There are position errors between the images from each split optical path of the amplitude-division imaging polarimeters, image registration between the images from each split optical path is a precondition for polarized detection. In order to solve the problem that the features of the target are not obvious and difficult to extract, and that the change of gray scale of the image is large, a function of similarity measurement was proposed and it was suitable for the image registration of the amplitude-division imaging polarimeter. Based on the function, the images from each split optical path had been registered. First, according to the principle that the position error between the images can cause the abnormal information areas in the polarized image, the algorithm of extracting the similarity metric function was discussed. Next, according to the characteristics of the imaging system, the parameters of the geometric transformation between images were determined. Then, the genetic algorithm as the parameter optimization search algorithm was used to get optimal parameters of the geometric transformation between images, and the image registration algorithm was finished. Finally, the algorithm was verified by using constructed images and collected images respectively, and the image mutual information value (MI) is determined as the metric of the accuracy of image registration. The experimental results show that the MI between reference image and registered sensed image is 2.692 5 for the constructed image, and it is 1.849 3 for collected images.It indicates that the accuracy of registration by using this method is higher than that by using method based on feature. It can satisfy the registration requirements of the system.