Reseach On Key Technology Of 3D Reconstruction Based On Random Field

3D reconstruction takes great parts in digital media area in and out China.Researchers aim to get efficient and adaptable methods to complete it. Reconstruction methods based on image are divided into 3D reconstruction for a single still image and for multi map. 3D reconstruction for multi map needs high quality images and limits the content of images, but we can get accurate result with it. 3D reconstruction for a single still image can’t get complete result because of overlaps. 3D reconstruction on stereo-vision and multi map always ignores depth information in a single image. In theory, depth cues in a single image have no conflicts with multi map’s cues. If we can combine cues in a single image with that in stereo-vision, the result of 3D reconstruction maybe more reality.Markov random field theory can be used to solve ill-posed problems. 3D reconstruction requires recovering from a two-dimensional image. Scene reconstruction is the reverse process of imaging, so it is an ill-posed problem. It is reasonable to use the Markov random field theory to solve 3D reconstruction problems. Due to the flexibility of Markov random field model itself, we can take advantages of local depth cues hidden in the picture. So it’s easy to combine single figure cues with 3D visual cues together to meet our requirements.In this paper, with the application of Markov random field on the basis of 3D reconstruction on single figure cues, we improve the 3D reconstruction presentation as follow:First, we add new constraints to modify the Markov random field model. We are able to use the new Markov random field model for multi map 3D reconstruction and the result of 3D reconstruction is truer. Because of containing all the content in multiple images, our method made up the disadvantage of a single graph object occlusion caused by voids.Second, we also proposed two more methods for 3D reconstruction of two images.Markov random method is applied to multi-dimensional reconstruction by changing the order of reconstruction and integration.The last but not the least, we did not make any assumptions on scene when model our Markov random field and have no requirement for images. Therefore our methodsare universality. We can get visually pleasing results with rich image sample library.This paper compared our three methods with each other. After experiments, we found that 3D reconstruction based on MRF model reconstruction had a best result in vision. 3D reconstruction based on image-mosaic is the fastest method. 3D reconstruction based on point fusion strike an average. Beside, we have compared our method with the traditional method — — SFM. Experiments showed that our three methods can get wider 3D reconstruction result but the accuracy is worse than SFM.