Abstract

From Computer Vision (RGB+NIR), to Medical (Diffusion tensor imaging) to defence applications (multispectral satellite images) there are many applications where images contain more than the 3 RGB colour channels. Such images are not easily visualised since we are trichromats (can only see with three degrees of freedom). Further, the more dimensions there are the more difficult it is to process these images. Thus, there is a large literature of techniques for fusing multispectral images to lower dimensional counterparts.

Conventional methods for image fusion suffer from two serious problems. First, they either map N-dimensions to 1 or their outputs are ‘pseudo colour’. In neither case is there a link to what we ourselves see (necessary for many applications including photography). Equally, many colour image processing algorithms are useful only if they do correspond to our own vision (e.g. they are based on perceptual notions such as ‘hue’ and ‘saturation’). The second problem of existing approaches is that – in fusing the images together – they often introduce artefacts that were not present in the original (i.e. in any of the multidimensional image planes).

What we call Spectral Edge is a new paradigm for image fusion which, uniquely, produces ‘true colour’ fused outputs and, by the construction of our algorithms, delivers image fusion without artefact. In our talk we will present the various theoretical insights underpinning our method. Our initial research has prototyped our work in the areas of surveillance (RGB+thermal), remote sensing (Landsat), medical imaging and improving displays for colour blind observers. We present the Spectral Edge outputs for each of these applications.

As part of the talk, I will also introduce some of the other research carried out in the ‘Colour lab’ at UEA .