Research On Point Structured Information Representation Learning-based Point Cloud Quality Assessment

The proliferation of 3D information capture technology has seen widespread use of 3D data forms such as depth maps,point clouds,voxels,and octrees in areas such as virtual reality,immersive presence,mobile mapping,and 3D information printing.Among these,the point cloud,characterized by its precision,density,and multidimensionality,is notably utilized.Each point in a point cloud represents a position in three-dimensional space,possessing unique attributes like color and normal vector.Realistic visual data representation necessitates point cloud models containing millions,if not hundreds of millions,of points.To enhance transmission rate and reduce storage resource requirements,lossy compression schemes are typically employed.However,these methods may lead to compressed sensing distortion.External interference during data acquisition and transmission may also cause downsampling perception distortion and Gaussian filtering distortion,further deteriorating the perceptual quality.To better manage and control the subjective quality of point clouds,it is imperative to research and propose high-performance color point cloud quality evaluation standards that align with human perception.Existing quality evaluation methods for point clouds simulate the human visual system,employing a variety of point cloud features and integrating machine learning technology.In image quality evaluation,the efficacy of structural information has been proven.However,the unstructured nature of point clouds poses a challenge for traditional methods to extract image-like structural information.This paper focuses on extracting structured information from point clouds and integrating key technologies such as deep learning into the quality evaluation of reference-less point clouds.The main research outcomes are as follows:Firstly,a color gradient-based,reference-less point cloud quality assessment algorithm is proposed.The farthest point sampling method obtains point cloud sampling points,while the K nearest neighbor method extracts the K nearest neighbors of each sampling point,creating a point cloud block.We calculate the normal vector of the point cloud block,perform spatial projection in its direction,and compute color structural information features.Moreover,traditional point cloud quality evaluation methods extract features like curvature,distance,and brightness of point cloud blocks.A dual-stream convolutional neural network is designed to learn the correlation between point cloud features and objective quality scores.Secondly,a reference-less point cloud quality assessment algorithm based on the point structured information network is proposed.The algorithm,inspired by the concept of point weight distribution in point cloud classification,designs a weighted module for extracting local point structural information from the point cloud.An end-to-end learning network is also designed,which reduces the complexity of the process of the reference-less point cloud quality evaluation algorithm.Thirdly,an unsupervised point cloud quality assessment algorithm based on a self-attention point structured information network is introduced.A self-attention module,inspired by the self-attention mechanism in images,is designed to optimize the point structured information network.Then,utilizing the dual-stream Transformer network,the point-structured information is regressed into quality perceptual scores.Finally,a self-adjusting learning method based on label confidence is proposed.This method adjusts the quality scores of current mini-batch samples,focusing on the lower quality sections of the point cloud and mitigating the impact of noisy labels.