Publications
Selected work spanning generative restoration, aerial imaging, scientific data and applied machine learning.
A degradation-agnostic face restoration method that optimises in the latent space of a pretrained generator, handling denoising, deblurring and inpainting without retraining per degradation.
Pivotal tuning inversion anchors a blurred observation to a generator's latent manifold, recovering sharp detail without paired supervision or a known blur kernel.
Denoising a single image with no training and no clean reference, using CLIP as a semantic critic to steer restoration towards a perceptually faithful result.
A dataset and recognition pipeline for historical stone inscriptions, where weathering and archaic glyph forms break standard scene-text OCR.
A stress test of unsupervised StyleGAN restoration under degradations it was never tuned for, showing where latent-space optimisation holds up and where it silently fails.
An aerial and ground image dataset covering a full cotton growth cycle from weekly drone missions at 10 m, 15 m and 115 m, for crop health and pest analysis.
A predictive framework for Parkinson's disease analysis pairing classification accuracy with descriptive feature analysis so clinicians can see which measurements drive each prediction.