基于神经网络的数字图像几何畸变矫正方法

在获取数字图像的过程中,由于摄像镜头的非线性,往往会导致获取的图像存在严重的几何畸变.因此,在对图像做进一步的处理之前,需要对发生畸变的图像进行几何矫正.介绍了数字图像几何变换的基本原理,并在总结已有的矫正几何失真的方法的基础上,提出了一种基于BP神经网络的新方法.该方法用最速梯度下降法训练网络,使之能以较高的精度拟和出畸变图像的形变模式,最后利用反向双线性灰度插值的方法恢复出原始图像.Matlab仿真结果充分表明:该方法不但有效,而且具有更高的精度.

Algorithm for geometric distorted image rectification based on neural network

In the process ofgaining digital images,the nonlinearcharacteristic of the camera lens may lead to severe geometric distortion ofthe captured images.Therefore,geometric rectification of the distorted images is necessary before they can befurther processed.The basic foundation of geometric transformation is introduced,and after intensive research of those existing approaches for geometric rec-tification,a novel algorithm based on BP neural network is presented.The proposed method train the network with fastest-gradient-decent algorithm,which precisely fit the distortion pattern.At last step,reverse bilinear gray-level interpolation is applied to recover the original image.Simulation results in Matlab reveals that the algorithm proposed is both effective and of higher precision.