A landmark-based framework for image comparison in the presence of various geometric misalignments

Image surveillance has now become an emerging research field with diverse applications across various real-life scenarios, such as medical science, satellite imaging, and so forth. Image comparison is a fundamental problem in this area, with the primary focus on detecting shape changes in image objects. However, the objects in the images are often not geometrically aligned, and hence require an image registration step during pre-processing. The ill-defined problem of image registration in the presence of a shape change in the image object poses significant challenges that have yet to be explored in depth. This gap is bridged by the proposed image comparison method for the foreground image object that is invariant under geometric mismatch due to translation, rotation, and scaling (TRS). Invariance to such transformations helps avoid the ill-defined problem of image registration, which leads to better computational efficiency, higher accuracy, and thus more reliability, compared to its state-of-the-art competitors. The central idea involves characterizing the shape of the image object by using the locations of selected landmark points, and then using a specially defined translation-rotation-scale invariant metric for comparison. Both theoretical justification and practical applications demonstrate its effectiveness and wide applicability.