An unbiased implementation of regularization mechanisms

Perceptual processes, in computer or biological vision, require the computation of “maps” of quantitative values. The image itself is a “retinotopic map”: for each pixel of the image there is a value corresponding to the image intensity at this location. This is a vectorial value for color images. A step further, in early-vision, the retinal image contrast is computed at each location, allowing to detect image edges related to boundaries between image areas. Such maps encode not only the contrast magnitude, but several other cues: contrast orientation related to edge orientation, shape curvature, binocular disparity related to the visual depth, color cues, temporal disparity between two consecutive images in relation with visual motion detection, etc.. There are such detectors in both artificial and biological visual systems (see e.g. [12] for a general introduction and e.g. [17] for an overview about biological vision). Such maps are not only parametrized by retinotopic locations, but also using 3D locations, or other parameters.