Two dimensional translation detection by comprehensively calculating three cross-correlations

This paper suggests a new method for detecting 2D translation between two images based on calculating three independent cross-correlations (CCs) on them. Such a method is conceptually different from other area based methods which generally perform only one CC or its variants for phase shift detection. The principle of traditional area based methods could be interpreted as a fast but simplified implementation of least squares (LS), by ignoring two summed squares of given images while keeping one CC component between them. It is argued by us that such an ignorance often inevitably results in the requirement of data pre-processing for robustness and accuracy. Keeping all the source information but calculating the whole LS by three CCs, the computation performance is kept as O(N log N). Without any data pre-processing, experiments on a dataset with rich application backgrounds and comparisons with widely recommended methods including both the area based and the feature based methods, show that our suggestion is very promising for general-purpose 2D translation detection.