Research Of Depth Estimation Upon Monocular Images Based On MRF Theory

Stereo Computer Vision is an important branch of Computer Visual Science. Depth of objects in the images, i.e. the actual distance between the object and the focal plane, is really apperceived with Imaging Technology. Depth information of a picture has been widely used in fields of image understanding, stereo reconstruction and computer vision, etc. At present, there are two kinds of depth perception methods: using radial sensors directly and using photography geometry based on Parallax. The advantages of these two methods are: they are of sufficient conditions and accurate depth data; however, the disadvantages are also clear: they need inner parameters of video camera and lots of physical working is necessary. The model in this paper base on MRF (Markov Random Field) gets over such disadvantages with low cost and more accuracy.As the framework of the whole article is based on theory of Pattern Recognition, we introduced it at first. The main purpose and effect of pattern recognition is putting an concrete object into a class which it should belong to. We view the process of Depth Estimation as an issue of Pattern Recognition. And the depth of monocular images is a pattern to be classified. At the same time, we combined absolute features and relative features of the monocular images in multi-scale.In order to match the depths and their features, we make use of the existed image library, and introduce DRF (Discriminative Random Field) Methodology, which is on the basis of MRF, to model the relationships between depths and features, and the relations among depths at different patches in the image. MRF methodology which is used to solve the ill-posed problem is exactly depicting the context relationships among a set of entities. DRF is a special express of Conditional Makov Random Fields, and it is composed of association potential and the interaction potential which could tell the relationship between depth and its features more clearly. Parameters of this model must be estimated by supervised learning using existed image library. According to Bayesian theory, issues of Depth Estimation could be transformed into problems of MAP (Maximum a posteriori). Thus, all algorithms are attributed to the model of DRF-MAP.The work which we have done is introducing DRF-MAP into field of Depth Estimation, and we have four kinds of algorithms with different parameters of the two parts in the model. We could tell the kinds of pictures each algorithm suited for. The last one named mixed model is firstly proposed by us, which did a good job according to experimental results. With the mixed model, we can get a more accurate depth-map in detail when we consider images with bigger complex texture patches.