| With the increasing popularity of stereo imaging technology and the continuous development of the 3D high-definition imaging industry,people have higher and higher requirements for image clarity and visual effects.However,in the process of acquiring,transmitting,restoring,and storing 3D images,different types and degrees of distortion will be introduced,which will cause the quality of stereo images to decrease,and affect people’s understanding and use of image information.Therefore,the research of stereo image quality assessment becomes more and more important,and stereo image quality assessment is also an effective method to evaluate the performance of stereo imaging system.Compared with2 D images,each view of 3D images not only suffers from the monocular distortion that 2D images will produce,but also has symmetrical or asymmetrical monocular distortion,including binocular confusion,depth perception errors,visual discomfort,etc.These distortions are in turn It is called binocular distortion.Therefore,stereoscopic image quality assessment not only considers the monocular distortion of the left and right views of the stereo image,but also considers its binocular stereo perception characteristics: depth,parallax,etc.This paper focuses on the application of stereo visual characteristics in the field of objective stereo image quality evaluation,and gives two reference-free stereo image quality evaluation methods considered from the aspect of stereo perception.First,a multi-feature-based stereo image quality evaluation method is proposed based on the visual characteristics of human eyes: parallax,saliency,and edge,which can characterize stereo image quality,are selected,features are extracted and formed into a feature map,and a three-channel convolutional neural network is constructed to train the feature map to realize the mapping from image features to quality scores,so as to predict the quality scores of stereo images.Secondly,in order to further improve the performance of the model,a non-reference stereo image quality assessment method combining VGG-16 and the distinctive features of stereo image depth is proposed.This method improves the saliency detection model SDSP and fuses the depth features of the stereo image through wavelet transform to obtain the depth saliency features;improves the convolutional neural network VGG-16 to increase the running speed of the algorithm;combines the depth saliency features,contrast features,and The normalized feature of the brightness coefficient is used as the input of the network,and the regression model is obtained through training to predict the quality score of the stereo image.Finally,the two algorithms proposed in this paper are verified and analyzed in the LIVE 3D IQA Phase I,LIVE 3D IQA Phase II and NBU 3D IQA databases.The experimental results show that the stereo image quality score predicted by the model conforms to human subjective perception,and has good applicability and robustness. |
