Research On Perceptual Quality Assessment Model For Stereoscopic Images Via Machine Learning

With the development of stereoscopic video display technology,stereo products are widely used in all walks of life,such as stereoscopic video conferencing system,3-Dimension(3D)movies,3D advertising,Virtual Reality(VR)experience,etc..3D-content making as well as stereoscopic image and video processing,have already brought great economic and social benefits to human beings,and have great potential for further using.Perceptual quality of stereoscopic images presented by stereoscopic display system,is definitely important to stereo technology application.Thus,perceptual quality accessment has been one of the key technologies for stereo technology development,which mainly includes two aspects: one is similar to the 2D image quality asscessment –finding the relationship between image distortion and quality;the other is about the visual comfort accessment,as the stereoscopic display technology can cause visual fatigue,dizziness and other uncomfortable phenomenon.Designing objective accessment model based on machine learning,is the main useful research method for perceptual quality prediction.On considering the characteristics of binocular visual perception,this paper has explored new perceptual quality analyzing agorithoms.The major contributions are as follows:(1)we introduce a new visual comfort assessment model based on multiple kernel boosting(MKL)method,by which a a strong classifier is learned to measure the preference probability of stereoscopic images and a mapping strategy between the probability and the visual comfort is analyzed.The traditional comfort predition algorithms,usually train a model via one kind of regression algorithms and face the problem to choose the most suitable one.From this view,the proposed model use MKL method and turns out to be very efficient.Experiments on publicly available datasets demonstrate that the proposed method outperforms state-of-the-art regression modles for predicting visual comfort.(2)We design a subjective perception experiment to study the masking effect working on asymmetric stereoscopic images,and construct a new category-deviation database for asymmetric stereoscopic quality assessment research.Due to the masking effect,when the difference between the left and right images is not larger than a masking threshold,the perceived quality will be not changed.From this point,stereoscopic images can be classified into two distinct categories.In the experiment,by fixing the distortion level of left image and increasing the the distortion level of right image for testing stereoscopic image pair,observers are asked if they perceive the overall quality variation,and thus,we construct a training database including the noticeable and un-noticeable stereoscopic image categories,and assign a label for each stereopair.We consider three different distortions for quality degradation in the experiment,and the pristine stereoscopic images are perceptually comfortable so that the factors affecting visual discomfort are excluded.(3)We propose a no-reference quality prediction method for asymmetrically distorted stereoscopic images,which firstly introduces the label consistent K-SVD(LC-KSVD)classification framework to supervise dictionary learning in the field of stereoscopic quality assessment.We use the newly constructed category-deviation database for dictionary training,and learn view-specific feature and quality dictionaries to establish a semantic framework between the source feature domain and target quality domain.In the testing stage,the qualities of the left and right images are predicted by sparse reconstruction,and the overall quality of the stereo image is obtained by quality pooling.Experimental results demonstrate the effectiveness of our blind metric-showing better performance for asymmetric stereoscopic quality assessment and also comparable results for symmetric stereoscopic images.