A Diffusion-based Data Augmentation Method for SAR Object Detection

Abstract

In recent times, diffusion models have garnered considerable attention as a key technology for generating images from text prompts. These models are especially recognized for their ability to produce high-resolution images with impressive accuracy. Another notable feature of diffusion models is their flexibility, as they can be easily fine-tuned through additional training to suit specific needs. This adaptability has made them useful in various tasks, one of which is data augmentation. In this study, we introduce a new data augmentation technique that enhances the training efficiency of object detection models. Specifically, our approach utilizes a diffusion model-based conditional generation method to create Synthetic Aperture Radar(SAR) images by taking input bounding boxes. This innovative approach demonstrates how diffusion models can effectively boost object detection performance.