Adaptive disparity map computation for stereoscopic video watermarking

Abstract

By reconsidering some 2D video inherited approaches and by adapting them to the stereoscopic video content and to the human visual system peculiarities, a new disparity map is designed. First, the inner relation between the left and the right views is modeled by some weights discriminating between the horizontal and the vertical disparities. Secondly, the block matching operation is achieved by considering a visual objective measure (NCC) instead of the traditional pixel differences (MSE or SAD). The processed corpus regroups 20 min of stereoscopic video sequences captured within the 3DLive French project (30000 stereo pairs). The new disparity map was benchmarked against two state-of-the-art algorithms, namely the NTSS and the FS-MPEG. The results exhibit quality gains in the reconstructed image between 5.67% and 5.37% in PSNR and between 7.58% and 4.63% in SSIM. The computational cost was reduced by factors between 1.3 and 13. This disparity map was finally integrated into a 3D-TV watermarking method.Keywords-3DTV disparity map, visual quality, watermarking I. PROBLEM STATEMENT In its widest acceptation, a disparity map provides information about the coordinates at which similar blocks are located in two images (the reference and target images). Hence, computing a disparity map requires to design a rule specifying how the reference blocks are searched for in a given area of the target image and to define a similarity metric establishing whether a reference block matches a target block. By exploiting the spatio/temporal correlation between successive frames in 2D video, several disparity maps have been advanced and proved their efficiency in various fields, like compression, indexing or segmentation. They generally assumed the differences between the target and reference frames being homogeneous in space.The Exhaustive Search (ES) [1] is widely used for block motion estimation in video coding, in order to determine effective similarity while providing minimal error estimation. However, for running the full search window, the algorithm requires a massive computation. Several fast algorithms were developed to reduce the computation time. Zeng and Liou [2] advanced the New Three Step Search (NTSS) algorithm, (Fig. 1.a) and showed that it provides smaller motion compensation errors comparing to the state of the art given by the Three Step Search [3], the 2D logarithmic search [4] and the conjugate directional search [5]. Zhu and Kuang [6] designed a Diamond Search algorithm achieving close performance to the NTSS algorithm in terms of reconstructed image quality and alleviating the computational constraints by a factor of 1.5. In all these studies, the block matching is based on measures of the differences between the two block pixels (Mean Square Error -MSE, Sum of Absolute Differences -SAD). (a) (b) Figure 1. NTSS (a) vs. 3DV-NTSS (b) searching areas.These principles inherited from the 2D video are commonly in use for stereoscopic video processing [7]- [9]. For instance, the study in Wang et al. [8] uses, a non-s...