Robustsparse Representation Algorithm For Depth Map Restoration

Depth map plays an important role in computer vision, since it provides a kernel cue(depth data) to describe and understand objects and scenes in real world, which isimpossible for ordinary natural images. Recently, as acquisition of depth is becomingeasy and population, applications of depth map have been growing to be a hotresearch topic. However, no matter depth map is captured directly by3D scanningtechniques or estimated by standard computer vision algorithm, it cannot be directlyused in real applications due to the gross noise obtained in depth. Therefore, there isan urgent need to construct an effective algorithm to restore noisy depth map.In recent years, the popular sparsity-based denoising framework has successfullyapplied for image denoising and restoration, which builds sparse representationframework on a self-adaptive dictionary by combining the merits of image sparserepresentation and dictionary learning techniques. Compared to an ordinary sparserepresentation model which uses a fixed dictionary (such as composed of DCT orwavelet bases.) the framework can more efficiently represent the given image byusing a dictionary training process which can adaptively capture useful detailstructures obtained in the image. Inspired by the successful application in imagedenoising, more recently, the dictionary learning sparse representation framework hasalso been used by researchers for depth map restoration. Although the framework canalso adaptively obtain detail structures in depth map, it’s not robust for depth maprestoration. Since the framework is built on the stationary white Gaussian noiseassumption, while this often holds for natural image but not for depth map. Depthmap is not only slightly contaminated by stationary white Gaussian noise but alsoseriously corrupted by non-Gaussian noise (so called replacement noise) such asoutliers, occlusions, and variable uncertainties in boundary. Experiment resultsindicate that the replacement noise pixels in depth map not only can misleaddictionary to obtain irregular structures but also seriously deteriorate the effectivenessof sparse coding, resulting in a poor performance for depth map restoration.To overcome these drawbacks in depth map restoration, in this paper we propose anovel robust dictionary learning sparse representation framework ILSR (IterativeLearning Sparse Representation). The ILSR framework can automatically detect thebad pixels (which are seriously corrupted by replacement noise) in depth map andgradually reduce their disturbance effects to the dictionary learning and sparse codingby constructing a mixture depth map. In the framework, the bad pixels detection and the dictionary learning naturally fuse into a mutually iterative sparse representationprocess, and a self-adaptive threshold value is used for sparse coding in the process.Through several times iteration, the framework can effectively remove noise fromdepth map while keep detail structures, ultimately a high quality depth map isrecovered. To further enhance ILSR algorithm, improve its robust to non-Gaussiannoise and speed its restoration paces, we combine the merits of the framework and themedian filtering technique to produce a two steps depth map restoration algorithmS-ILSR (Speed Iterative Learning Sparse Representation). The first step is called fastrestoration step, in which a median filtering is used to quickly remove gross noise indepth resulting in a coarse restoration for depth map by constructing a mixture map.And the second is called deep restoration step, in which an iterative dictionarylearning spare representation process is employed to further remove remaining tinynoise resulting in a subtle restoration for depth map.To demonstrate effectiveness of the proposed framework a series of comparisonexperiments are conducted on the public USF and Middlebury depth map databases.And the results indicated that our proposed robust dictionary learning sparserepresentation framework can efficiently remove noise from depth and recover a highquality depth map.