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

Abstract

To address this challenge, we propose StyleGAN3-T, the translation-equivariant alias-free variant of StyleGAN3, as a generative framework for producing high-fidelity synthetic plant disease images, integrated with a hybrid Swin Transformer-ResNet50 classifier for precise recognition. Accurate detection of plant leaf diseases is essential for sustainable agriculture and early intervention. However, deep learning models often struggle with small, imbalanced datasets that limit generalization and robustness. To address this challenge, we propose StyleGAN3-T, a novel alias-free generative framework for producing high-fidelity synthetic plant disease images, integrated with a hybrid Swin Transformer-ResNet50 classifier for precise recognition. The proposed approach ensures translation-equivariant, artifact-free image synthesis and enhanced feature diversity. A balanced dataset of 18,000 images was developed by combining real and StyleGAN3-T-generated samples. In pooled GAN benchmarking, StyleGAN2-ADA achieved the strongest generative-quality metrics, whereas StyleGAN3-T was selected as the preferred augmentation model because its alias-free synthesis and spatial consistency yielded superior downstream classification performance in the proposed pipeline.

Plant phenotyping relevance

植物病害画像を合成・認識する画像解析手法が研究の中心であり、植物の病害状態を画像から推定するフェノタイピング手法に該当する。

Code and data availability

The paper's grape leaf disease image inputs are two publicly available Kaggle datasets explicitly named in the Data Availability statement. No author code, models, or synthetic dataset deposit is provided; other processed data is request-only.