Stereoscopic Image Coding Performance using Disparity-Compensated Block Matching Algorithm Imen Kadri *† , Gabriel Dauphin * , Anissa Mokraoui * * L2TI, Institut Galil´ ee, Universit´ e Paris 13, Villetaneuse, France Email: imen.kadri@enit.rnu.tn, gabriel.dauphin@univ-paris13.fr, anissa.mokraoui@univ-paris13.fr Zied Lachiri † † LSITI, Ecole Nationale d’Ing´ enieur de Tunis, Universit´ e de Tunis El Manar, Tunis,Tunisie Email: lachiri.z@gmail.com Abstract—This paper focuses on the disparity-compensated stereoscopic image coding. Such approach takes advantage of the existing redundancy between the two views as they are intended to render the visual impression of a 3D-scene, in which inter-view object displacements are understood as depth- related information. The classical approach is based on Block Matching (BM) algorithm, yielding a disparity map with which the predicted image is most similar to its original version. Then, with no modification of the disparity map, the residual image is encoded, yielding a refinement added to the predicted image. The proposed approach, first, improves all the possible predicted images taking into account this refinement, and then, estimates the disparity map as the one with which the predicted image resembles most that same view. Despite the significant increase in the numerical complexity, the substantial improved performance in terms of Peak-Signal to Noise-Ratio (PSNR) of this new approach is evidence of ongoing progress in this field of research. Index Terms—Stereoscopic Image Coding, Block Matching Algorithm, Disparity, Disparity-Compensation. I. I NTRODUCTION During the last decades, the use of stereo imaging technology has greatly increased. Applications concern the entertainment industry (3D cinema), video games, medical field (stereoscopic displays) and cartography (aerial stereo- photography). A stereoscopic image is composed of two views. A specific device is required to make them perceived as two viewpoints of a single 3D-scene, where same points have different positions on each view. This spatial displacement is called disparity. Estimating the true disparity map (i.e. the true depth) remains an extensive research field [1]. In this paper, our concern stems from the increased storage needs when using stereoscopic images, and the existing re- dundancy between the two views. Improving the performance of stereoscopic image coding technique is also an extensive research field. Hence disparity compensated coding is a clas- sical technique [2], [3]. Additional benefits is obtained when depth-based video coding is combined with view synthesis prediction, and this includes lifting schemes [4]. Higher perfor- mances are achieved when different techniques are combined as in H264 or in HEVC, which has been subjectively evaluated in [5]. Two extensions of H264, AVC and MVC are concerned with stereoscopic-image coding, note that they make use of disparity compensated coding [4]. This paper focuses on improving the disparity compensated coding scheme, which consists in coding separately a reference view, losslessly encoding an estimated disparity map and en- coding a residual image. The transmitted information enables the decoder to reconstruct the reference view, and using the disparity map to compute a predicted view, to which is added the decoded residual image. This scheme is very similar to motion compensation. Research within this framework has achieved increased performance when estimating the disparity map, by taking into account its own bit-cost in [6], by using blocks of arbitrary shapes in [7], and by addressing also the illumination compensation in [8]. Reducing the numerical complexity is also a significant research issue. Examples include selecting optimal hyper-parameter values thanks to allocation modeling as opposed to an exhaustive search in [9], and reducing the search area in [10]. This paper is a proof of concept showing the substantial progress to be expected when the disparity map is estimated by taking into account the encoding of the residual. To this end, an algorithm using greedy search has been designed. Section II shows how finding the best performing disparity map can be regarded as solving an optimization problem. The classical approach is derived as a suboptimal solution. Section III derives from a different suboptimal solution a new algorithm. In section IV, experimentations show significant in- creased performance on some stereoscopic images. Section V concludes the paper. II. PROBLEM STATEMENT This paper is concerned with coding rectified stereoscopic images using the disparity-compensated coding scheme based on a closed loop. The reference view encoded separately is assumed to be the left view. The right view is then decomposed into K non-overlapping blocks of same size. In traditional stereoscopic image coding, pixel values of each block are first predicted using pixel values of the corresponding block