基于红外热像图的变电站远程智能监控系统，能够对变电站的电气设备进行同时、连续、实时的在线监控与故障诊断，对异常设备采取停电处理或监视运行等措施，从而保证电力设备的安全运行。 在电气设备的运行状态监测中，需要把同组设备相同部位在不同时期检测的结果进行比较，以便掌握设备运行状况及故障隐患的发展变化。然而，系统成像模块在经过一段时间扫描运行后，在预置点处会产生位置上的偏差，以及景深效应产生的透视影响。这些会对后台图像处理造成影响，难以对要检测的对象准确定位，造成漏检和误判。因此应对发生偏差的图像进行配准，消除位置差异和透视影响。 本文研究了图像配准的基本框架和常用方法，针对本文图像特点，采用基于特征的图像配准方法。作者首先对传统的基于Harris角点匹配的方法进行了改进，将基于傅里叶变换的相位相关法应用于其中，通过得到图像间的主运动方向，减少了Harris角点匹配时的搜索区域和时间，并能够剔除误匹配点对。实验表明该算法具有一定的适用性和较高的匹配精度。其次，本文将当前图像配准和识别领域的新技术——SIFT算法应用到红外遥视图像的配准中，并将算法中的匹配策略进行改进，使用相位相关法引导匹配，减小了搜索时间；并且通过调整匹配时的比例阈值使匹配的点对控制在5~10以内，提高配准效率。最后本文对两种配准方法在适用性、配准精度和配准时间上进行了详细比较。在适用性方面，基于相位相关和Harris角点匹配的配准方法，有一定的适用性，但存在失配现象，总体正确配准率达到90.9%；基于改进的SIFT算法的配准方法能够正确配准所有图像，正确配准率为100%。在配准精度方面，对于正确配准的图像（除失配图像外），两种方法均能够达到亚像素级，但基于改进的SIFT算法的配准方法的配准精度总体上更高。最后，在配准时间，基于改进SIFT算法的配准方法也更稳定。基于此，本文将基于改进的SIFT算法的配准方法确定为系统的配准算法。

The substation remote monitoring system which based on infrared thermal images is a new kind of monitoring----visual surveillance. The system can carry out real time online monitoring and fault diagnosis to multiple substation equipments simultaneously and continuously, it cuts down the out-of-order equipments or keeps monitoring them. Thus ensures the safe operation of electrical equipments. However, the acquired sequence infrared images have position deviation, this disable the system normally monitoring the interested equipment. Thus the deviated images need to be registered. In this paper, we have researched the basic framework and common methods of image registration. And adopt feature-based image registration method according to the character of our images. The writer firstly improved the classical Harris-corner matching approach, using the phase correlation method which based Fourier transform to acquire the main direction of movement between two images. This greatly reduced the searching area and computing time of matching, and could eliminate false matching points. Experiments show that this algorithm has some applicability and high matching accuracy. Secondly, this article presented a new method—SIFT algorithm which used in image registration and pattern recognition recently. We also have improved the matching strategy of the algorithm, using the phase correlation again to guide the matching step to acquire a small searching area and consuming time. And controlled the matching points within five to ten by adjusting the ratio of matching threshold, this greatly improved the efficiency of registration. Combine these two methods, the improved SIFT algorithm demonstrated better performance in applicability, accuracy and consuming time.