High Quality Dynamic Occlusion Computational Ghost Imaging Guided Through Conditional Diffusion Model

High-quality imaging under low sampling is the key to the practical application of computational ghost imaging, and significant progress has been made. However, as far as we know, most studies focus on ideal scene imaging, and research on computational ghost imaging under dynamic occlusion inter-ference has not been mentioned. In this work, we propose a high-quality dynamic occlusion computational ghost imaging method guided by a conditional diffusion model. This method can extract target features through one-dimensional bucket signals when the target object is occluded and then directly output high-quality reconstruction results. By embedding the CGI process as an imaging layer into the network in a conditional input manner, an Ada-ResBlock module is proposed to replace the ResBlock of the traditional U-Net, and the time step ${\bm{t}}$ of the diffusion model is embedded in the normalization of the Ada-ResBlock using the Deep-emb method, which improves the accuracy of the noise prediction. Experiments show that this method has a significant improvement over the traditional conditional diffusion model. It can not only have better reconstruction capabilities under low sampling rates and 30% occlusion area of the target, but also maintain stability when the target is occluded by obstacles at different angles. This method helps promote the practical application of CGI and has application potential in target detection scenarios such as unmanned driving and lidar.