Research On Stereoscopic 3D Image Quality Assessment Based On Deep Learning And Visual Perception

The development of contemporary science and technology has driven the rapid development of stereoscopic imaging technology and multimedia applications.Compared with plane images,3D content brings people an immersive visual experience and greatly enriches people’s daily life.However,during the process of generation,transmission and display,stereoscopic 3D(S3D)images can be contaminated by various distortions,resulting in the decline of S3 D image quality,and affecting people’s visual experience.Therefore,it is of great significance to evaluate the quality of S3 D images.Nowadays,the S3 D image quality assessment algorithms based on convolutional neural network have attracted more and more attention and achieved excellent results.In this thesis,human visual characteristics are simulated and used for S3 D image quality assessment.The following research results are obtained:1)A no-reference S3 D image quality assessment algorithm based on global and local content characteristics is proposed.This algorithm mainly includes global feature fusion sub-network and local feature enhancement sub-network.The corresponding features are extracted from the fused view and single view respectively,and then the perception of human visual system to image quality is simulated.For the fusion view,the cross-fusion strategy is applied to model the process in V1 visual cortex,and multiscale pooling is utilized to integrate the context information under different sub-regions,so as to effectively extract global features.For the single view,the asymmetric convolution block is introduced to enhance the local information description.By comprehensively considering the fusion view and single view,the network can effectively extract features.Finally,the weighted average strategy is adopted to estimate the visual quality of S3 D images.Experimental results on LIVE 3D and Waterloo IVC3 D databases show that the proposed method is superior to the state-of-the-art metrics,and achieves an excellent performance.2)A no-reference S3 D image quality assessment method based on multidimensional attention mechanism is proposed.The method is divided into view fusion sub-network and multi-scale feature enhancement sub-network based on multidimensional attention.Firstly,the multi-dimensional attention is applied to calculate the weight of left and right views to obtain the fused view,which can simulate the binocular fusion and binocular rivalry mechanism of human visual system.In the multiscale feature enhancement sub-network,multi-scale features are extracted from the fusion view,and different scales features are enhanced by multi-dimensional attention,so as to assign weights to different scales information.Finally,the quality score is estimated through full connection layer.Comparative experiments are carried out on the LIVE 3D and Waterloo IVC 3D databases.The results show that the algorithm can accurately predict the quality score of S3 D images and has high consistency with subjective assessment.