| In recent years,with the landing of 5G technology and the maturity of ultra-high-definition video technology,panoramic video,a new high-tech video industry,has begun to flourish.Panoramic video enriches users’viewing content and meets different visual needs,but the mapping,encoding and transmission process will cause different degrees of distortion,affecting the video quality and user’s viewing experience.Therefore,the study of effective panoramic video quality evaluation methods will help improve users’ experience and promote the development of panoramic video industry.Panoramic video quality evaluation can be divided into subjective and objective methods,of which objective quality evaluation is an important research direction in the field of video quality evaluation,because of its low cost,convenient application and other advantages,and its application scenarios are more extensive.This thesis makes in-depth research on the full-reference quality evaluation method and the no-reference quality evaluation method in the objective evaluation of panoramic video,and proposes improved objective quality evaluation algorithms for panoramic video.The main contributions of this thesis are as follows:1.For the full-reference video quality evaluation of panoramic video,this thesis proposes a full-reference evaluation algorithm based on three dimensional convolutional neural networks and attention mechanism.This algorithm is based on the spatiotemporal features and salient content features of panoramic video,and applies 3D-CNN to automatically extract and analyze spatiotemporal features,which improves the problem of inadequate time domain analysis of 2D-CNN,and introduces a mixed attention mechanism in three-dimensional space for salient content analysis and feature weighting.The full-reference quality evaluation algorithm proposed analyzes the significant content information and time domain characteristics of panoramic video based on H.265 encoding distortion.The model is verified by experiments and the evaluation accuracy reaches 0.865,which is higher than other commonly used classical algorithms.2.For the no-reference video quality evaluation of panoramic video,this thesis proposes a no-reference evaluation algorithm based on the three features of panoramic video and the support vector regression model.The algorithm extracts texture features,time domain features and natural scene statistical features according to the structure and content features of panoramic video.Texture feature extraction uses the method of combining Local Binary Pattern and Histogram of Oriented Gradients,and then add panoramic weighting;Three-dimensional Discrete Cosine Transform method is used for time-domain feature extraction;The statistical feature extraction of natural scenes uses the mean Subtracted Contrast Normalized statistical feature.Then,the three feature sets are dimensionally reduced,and finally input into the support vector regression model together with the video subjective score for training to obtain the evaluation model.The noreference quality evaluation algorithm proposed in this thesis aims at the panoramic video with H.265 coding distortion.Based on the three features of panoramic video and SVR model,the no-reference quality evaluation of panoramic video is achieved.The evaluation accuracy PLCC has reached 0.829 after experimental verification,which is more effective than other commonly used evaluation algorithms. |
