Depth Map Recovery Based On Matrix Of Low Rank And Local Image Model

In recent years, with the rapid development of visual imaging technology, the general video based on color images does not accurately depict the reality of three-dimensional scenes, so it has become increasingly unable to meet the people's needs, and the three-dimensional imaging technology arises at the historic moment. Three-dimensional images contain information of three dimensions naturally, which not only contain the traditional color images, but also distance information of the object scene, that is, the depth of information. The image that only contains depth information is named depth image, and because depth information is not affected by the radioactive nature and the light of the surfaces in the scene, which can more accurately represent the three-dimensional information of the object surface, and make up the shortcomings of conventional image to describe three-dimensional information.Despite the development of depth technology, comparing with the color image of the same scene, the quality of depth image is still poor(lower resolution, a lot of noise and missing data). Thus, the practical application of the depth camera has certain limitations; hence it is necessary to develop an effective depth image recovery technology to support their broader application in the real world. Depth camera can obtain a color image and depth maps, so depth recovery by the color image-oriented technology becomes popular. This paper focuses on depth recovery based on the matching color and depth image.1. The dependency property between color image and depth image varies partly, especially in the edges and the geometric details of the scene. For example, smoothed region in the color image is often corresponding to the flat surface in the depth image(because the reflectivity and shape correspond to each other). Furthermore, the depth image often has a strong non-local similarity, particularly in discrete areas of the scene geometry. Therefore, making the color image oriented and effectively using such non-local similarity is very important. These findings inspired us to invent a recovery algorithm which joints local and nonlocal regularization –a recovery algorithm(LRL) based on matrix of low rank and the local image model, which can adaptively utilize the dependency between color and depth maps. LRL outperforms the existing method of depth recovery on Middlebury dataset.2. LRL still has some weakness: using low rank in the image globally would increase the operating time. Therefore, with the introduction of total variation regularization, flat areas would not need low rank, which can get a good recovery on the flat area with eliminating noise effectively. For the contrast deletions and texture deletions caused by total variation, we applied different weights on the total variation, namely the weighted total variation(weighting TV). Because these weight coefficients is adaptively determined by the color image that matches depth image, which can adapt well to the image, not only greatly accelerate the speed of the algorithm, and also performs better than LRL. This new algorithm is superior to other methods On Middlebury depth data sets and real data.