Algorithm Of RANSAC And Its Application In Remote Sensing Image Processing

In computer vision, engineering and many other fields, there are needs to determine the model parameters based on related data sets. However, among the data sets, there are a lot of abnormal data, which do not meet the model (also known as outliers). They have great interference on the robust of model parameters. Classical parameter estimation algorithms, such as least squares, are difficult to rule out the impact of abnormal data. RANSAC algorithm is widely used in refining the data, thus reducing the outliers'interference on the model parameters.In this paper, it will describe the development of RANSAC algorithm, its principle and its improved algorithm. It mainly discusses RANSAC application in the geometric correction and radiation correction of remote sensing image processing. In geometric correction, it is generally to fit the correction model after the control points are obtained without thinking about the impact of abnormal data. We will use RANSAC algorithm to remove the bad control points before deciding the model parameters. Before using RANSAC iteratively obtaining robust model parameters, we should firstly select the appropriate geometric model. The fundamental matrix and homographic matrix are the two most used geometric models. With respect to geometric processing of remote sensing images, this paper experiment focuses on image registration, one of the important aspects of geometric processing. There are primarily two methods to image registration:based on gray and based on feature points. In this experiment, we realize image registration based on feature points to calculate fundamental matrix, thus removing the bad matching points. Experiment result shows that image registration accuracy is greatly improved to 1 or 2 pixels error after using RANSAC algorithm to remove the effect of mismatching points.Remote sensing image radiometric correction can be divided into absolute radiometric correction and relative radiometric correction. In this paper, we think out the relative radiometric correction, using RANSAC algorithm to remove the abnormal pixels in two image overlap, thus find out the best linear regression with maximized data points to support the linear transformation model parameters. As to some radiometric correction with bad color balance, this paper puts forward the method of first cluster of pixels and then fitting transformation model.Finally, there are some improvements on the RANSAC algorithm to increase its operational efficiency. Based on the principle of RANSAC, two major improvement directions are the sample selection of each iteration and model parameter testing. In addition, describing pre-testing model parameters, this paper proposes sample pre-testing model based on the selection of sample. Through the sample pre-testing, we can not only greatly reduce computation, but also improve the sample selection efficiency as well as RANSAC operation efficiency.