Research
My work sits at the intersection of computational colour, image physics, and deep learning. I am especially interested in problems where physical constraints can guide model training.
A single photograph cannot tell a dark surface from a dimly lit one, so lighting leaks into the recovered albedo. CIAI trains on pairs of real photographs of one scene under different lighting and requires their albedos to agree. As a result, the albedo becomes much more stable in luminance. However, it is not yet stable in colour, because the explanation loss uses brightness only. The next stage splits the shading into a grey part and a colour part.
Fine-tuned a vision-language quality model to score portrait noise, texture, and contrast separately, for camera portrait-mode benchmarking.
Keeps a quality-1 JPEG as the base layer and adds a 10M-parameter neural codec for what the JPEG got wrong, so every file is still a valid JPEG. On Kodak it reaches 37.41 dB at 1.56 bpp, 2.1 dB above JPEG at the same file size.
Sends only the informative patches of an image plus a one-sentence caption. A Masked Autoencoder and a caption-guided Stable Diffusion refiner rebuild the rest at very low bit rates.
Coursework
Selected technical reports from COSI coursework across the consortium — Université Jean Monnet (France), NTNU (Norway), and the University of Eastern Finland.
A full classical stereo pipeline (calibration, SGBM disparity, ICP fusion, meshing) on ETH3D and Middlebury Chess. The fused point cloud reached a Chamfer distance of 0.043 m and F1 0.638.
A 450–950 nm hyperspectral inspection pipeline for AR-waveguide gratings, built with Dispelix Oy. Three metrics separated all 5 defect types, and a 4-operator Gage R&R study kept measurement error under 12%.
Benchmarked ANN, DNN, SVR, and RBF regressors against ICC-profile colour transforms on FOGRA51. The classical DemoIccMAX transform won on every metric (median round-trip ΔE00 0.5 vs 2.38 for the best ML model).
Measured an Ishihara plate with a hyperspectral camera and ten LED channels with a spectrometer, then used particle swarm optimisation to find LED mixes that cut the numeral–background ΔE from 31.8 to 7.5 or raised it to 57.7. All six lights were reproduced on a tunable LED source and matched the predicted spectra at r ≥ 0.997.
A controlled eye-tracking experiment on emotion-recognition gaze patterns, using dispersion-threshold fixation detection.