Assessing automatic registration methods applied to digital analysis of historical documents

We first review various cases of interest in document image analysis where spatial registration is necessary. For reasons related to both the optical sensors and the acquisition procedures, the different image components in multispectral-multiview images can be spatially misaligned. This happens commonly in color or multispectral imaging and in recto-verso acquisition, and many image processing and analysis tasks need to work with precisely located data points. We report the results of our evaluation of several registration methods that do not require human intervention, showing that registering recto-verso pairs is normally much more difficult than aligning multispectral channel images.