Infrared target generation based on conditional diffusion inpainting

A novel method for generating infrared targets is proposed, based on conditional diffusion inpainting. Firstly, we introduce a transformation framework designed to translate detection annotations from text to image, serving as the conditional input for the diffusion inpainting convolutional neural network. Secondly, a novel multi-scale image-condition Unet is designed as the mainframe of diffusion model. Finally, we conduct numerous direct and indirect evaluation experiments to assess the proposed algorithm. Experimental results demonstrate that the proposed algorithm generates high-quality infrared targets. Furthermore, as an augmentation, the generated images significantly enhance the detection accuracy of few-shot thermal infrared targets.