Edge-Guided Human Depth Image Inpainting Based on Diffusion Models

In the context of depth images, fusing human segmentation images with depth background images provides critical data for human detection tasks, where accurate localization of the human head is a key indicator. However, in real-world scenarios with occlusions and complex backgrounds, human segmentation images often lack head pixels. This missing data leads to subpar performance in human recognition tasks for deep learning models. To address this issue, this paper proposes an efficient approach that combines edge detection, projection analysis, and template matching for upper-body localization and head identification. Additionally, a refined diffusion model is used to repair the human head edge mask obtained from the initial processing. The main advantage of this approach is its independence from complex deep learning model training. It integrates traditional image processing techniques, morphological analysis, and diffusion models, simplifying head localization and effectively avoiding interference from the shoulder and arm regions. This approach enables more accurate head localization and uses diffusion models to restore missing portions of the human head, thereby improving overall pose recognition accuracy. Experimental results demonstrate that the repaired depth images outperform the original images in human pose model training. When trained with the repaired images, detection accuracy (mAPVal50-95) and precision improved by 10.1 % and 7.1%, respectively, significantly enhancing downstream object detection performance. These findings validate the effectiveness of the proposed method in repairing missing human head depth images. They further highlight its potential in improving pose recognition and object detection performance, offering reliable technical support for related tasks.