Research Line:
Visual-statistical models for image representation.
Application to texture analysis/synthesis, image enhancement and restoration.
Summary:
Choosing a proper image representation is a key issue for the success of any computer vision (CV) or image processing (IP) task. In the last years, such representations are being designed based on two fundamental pillars 1) the statistics of natural images, and 2) the human visual system (HVS) non-linear and non-uniform response to visual stimuli. Indeed, both issues are closely related, as biological visual systems have evolved towards an optimal information extraction from the environment, according to the relative frequency and relevance of the input stimuli. In addition, matching digital image representation to human visual representation may only increase the performance of both computer vision tasks (aimed to emulate HVS scene analysis abilities) and image processing tasks (for which the final target is precisely the HVS).
We know from neuro-physiology (experiments on apes) and visual psycho-physics that early stages of the HVS involve a local image decomposition into different orientations and levels of detail of the input image. According to recent studies based on information theory, that kind of representation, which can be modelled as a wavelet linear transform, is specially suited for natural image statistics. In fact, wavelets have become a standard tool for image compression, texture analysis/synthesis, restoration, etc. However, such linear transformation of the input images is just a first approximation to the internal representation of the early HVS. There remains a gap between low-level vision, whose goal is performing a common, multi-purpose, pre-processing of the input, and mid-level vision, responsible for extracting coherent features from the processed input (such as texture labels, contours, etc.). In that sense, higher order statistical dependencies found among wavelet coefficients point out to the necessity of including non-linear interactions in the representation to account for that statistical coupling. Important visual effects, such as visual masking, depend directly on these interactions and have been recently modelled by normalisation schemes. My research is about that type of models and how they may be used to increase the performance of the applications.