Depth Information Processing And Hole Filling In 2D-to-3D Conversion

Since the hot showing of "Avatar" in 2009,people want to see all films in 3D,such as "Transformers ".What is more,people even want to see cartoons in 3D,too.3D movies can be shot directly nowadays,but it costs a lot in time and money,as a result,consumers' desires for seeing 3D movies can't be satisfied.With the technology of computer graphics and image processing,we can do something on 2D movies to generate 3D movies,which largely making up for the vacancy of 3D movies.In this paper,we did a lot of work on the whole procedure of 2D-to-3D conversion,mainly studied about the generation of depth image and virtual viewpoint image,to achieve high-precision conversion effects.In depth map generation section,a B-Spline based image segmentation technique is presented to generate accurate depth map.Using B-Spline one can draw the object contour easily,based on which foreground and background seed points can be generated automatically.The seed points are used in Graph Cut algorithm,which will lead to a segmentation line.But the segmentation line got this way can be Inaccurate locally and with some sawtooth.In this paper,the result of segmentation is used to rectify the fitting of B-Spline for smooth and accurate segmentation line.At last,some detail adjustment is done locally to achieve better accuracy.In the section of generating a virtual viewpoint,a hole filling algorithm based on image segmentation is presented,which making using of the relevance of contiguous video frames.The edge of target object in video frames can be blurring frequently,which causing a great impact on hole filling.In this paper,image segmentation is carried on to get rid of ambiguous pixels on the first step.Secondly,matching is performed on neighboring video frames,which takes many factors,such as,color,texture and structure into account for better degree of matching.When got the most matched patch,we simply copy and fill the hole instead of computing the weighted mean value,avoiding of introducing blur.After hole filling,we may come across several problems affecting the 3D image quality.There will be some holes not filled still,and to solve this problem,we can make use of bilateral filtering algorithm.Improved Gaussian blur algorithm can be used to solve the wrong pixel problem.Experimental results show that with the combination of the two parts proposed in paper,we can obtain highly accurate virtual viewpoint image,it is also true for scenes with blurred edges.