Research On Perceptual Quality Metrics Of 3D Meshes

With the rapid development of 3D modeling technologies,3D model and its applications have received extensive attention.As the most commonly used representation of 3D model,3D mesh has been widely used in various applications,such as computer aided design,medical imaging and digital entertainment.After a 3D mesh goes through mesh processing operations,the 3D mesh is always subject to geometric distortions,which may degrade the visual quality of the 3D mesh to some degree.It is necessary to assess the visual quality of 3D meshes.As the ultimate receptor of 3D meshes is typically human eyes,a straightforward way to assess the visual quality of 3D meshes is subjective scoring.However,subjective scoring is usually time-consuming and has a certain randomness,which makes it unsuitable for most practical applications.Thus,it is urgent to develop mesh quality metrics that model the perceptual characteristics of human visual system,which could improve the efficiency of mesh quality assessment dramatically.Therefore,it is of great significance to investigate the perceptual quality metrics of 3D meshes.The thesis investigates the perceptual quality metrics of 3D meshes.The research contents of the thesis focus on the perceptual importance of local regions on the mesh for mesh quality.We analyze the statistical distribution of local distortions of the vertices and investigate the relationship between mesh quality and the statistical descriptors of local distortions.A visual importance model is built based on the local distortions of the vertices.We investigate the relationship between mesh saliency and mesh quality,and build another visual importance model based on mesh saliency.The main research contents of the thesis include: mesh quality prediction based on the statistical descriptors,the investigation of a spatial pooling method using the percentile weighting strategy and mesh quality assessment by incorporating mesh saliency.Based on the research contents,the thesis proposes three full-reference mesh quality metrics:TPDMSP,TPDMPW and TPDMVS.The thesis investigates the relationship between the overall quality of the distorted mesh and the statistical distribution of local distortions of the vertices.The statistical descriptors are extracted from the local distortion map as a feature vector to represent the overall quality of the distorted mesh.We use the support vector regression model to learn the function that maps from the feature vector to the quality score.The thesis proposes a mesh quality metric TPDMSP based on the statistical descriptors.Three publicly available subjective databases are used to demonstrate the prediction accuracy,the cross-model generalization capability and the cross-database generalization capability of the metric TPDMSP.We analyze the contribution of each element in the statistical descriptors to the feature vector.Experimental results demonstrate that there is a close correlation between the overall quality of the distorted mesh and the statistical distribution of local distortions,and the regions with more severe local distortions on the mesh have a greater impact on the overall quality of the mesh.The metric TPDMSP performs better than state-of-the-art mesh quality metrics.The thesis conducts an in-depth investigation of the contribution of the regions with severe local distortions on the mesh to the overall quality of the mesh.We build a visual importance model based on local distortions by assigning a greater weight to a small portion of vertices that suffer from severe distortions.The local surface area of the vertex is used to weight the local distortion of the vertex in order to reflect the influence of the local surface area on the overall quality of the mesh.The thesis proposes a spatial pooling method using the percentile weighting strategy and proposes a mesh quality metric TPDMPW based on the percentile weighting method.We investigate the influence of the parameters of the percentile weighting method on the performance of the metric TPDMPW,and determine the unified values and optimal values for the parameters through empirical tests on three publicly available subjective databases.Experimental results indicate that the percentile weighting method has a strong capability to emphasize the impact of the regions with severe local distortions on the mesh quality,and the local distortion-based visual importance model contributes to improving the performance of mesh quality metric.The metric TPDMPW shows superior performance over state-of-the-art mesh quality metrics.The thesis investigates the relationship between mesh saliency and mesh quality assessment.We optimize the design of the mesh quality metric at the stage of spatial pooling by using the saliency value of the vertex to weight the local distortion of the vertex.A mesh saliency-based visual importance model is built to emphasize the impact of salient regions of the surface on the overall quality of the mesh.The thesis proposes a mesh quality metric TPDMVS by incorporating mesh saliency into mesh quality assessment.Three well-known mesh saliency detection methods are used to generate the saliency maps for analyzing the performance of the metric TPDMVS.A saliency map synthesis method is developed to synthesize the saliency maps by assembling the salient regions from different individual saliency maps and finally generate a synthetic saliency map for the mesh.We analyze the performance gain of the synthetic saliency map over the individual saliency maps by using each saliency map in the metric TPDMVS,and investigate the dependence of the performance gain on the similarity between the individual saliency maps.Experimental results demonstrate that the mesh saliencybased visual importance model can improve the performance of the mesh quality metric,and the mesh saliency-based visual importance model achieves better performance than the local distortion-based visual importance model for mesh quality assessment.The metric TPDMVS shows better performance than state-of-the-art mesh quality metrics.