ISTDiffuser: Infrared small target image generation via conditional denoising diffusion with contrastive learning

Abstract

Infrared small target detection (IRSTD) technology is widely applied in modern military systems, such as missile guidance, night vision reconnaissance, target identification, and tracking. Due to the high costs in terms of time, labor, and finances, field experiments for IRSTD are limited and not comprehensive, making sample image collection challenging. Most publicly available datasets are synthesized and manually annotated, which is labor-intensive. Consequently, the number of images in these datasets is limited, and they fail to adequately cover the wide range of scene variations. To overcome these challenges, we propose ISTDiffuser, a novel single-frame infrared small target (SIRST) image generation method based on conditional diffusion and contrastive learning. This is the first time a diffusion model has been applied to SIRST image generation tasks, making this work pioneering in the field. In our approach, we introduce a scene-target interaction generation module that ensures the seamless integration of targets into their backgrounds, enhancing the realism and diversity of the synthetic data. Additionally, to better generate images consistent with the style of the instance images, we propose a style-type contrastive learning module. This module employs contrastive loss to guide the diffusion model, enhancing the realism and coherence of the generated images. Experimental results indicate that ISTDiffuser exhibits outstanding image generation capabilities. Furthermore, the generated images show high quality and stability in detection tasks, offering a promising solution to data scarcity in practice. The complete code will be made accessible via the repository at https://github.com/Tianzishu/istdiffuser.