Research On Depth Image Enhancement Based On Soft Clustering

Depth information is an important means for people to perceive things in 3D space and is widely used in the field of computer vision.Most of the existing depth image acquisition devices are not very complete,leading to problems such as missing data and low resolution.Depth image enhancement technology utilizes depth information to solve these problems,which belongs to the category of edge-aware smoothing.The processing of depth images can be roughly divided into filter methods and solver methods based on whether confidence is required.Traditional filters may cause halo artifacts,fail to obtain depth information with high accuracy,or reduce computational efficiency when solving complex computer vision tasks.Some existing solver methods may be limited by their algorithm optimization,resulting in significant differences between the output and the actual value.Therefore,the effect of depth image enhancement needs to be improved.To conduct in-depth research on the problems of the existing depth image enhancement technologies,this thesis proposes a study of the depth image enhancement method based on soft clustering.The specific work content is as follows:(1)To suppress halo artifacts and improve certain solver methods that may be limited by their algorithm optimization,this thesis proposes a depth image enhancement method based on low-level features soft clustering,called the soft clustering solver.The objective function of this method is constructed as a confidence-based weighted least square model.The key point of the model is that the optimization can be formulated by using the properties of soft clustering.Apply soft clustering to define the similarity between pixels and improve the accuracy by iterative execution.This thesis conducts experiments on several depth enhancement tasks to evaluate the proposed method,including depth inpainting,depth super-resolution,and depth rectification.Quantitative and qualitative experimental results show that the proposed method outperforms the state-of-the-art methods in accuracy.(2)To improve the proposed depth image enhancement method,which occasionally produces texture duplication.Considering the filtering processing of the color-guided images,this thesis proposes a depth image enhancement method based on the fusion of high-level features soft clustering,which combines high-level semantics with low-level features.The principle of the proposed edge-aware method is to first perform soft clustering on the input image according to the high-and low-level features to derive the affinity matrices,which are then fused for image smoothing.This method has higher computational efficiency and improved accuracy compared to the existing methods.