Obtaining image sequences has become easier and easier thanks to the rapid progress on optical sensors and robotic platforms. Processing of image sequences (e.g., mapping, 3D reconstruction, Simultaneous Localisation and Mapping (SLAM)) usually requires 2D image registration. Recently, image registration is accomplished by detecting salient points in two images and nextmatching their descriptors. To eliminate outliers and to compute a planar transformation (homography) between the coordinate frames of images, robust methods (such as Random Sample Consensus (RANSAC) and Least Median of Squares (LMedS)) are employed. However, image registration pipeline can sometimes provide sufficient number of inliers within the error bounds even when images do not overlap. Such mismatches occur especially when the scene has repetitive texture and shows structural similarity. In this study, we present a method to identify the mismatches using closed-loop (cycle) constraints. The method exploits the fact that images forming a cycle should have identity mapping when all the homographies between images in the cycle multiplied. Cycles appear when the camera revisits an area that was imaged before, which is a common practice especially for mapping purposes. Our proposal extracts several cycles to obtain error statistics for each matched image pair. Then, it searches for image pairs that have extreme error histogram comparing to the other pairs. We present experimental results with artificially added mismatched image pairs on real underwater image sequences.