Cet article a été publié dans un numéro de la revue, cliquez ici pour y accéder
Multispectral imaging techniques are widely used to analyse and restore digital images of ancient documents degraded over time. In particular, acquisitions in infrared and ultraviolet bands can reveal information invisible by naked eye, which is not captured by conventional RGB imaging. Multispectral acquisitions are usually performed with filter-wheel cameras that mount a series of interference filters in front of the sensor to select the spectral band in which the document has to be acquired. Due to the use of different filters, the focus of the lens changes and it is necessary to adjust it manually for each filter, because such imaging systems are devoid of automatic focusing. Thus the acquired images may not be perfectly in-focus and have misalignments. In this work we present an automatic solution to acquire a multispectral data cube of aligned images that can be used in document analysis to extract and/or separate information through enhancement techniques. A custom-made motorized autofocus system controlled by software has been installed on a filter-wheel camera, which allows us to acquire in-focus images automatically at each filter changing, and a registration method based on Fourier-Mellin transform aligns these images. A preliminary calibration is performed to set the focusing of each filter with the working distance. As a result, the camera can capture images independently from the contrast of the scene, also in presence of documents with homogeneous texture. The proposed solution reduces considerably the acquisition time and offers a tool to acquire automatically the data cube to be used in further image analysis techniques. The results of experimentations are hereby presented and discussed.
Keywords : Multispectral imaging, Focusing, Image registration, Document analysis