Research Of The Camera Distortion Correction Based On Neural Network

Geometric model of the camera imaging must be established in the fields of computer vision and image measurement applications for determining the relationship between the world coordinate of any one point of the surface of the object and the corresponding point in the image. We solve parameters of camera geometric model for calibrating camera. Camera calibration is an important field in computer vision and is the first step of computer vision too. Camera distortion correction is the. key point of camera model construction. According to camera distortion correction, we establish camera model and this step is the basis of the follow-up work. At the same time, this is the important way to improve the calibration accuracy. In this thesis, we mainly study the chessboard corner detection method and the distortion correction method.In this thesis, an algorithm for chessboard corner detection based on symmetry is proposed. Chessboard has its inherent characteristics, and we realize the checkerboard corner detection by using the checkerboard symmetrical characteristics. This method does not calculate gradient. So this algorithm greatly reduces the computation time, and is faster than the algorithm which is using gradient, and can avoid the disadvantages that the SUSAN algorithm can not detect the checkerboard corners.The research on description of camera distortion by distortion model had made progress, but it is difficult to precisely reflect the distortion of the camera. Because neural network has the approximate arbitrary nonlinear mapping capability, the method of using the BP neural network to describe lens distortion based on the traditional calibration is proposed. Then we use the trained neural network instead of the distortion model to realize lens distortion correction. Thus the camera calibration can get to higher accuracy.This thesis introduces PSO to improve the BP algorithm, and combined with chaos algorithm in order to achieve better distortion correction. The stability of BP neural network is relevant to the network value of the initial training. BP algorithm is easy to fall into local minimum. Particle swarm optimization algorithm is a global optimization algorithm, and this algorithm is easy to realize and has the advantage of fast convergence. Chaos is a common phenomenon in the non-linear system. Within a certain range, chaotic variables change with randomness, ergodicity and regularity. We use these characteristics of chaotic variables to search optimization, and make the algorithm jump out of the local optimum. This can maintain the population diversity and can improve the performance of global search algorithms. In the simulation studies we also show that the accuracy of the model that we established is improved.