| Remote sensing is an important earth observation technology.The image data obtained contains rich feature information,which can present many natural and human scenes on the earth.Therefore,we can identify,judge,and extract remote sensing image data.Obtain the useful information it contains.With the continuous improvement of remote sensing technology,the image data we obtain has a higher spatial resolution,and the amount of data is also increasing rapidly.How to quickly and effectively obtain useful information from the huge amount of data Information is still the focus and difficulty of current research.Segmentation is one of the most important and challenging tasks in the field of remote sensing image processing.The purpose of image segmentation is to divide the image into several uniform areas with some same properties in a certain way.In the past few decades,Have proposed many methods to solve this problem,such as threshold method,active contour method,Markov random field model(MRF),etc.Among them,MRF model can capture the spatial context relationship between the basic units of the image,so it is widely used However,the current Markov random field method also has its needs for improvement.For example,the pixel-level method has too small extraction of image features due to the small neighborhood,which makes the segmentation results too prominent.Detailed information is often misclassified;although the object-level method has been improved on the above problems,using the over-segmented area as the primitive,but due to the inaccuracy of the initial segmentation and the influence of the irregular spatial context relationship,it Object-level segmentation accuracy is also unsatisfactory.In view of the shortcomings in the above MRF model,this article The introduction of the Alternative Direction Method of Multipliers(ADMM),the main research work is as follows:(1)When solving the classic object-level MRF model,the feature field energy and the label field energy are considered together,and the effect of the label field energy on the image features is adjusted by the potential function ?,but the disadvantage of this is that it is in the two energy The important information of the characteristic field or the labeling field may be blurred when the summation is performed.The first work of this paper is to introduce the discrete ADMM algorithm in the MRF solution process,and separate the feature field energy and the labeling field energy by the discrete ADMM algorithm.In the framework of the ADMM algorithm,the results of the feature field and the label field affect each other,and the continuous solution is continuously obtained to obtain the final result,which verifies the feasibility and effectiveness of the discrete ADMM algorithm in the MRF model.(2)After obtaining the feasibility and effectiveness of the discrete ADMM algorithm in the MRF random field model,consider using the discrete ADMM algorithm in the MRF random field model with multigranularity features.The segmentation results obtained by the pixel-level method have good marginality.The result obtained by the object-level method has good regionality.In order to obtain both good marginality and good regional segmentation results,the second work of this paper considers the discrete ADMM algorithm for both pixel-level and object-level.The granularity solution process combines the pixel-level results and the object-level results through the ADMM algorithm to influence each other.The results of the two granularity methods are continuously iterated and iterated to obtain the final segmentation results.In order to verify the effectiveness of the two methods proposed in this paper,we will apply the above two methods to the solution of remote sensing image segmentation,and perform experiments on different data sets,and compare the obtained segmentation results with existing ones.Based on the comparison of MRF methods,the quantitative analysis index shows some advantages of this method in remote sensing image segmentation. |
PDF Full Text Request