3D Reconstruction And Digital Image Authenticity Check Base On Stereo Vision

The goal of computer vision is using computer to models and automate the process of visual recognition. As an important branch of computer vision, stereo vision imitates biological vision to obtain the description of the external scene and dimensional camera movement by analyzing the information from two or more images of the same scene. Base on the perception from the images, the scene can be reconstructed. stereo vision received increasing attention in recent years, especially in medical image processing and understanding, robot control, industrial manufacturing, virtual reality.Three-dimensional reconstruction based on stereo vision can be divided into the following steps:image acquisition, feature extraction, image matching, camera calibration and 3D reconstruction. This paper focuses on the robust estimation of fundamental matrix which gives expression to the epipolar restriction between the pair of images. Base on the study the strengths and weaknesses of several common robust methods, such as RANSAC, M-estimate and LMeds. an improved self-adaption LMeds algorithm base on random sampling (RSALMeds) Is proposedThe main idea of RSALMeds algorithm is:first, an explore space is formed by using random sampling, in this explore space, the optimal solution meeting to a certain Confidence level can be found. Unlike RANSAC algorithm, the optimal solution of fundamental matrix is not based on the number of inliers but the Least Median of the geometric error. Then M-estimator and cost function are given base on the optimal solution. Through the optimization on global data, the exact solution of fundamental matrix is obtained. In the other hand, Least Median of the geometric error given gives a threshold for inliers and outliers, base on the threshold, the error matches are eliminated. Finally, dense matching is completed by guided matching.RSALMeds algorithm need multiple sampling to get the optimal solution, sampling frequency depends on the desired confidence level and error assumptions. I n order to improve the efficiency of the algorithm, A self-adaptive process is used to determine the number of samples. Through real-time update error estimation and sampling frequency, unnecessary iteration processes are avoided.In order to improve the stability of the algorithm, uniform random sampling is used in RSALMeds, in which the data points for each sample is not particularly close. In the other hand, each sample is pre-tested to exclude the bad sample contain collinear points redundancy and coplanar points redundancy.The procedure of RSALMeds algorithm is as follows:Step 1. Base on image blocks, indexing the distribution of matching,Step2. Start RSALMeds robust estimation, get the minimum set of sampling points using uniform sampling method, in each sampling, execution the next steps:a) Detect the stability of the sampling (include collinear points redundancy and coplanar points redundancy), if the pre-test fails, turn to re-sampling, else, go to step b)b) Estimate the fundamental matrix Fi c) Calculate the robust standard deviation of estimated Fundamental Matrix Fi record the amount of inliers and proportion of error matching.d) Updated optimal solution F, if Fi has smaller robust Standard deviation and adaptive the sampling frequency K.e) If K(?)currently sampling frequency, go to step3. Else return to step2.Step3. Base on the optimal solution, structure the M-estimator and cost function, optimize the cost function on global data. Obtain the exact solution of fundamental matrix. In the mean time, eliminate the error matchings.As a robust statistical method, RSALMeds algorithm can be applied to other estimation of model parameters which bases on over-determined data with errors.In next part of the paper, a three-dimensional reconstruction system based on stereo vision is designed base on RSALMeds and the reference to other matured three-dimensional reconstruction system, This reconstruction system, can achieve sub-pixel corner feature detection and robust image dense matching. Using plane template calibration, inner parameter matrix will be calculated. Then the camera position and posture in 3D space is recovered from essential matrix. Through triangle intersection, we can reconstruct the spatial structure of the object. In the end, by using computer graphics technology, the object will be real reconstruction.In the final part, the paper focuses on stereo vision based digital image authenticity check techniques. Base on perception the geometric information from the images, quantified and visual test results can be used to check digital image authenticity. the stereo vision based digital image authenticity check in this paper base on epipolar geometry and automatically obtained three-dimensional information of the object.1. Epipolar geometry based check method: the epipolar error: l1= Fx2 is the epipolar line match to x2,l2=Fx1 is the epipolar line match to x1,d(*,*) is Euclidean distance.Compare the epipolar error between the checked object and the scene. Come to conclusion whether the checked object and the scene came from the same imaging process, achieve the purpose of testing the authenticity of the image.2. automatically obtained three-dimensional information check method:Base on the spatial structure of checked object restored from images, analyze their structure and location information (including size, depth, angle, etc.) to determine the authenticity of the image.