Underwater Image Depth Extraction Based On Markov Random Field Algorithm

Computer vision is mainly used to study the application of computer to simulatehuman and biological technology subject function of the visual system. Its researchgoal is: through a computer, we can have a picture including3D scenes out of a2Dimage cognition, which uses the computer realization of the objective world of3Dscene perception, understanding and recognition. Computer vision is not only anengineering field, but also a challenging research area in the field of science.Many computer vision algorithms have ignored the underlying3D geometrystructure of the image. The algorithm in this paper is introducing that through thegeometric model based on appearance matching class learning, even with a denseclutter natural scene, can be robustly estimated in a scene about their geometricproperties. Relative to the camera, the geometry classes describe the threedimensional positioning in the image area of. This paper provides a framework thathas a variety of assumptions, the framework can be implemented based on a picture ofthe robustness of the scene structure to estimate and label each of the geometries toincrease confidence level. These worthy confident labels can also be applied to manyother applications of the enhancement effect.This paper briefly introduces the depth extraction technology development at thepresent situation. Also this paper introduces the algorithm principle and technicalindex, and based on this, the "underwater3D image reconstruction algorithm based onmachine learning" is a key content which were described in detail here.This article focuses on the use of Markov Random Field (MRF) model to modelthe monocular visual clues and finding the relationship between the different parts inthe same image. Previous depth information extraction methods of underwater imagecharacteristic is condition fully, to get more accurate depth information, thedisadvantage is that it requires specialized instruments and the complex parameteradjustment. This paper discusses the underwater depth of Markov Random Field information extraction method to overcome the above problem, and has been appliedto a variety of problems. These messages are often found in local characteristics thatare low and need to be a lot of related information from the image. It has a very goodfeasibility and much lower cost. And it is a heavy load of machine learning methods,and can improve the accuracy of underwater depth information. In addition, theapplication examples are stereoscopic observation, image segmentation and objectclassification, etc.This paper first introduces the basic theory of pattern recognition. Then, theauthor introduces the characteristics of underwater image feature extraction anddifficulties, and then introduces the principle of image information extraction basedon Markov Random Field about underwater method.Considering the task that from a monocular image we can estimate the depth, thisarticle adopts the method of supervised learning to study the problem. In this paper,our experiment process is, first of all, collect a set of trained underwater monocularimages, and also the corresponding real depth information of each other. Then, wewill monitor the study application which is to map the predict depth as a function topredict the image depth. Since from only one point on the internal characteristics ofimage to estimate depth is inefficient, it also make the monocular images depthestimation has become a challenging problem. We need to consider not only theimage of the depth of the single point, we also need to consider the fluctuation of theglobal information in the whole image. Our model which is using a different trainedMarkov Random Field (MRF). It contains not only the multi-scale andmultidimensional local and global properties, but also the depths’ relationshipbetween different points. This paper shows that even in unstructured scene, thealgorithm is to restore accurate depth map in a higher accuracy.The image can be divided into a few small patches. For each of them, we use aMarkov Random Field to infer a set of "plane parameters". And these parameters canshow the space position and the direction of the patch. The Markov Random Field,through the supervised learning, the depth cues about this image and the relationshipsbetween different regions of the image can be modeled. In addition to assuming that it is composed of many pieces of small planes in the environment on the photo, themodel did not clearly assume that the scenario of concrete structure. Changes aboutthis are made in this paper. Comparing with the previous technology, the algorithmcan obtain more detailed3D structure. And, by some image rendering, the algorithmcan output a much richer wandering results. And even in the absence of obviousvertical structure of pictures, this algorithm is still valid in the scene.This algorithm has been successful in70of120images to output qualitatively3Dmodels. Underwater image depth extraction method based on Markov Random Fieldhas the characteristics of low cost and is easy to use. It has a very good applicationprospect.