| Summary: | 碩士 === 中華大學 === 資訊工程學系碩士班 === 102 === Following the development of Autostereoscopic 3D Displays, using one color image and its depth map to generate a multi-view stereoscopic image has always been an important research direction. In the process to produce a stereoscopic image, if the quality of the depth map is poor, the resulting multi-view stereoscopic image will be incomplete. Therefore, the depth map greatly affects the quality of the stereoscopic image. Generating a depth map can be divided into two parts; estimating the initial depth map, and refining the estimated initial depth maps. This thesis research proposes a specially shaped mask to refine the estimated initial depth map, thus enhancing the quality of the depth map. In the methods for refining the depth map, the focus is usually on entire depth map, or the objects are segmented and then their depth map values are redefined. A giant mask is generally used in the algorithm part; or belief propagation iterations and some traditional filters are typically used. If a depth map of better quality is desired, the complexity of the algorithms used increases, thus increasing the difficulty to achieve real time. The goal of this thesis is to use simple algorithms to quickly determine which areas need refinement in order to improve the quality.
The proposed algorithm of this thesis can be divided into three parts. In the first part, we use three images, the left view color image, the right view color image, and initial depth map, then differentiate between the two color images. An image formed by the strong color changes becomes the feature image. By inspecting the feature image, a symbol is used to define the areas that include errors or are occluded if the color difference is over a threshold. After the symbol is defined, the cross-shaped mask and traditional histogram-based voting method is used to refine it. In the second part, the focus is applied on those scrambled areas. Different from previous method which segment the objects in detail, the whole picture is divided evenly, and the number of pixel in each area is accumulated. If the pixel number is lower than the threshold, the area is marked to be rectified. Then the refine method from part one is used. Finally, the canny edge and N x N mask are used with the histogram-based voting method to refine all the images.
From the experimental results, it is concluded that the proposed method compared with previous methods only processes the areas that are marked for refinement. This method could not only reduce the number of useless corrections, but could also reduce number of processes in the operation. Simultaneously, the algorithm has a two percent growth in its accuracy rating.
|
|---|
