由二维图像恢复三维空间信息，首先要解决三维物方点和二维像点之间的对应关系问题，因此必须对视觉系统中的相机进行标定。相机标定技术是近景摄影测量和计算机视觉中首先需要解决的关键问题，为计算机视觉在相关的各个领域的广泛应用奠定了基础。现有相机标定方法主要分为传统的标定算法、自标定算法和主动视觉标定法。由于相机镜头存在着畸变，对于高精度测量应用，需要进行畸变校正。传统的标定算法中存在计算量大、模型复杂等不足。本文在分析相机标定原理的基础上提出基于神经网络的相机标定技术，该法比传统的标定方法快速，方便。在误差允许范围内能满足实际应用的需要。 人工神经网络是近年来发展起来的一门新兴技术，它可以处理那些难以用数学模型描述的系统，具有很强的自适应、自学习能力。而将神经网络技术应用于相机标定，无疑为标定技术的发展提供了一个崭新的思路。该方法由神经元传递网络信息，通过网络学习决定各神经元的网络权数，只需确定网络输入输出和约束条件，无须考虑中间过多的不确定模糊因素，即可建立空间点与图像点的映射关系，使系统不经过复杂的相机标定，直接恢复物体的三维信息，为立体视觉在测量和精确定位等方面的应用奠定了基础。

Two-dimensional image restores the three-dimensional information. We must firstly solve correspondence problem between the three-dimensional object space point and two-dimensional image point, so the camera of the visual system must be calibrated.Camera calibration technique is the key issue which close-range photogrammetry and computer vision must firstly solve, establishing the foundation of related areas which computer vision will be used extensively. Existing camera calibration method consists of the traditional calibration method, self-calibration algorithm and active vision calibration. Because there is the camera lens distortion, the high-precision measurement applications, the need for distortion correction. Calibration algorithm in the traditional calculation of large, complex models such as inadequate.Based on analyzation of the principles of camera calibration, this paper comes up with the camera calibration technique on account of neural network. The method is faster and more convenient than the traditional calibration method. And it can meet the needs of practical application in the range of allowable error. Artificial neural network is a newly developing technology in recent years. It can deal with the system which mathematical model is hard to describe. It has the powerful self-adaptive and self-learning capability. The neural network technology is used for camera calibration, undoubtedly providing a brand new thinking for development of calibration technology. The method is that the neuron conveys network information, network learning decides network weights of neuron, it only determines network input / output and constraints requirements without consideration of too many unstable and fuzzy factors, it sets up the mapping relationship between space point and image point, the system directly restores three-dimensional information of object without complex camera calibration, establishing the foundation for the stereo vision application of measurement, accurate positioning and other respects.