Depth Image Inpainting Algorithm Based on Improved Non-Local Means Filtering
Jian Chen, Jianfeng Zhang · 2024
Depth images captured by depth cameras often suffer from issues such as holes and noise, which can have adverse effects on subsequent practical applications, including the accuracy of 3D reconstruction. To address these issues and enhance the quality of depth images, we propose depth image restoration algorithm based on improved non-local means (NLM) filtering. Initially, the algorithm utilizes fuzzy C-means (FCM) clustering to cluster the color image and generate a guidance image. Subsequently, neighborhoods are defined centered around the hole points in the depth image. A weight function for NLM filtering is then constructed using bilateral filtering. By calculating the weighted average of all non-hole points within the neighborhood, the holes are repaired. Finally, median filtering is utilized to smoothen and reduce noise in the depth image. Experimental findings show that the proposed algorithm efficiently repairs holes while preserving the edge details of image.