Paper Title: HiClassGen: High-Resolution Image Augmentation with Class and Shape Controllable Diffusion Models
Conference Name: International Conference on Artificial Intelligence in Information and Communication (ICAIIC)
Abstract: Deep learning models generally require large quan-tities of data to ensure effective generalization and capture complex patterns in diverse scenarios. However, acquiring sufficient labeled data is often costly and time-consuming in many domains, such as medical imaging, industrial defects, remote sensing, and other specialized fields. To address this challenge, we present HiClassGen, a novel image augmentation framework that leverages class- and shape-controllable diffusion models to generate high-quality, domain-specific synthetic data. HiClassGen integrates class-based semantic guidance and shape constraints during the image generation process, enabling precise control over both the augmented images' class labels and morphological features. This customization ensures that the synthetic data generated aligns with the specific needs of the target task, leading to better performance and generalization. Additionally, HiClass-Gen automatically generates corresponding annotations for the synthetic data, ensuring seamless integration into supervised learning pipelines. Experiments demonstrate the effectiveness of our proposed model, showing a 10% increase in accuracy on benchmark datasets.