Year: 2024
Type: conference-paper
Abstract: In the realms of computer vision and image processing, image-to-image translation is pivotal to data augmentation. Due to the high cost of hardware for direct registration of optical and infrared images and the limitations of conventional model-driven approaches in cross-domain registration, this paper focuses on generating infrared images from optical images using generative models. Existing image translation methods often treat target domain images as conditions, leading to inefficiencies. We propose a diffusion-based method for translating optical images to infrared. Using VQ-VAE for latent space formation and a bidirectional stochastic Brownian Bridge diffusion process for modal alignment, we validate its effectiveness through extensive simulations, showing superior performance compared to existing methods.
DOI: 10.1117/12.3036995
Authors: Yuxin Zhao, Zihao Gao, Huaizhang Liao, Tao Han, Jingyuan Xia
Institution: National University of Defense Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4404992362