T with a hybrid Swin Transformer-ResNet50 classifier for precise disease recognition.
• To improve the robustness and generalization of deep learning models for plant disease detection.

Main Methods:

• Proposed StyleGAN3-T, a translation-equivariant, alias-free generative model.
• Developed a hybrid Swin Transformer-ResNet50 classifier.
• Created a balanced dataset of 18,000 images using real and StyleGAN3-T-generated samples.
• Conducted pooled Generative Adversarial Network (GAN) benchmarking.

Main Results:

• StyleGAN3-T demonstrated alias-free, artifact-free image synthesis with enhanced feature diversity.
• StyleGAN2-ADA showed stronger generative-quality metrics in benchmarking.
• StyleGAN3-T was chosen as the preferred augmentation model due to superior downstream classification performance.
• The combined dataset and pipeline achieved superior classification accuracy.

Conclusions:

• StyleGAN3-T effectively generates high-fidelity synthetic plant disease images, addressing data limitations.
• The alias-free synthesis and spatial consistency of StyleGAN3-T are critical for improved downstream classification tasks.
• This approach offers a promising solution for enhancing plant disease detection in sustainable agriculture.