Advancing Chromagenic Computer Vision (FINLAYSONU16SCI)
Film and digital are, respectively, the defining technologies of the 1st and 2nd eras of photography (and imaging). We are just entering the third and most disruptive era of so-called computational photography. Unlike film and digital cameras, which are built in analogy to how we see (a lens, aperture and sensor array), in computational photography the basic design of the camera is changed so that certain problems are made easier to solve. Novelly, in this project we will develop a computational photography solution which also has relevance to how we ourselves see.
This project will develop a camera architecture where two images are taken of each scene, the first as normal and the second through a specially chosen coloured filter. This ‘chromagenic’ design was proposed previously [i] with the specific purpose of helping to find light changes in images. However, the performance in solving this task – e.g. shadow detection – required precise (laborious and very difficult) calibration and then only worked well for small exposure ranges. The central aim of this project is to re-engineer the chromogenic camera, incorporating recent work on exposure independent processing [ii] and pattern recognition [iii], to work robustly.
Plausibly, the eye itself, with eye movements, is also a chromagenic camera [iv] as the central foveal region is pre-filtered with the macular pigment (while there is no pre-filtering of the periphery). Equally, recent work [v] proposes that we are at least tetra-chromatic (four sensors rather than three) in our peripheral vision. In the research project a link will be made to the developed computational theory and studies of human vision.
This PhD is funded by Faculty of Science and Apple. The studentship is funded for 3 years and comprises of home/EU fees, an annual stipend of £14,057 and some funds to support research training. Overseas applicants may apply but they are required to fund the difference between home/EU and overseas fees (in 2015/16 the difference is £9,498 but fees are subject to an annual increase).
i) Finlayson, G.D. et al (2005). Illuminant Estimation using the Chromagenic Constraint. IEEE conf. on Comp. Vis. And Patt. Rec., 20-25.
ii) Finlayson, G.D. et al (2015). Root Polynomial Colour Correction. IEEE Transactions on Image Processing. 1460-1470
iii) Lou, Z. et al (2015). Color Constancy by deep learning. British Machine Vision Conference, 76.1-76.12
iv) Finlayson, G.D. and Morovic, P.M. (2005). Human visual processing: Beyond 3 sensors. In IEE Visual Information Engineering, 1-7.
v) Horiguchi, H. et al (2013). Human Trichromacy Revisited. Proc. Nat. Acad. Sci, 110(3): E260–E269