A Supervised Learning Method for Solving Space Resection of Single Image

The space resection of single image can be described as a problem of non-linear least squares, and the non-derivative, ill-conditioned coefficient matrix of normal equation and local extremum are main reasons for non-convergence in its numerical procedure. The spatial distribution of control points in different regions is not similar. If we put down the multiple images and have known their exterior orientation elements regarding the same control points under the same region, as a sample set, the overall descent direction can be obtained by supervised learning under the circumstance that every initial values of exterior orientation elements have been given. What's more, in the case of non-convergence in original space resection of single image because of the reasons mentioned before, it can be approximately solved by using the overall descent direction. From this angle, a supervised learning method for solving space resection of single im age is proposed. The process mainly includes:①Training stage, in which supervised learning process is utilized and the descend direction set of exterior orientation elements is obtained by solving the overall exterior orientation elements of images set with different attitude in the same survey area. ②Testing stage, in which for any image in study area, the exterior orientation elements can be calculated iteratively if the initial value and the descend direction set were given from the process of supervised training. Experimental results show that the method in this paper is more efficient in numerical procedure convergence and dependence of the initial value than the current there realized. Besides, it can overcome the non-convergence of Euler angles caused by the ill-conditioned coefficient matrix of normal equation, which is essentially the gradient matrix.