| Three Dimension(3D)multimedia service can provide users with a more realistic visual experience,which is one of the development trends of multimedia technology.However,the quality of 3D images/videos on the market is currently uneven.People often experience a certain degree of visual discomfort when viewing 3D images/videos.Therefore,it is necessary to evaluate the comfort of 3D images / videos.It is time-consuming and labor-intensive for the human eye to judge whether the 3D images/videos is comfortable,so we need to explore a method for predicting the comfort and to improve the comfort level in a certain extent.Existing 3D visual comfort prediction methods are mostly based on spatial domain or time domain.And they cannot be predicted combining the information on compressed domain(frequency domain).The available information of the 3D images/videos has not been fully explored.Based on the above research background,this dissertation mainly studies how to use the compressed domain(frequency domain)information to establish an objective model of 3D images/videos comfort for prediction and enhance the comfort of 3D video effectively.The main contents of this paper are as follows:1)A 3D image comfort prediction model based on DCT transform(3D-Ms DCT)is proposed.Since Discrete Cosine Transform(DCT)is used in image/video compression widely,this paper performed multi-scale DCT transform on disparity maps and extracted 3D visual comfort features from multi-scale transform coefficients.The model defined three types of multi-scale visual comfort features,which are basic disparity intensity at 8 different scales,disparity gradient energy at 7 scales,and disparity texture complexity at 8 scales.The random forest regression algorithm trained and tested 23 multi-scale visual comfort features to obtain the final visual comfort of 3D images.This paper used compressed domain information to predict the visual comfort of 3D images firstly.2)Based on the three types of frequency domain features extracted from 3D image comfort prediction,a 3D image comfort prediction algorithm(3D-SF)combined spatial and frequency domain features is proposed.In order to fully exploit the visual discomfort of 3D images,this thesis extracted four traditional features of disparity amplitude based on the spatial distribution features of disparity maps.They were disparity gradient,disparity maximum and disparity range,which are classified into low-level disparity features of spatial domain.On this basis,we also used the five new parameters as the high-level disaprity features that affect visual comfort,such as image complexity,object boundary abrupt,object lateral distance and the boundary disparity of left and right.Then,combined with the three frequency domain disparity features at a single scale a total of 12 comfort features are used for 3D image comfort prediction.About features'fusion,we used extremely randomized trees regression models to train.Experimental results show that the algorithm can improve the efficiency and accuracy of prediction.3)A 3D video comfort prediction algorithm(3D-SFT)with multi-dimensional features is proposed.In the 3D scene,the motion features of the object are the key factors affecting the comfort of the human eye to view 3D video.We explored the depth change information of the object moving in the depth direction and the two-dimensional space in the 3D video to measure 3D comfort.Then,the 3D video comfort prediction is performed that combined with the DCT transformed frequency domain features and spatial comfort features of the disparity maps,which improves the accuracy of 3D video comfort prediction.In addition,a 3D video comfort improvement method based on depth adjustment is proposed.The experimental results show that the method can improve the comfort of 3D video to some extent. |
PDF Full Text Request