Image inpainting algorithm based on rapid latent spatial diffusion model

• 1、Shanghai Aerospace Electronic Technology Institute, Shanghai 201100, China

Abstract：To address the decline in indoor visual localization and navigation performance caused by pedestrian occlusions, this paper proposes an image inpainting algorithm based on a fast latent space diffusion model. This algorithm repairs and completes the pixel gaps caused by pedestrian occlusions, aiming to restore the original structural information of semantic targets as much as possible. The image inpainting algorithm based on the rapid latent space diffusion model transforms input data from pixel space to latent space and simplifies the forward diffusion process of the diffusion model, significantly reducing the time overhead of the diffusion model while ensuring the quality of inpainting. Extensive experimental results demonstrate that the proposed image inpainting method effectively restores the semantic structure of foreground targets obscured by pedestrians, thereby improving the accuracy of scene recognition and ultimately reducing the localization errors caused by pedestrian occlusions.

Keywords： visual localization image inpainting diffusion model pedestrian occlusion