Abstract: Image registration is a key step in many real applications, such as image fusion and change detection. However, after transforming the reference image, it is usually difficult to assess whether the transformed image is indeed registered with the target image. The underlying reason is that a suitable measure is unavailable currently in literature to adequately evaluate the similarity of two images. In contrast to the traditional evaluation methods, in this paper, the registration evaluation problem is converted into a classification problem through machine learning. Experiments based on 400 pairs of images demonstrate the validity and reliability of our proposed method. Although the proposed method is specifically designed for Fourier-Mellin transform based registration technique, its basic principles could also be applied to other registration techniques.