Generation of microwave images for sensitive applications with diffusion models

The surge in artificial intelligence advancements has led to remarkable progress in image generative models, which are constantly being applied to new domains and fields of study. This work delves into the usage of Diffusion Models(DMs) in generating intricate synthetic samples in the domain of microwave imaging for security applications. The research explores various techniques that can be applied to image generation with DMs, starting from generating random images and gradually incorporating control mechanisms such as text conditioning and element specification. The main objective of such an activity is to generate synthetic datasets used to augment the training of detection models for security body scanners. The results demonstrate the robustness of diffusion models across diverse architectural and hyperparameter choices. Although synthetic data proved less informative than real data for training detection models, a combination of both enhanced performance. This project sheds light on the potential of synthetic data augmentation in scenarios where data scarcity hampers conventional machine learning approaches, emphasizing the relationship between human guidance and generative models.