Hyperspectral imaging is becoming widely adopted in the analysis of paintings and
there is a trend towards capturing hyperspectral images at higher spatial resolutions.
At higher spatial resolution, it is possible to identify more detail in an image
(e.g. painting technique or pigment fragments) and better inform conservation
treatment. There is a trade-off between hyperspectral imaging parameters such as
spatial resolution, spectral quality and acquisition time, where an improvement in
spatial resolution can be traded for improvements in other imaging parameters such
as spectral quality. This work presents two methods towards the development of
high-resolution hyperspectral imaging in chapter 3 and chapter 4.
In chapter 3, style-transfer machine learning techniques are investigated to
improve the spatial resolution of short-wave infrared sensors in conjunction with
higher resolution image sensors such as a DSLR camera. A spectral imaging
pipeline is proposed which uses style-transfer techniques, with reduced acquisition
time compared to image mosaicking approaches.
Chapter 4 investigates image mosaicking, combining multiple high-resolution
hyperspectral images covering a narrow field-of-view. The images are acquired
from a line-scan hyperspectral camera which can exhibit dropped frames, resulting
in misalignments in the composite image. A method is developed to correct for
dropped frames in high-resolution line-scan hyperspectral images. At high spatial resolution, three-dimensional aspects of a painting surface become
more apparent. There are also numerous cultural artefacts which are threedimensional
and would benefit from hyperspectral imaging, though methods to
achieve this are not well established. One specific challenge comes from the spectral
calibration process which often assumes a flat two-dimensional surface. The work
in chapter 5 introduces a method to geometrically calibrate a line-scan hyperspectral
camera with a video projector, enabling depth information to be extracted from
the hyperspectral image via triangulation. A new method for spectral calibration in
three spatial dimensions is also proposed.