分时型长波红外偏振成像系统图像配准研究

Study of Image Registration in the Division-of-Time Long-WaveInfrared Polarization Imaging System

针对分时型长波红外偏振成像系统在图像采集过程中,由于在不同偏振方向获得的原始图像存在平移或者旋转的问题,使用自行搭建的分时型长波红外偏振成像系统拍摄的2组不同复杂程度场景图像作为实验对象,利用3种经典的特征检测算法(Harris角点、尺度不变特征变换(Scale-Invariant Feature Transform,SIFT)和KAZE)和2种匹配器(Brute Force和Flann)进行图像配准实验。结果表明:KAZE算法检测出的特征点数比Harris和SIFT算法少,但与Brute Force匹配器搭配时的配准正确率分别为79.58%和59.53%,远大于Harris和SIFT算法的配准正确率;KAZE与Brute Force组合作为偏振成像系统的图像配准算法时,配准后的图像中物体的几何形状和边缘轮廓更加清晰,很好地减少了像素偏差带来的影响。

In the process of image acquisition by the division-of-time long-wave infrared po-larization imaging system,original images obtained in different polarization directions may have small translation or rotation.To solve this problem,two sets of images with different complexi-ties of scenes captured by the division-of-time long wave infrared polarization imaging system built by the paper were taken as experimental objects.Thereby,image registration experiments were conducted by three classical feature algorithms including Harris corner detection algorithm,scale invariant feature transform(SIFT)and KAZE as well as two matchers including Brute Force and Flann.Results showed that the number of feature points detected by the KAZE algo-rithm is less than that of the Harris and SIFT algorithms.However,the registration accuracy of the KAZE algorithm with Brute Force matcher are 79.58%and 59.53%respectively,which are much higher than the other two algorithms.Moreover,the geometric shape and edge contour of the object in the registered image obtained by the image registration algorithm combing KAZE and Brute Force are more clearly visible.It indicated that the proposed method could greatly re-duce the influence of pixel deviation.