Enhancing Few-Shot Plant Disease Classification with Diffusion Model

Plant diseases significantly threaten global agriculture, impacting crop yield and quality, thus affecting food security and economic stability. Traditional detection methods, reliant on expert knowledge and manual inspection, are time-consuming and error-prone. Few-shot classification methods, including meta-learning, transfer learning, and data augmentation, have gained attention. Meta-learning adapts models quickly across tasks but requires substantial computational resources. Transfer learning enhances related task performance, reducing data needs but risking negative transfer. Traditional data augmentation, like image flipping and rotation, offers limited effectiveness. We propose an advanced generative model-based plant disease data augmentation method using Stable Diffusion models to generate high-quality synthetic data, integrated with ControlNet and Embbedings for precise control, and fine-tuned by Low-Rank Adaptation (LoRA). This method overcomes traditional limitations by generating diverse, high-quality samples, enhancing dataset diversity and model generalization, thereby facilitating robust and accurate plant disease detection model development.