Scene variability and perception constancy in the visual system: a model of pre-processing before data analysis and learning

International audienceHell in data analysis is paved (at least) with variability and noise. Is there some lost garden of Eden? Is there some way to approach it? In this paper, we take the example of Human visual perception and show how our visual system manages to process visual information in such a highly efficient way that it is able to categorize images or scenes within ranges of 100-150 ms, whatever the viewing conditions. In fact, before any high-level recognition task, the visual system unfolds a series of pre-processing stages to reduce image variability: * In the retina: a first adaptation process to global and local illuminant intensity and color and second one to local contrasts allow to extract information equally in the whole image. A spatio-temporal filter allows a spectral whitening of images so that all spatial frequencies are equally represented. * In the primary visual cortex: the estimation of the local power frequency spectrum allows a relative insensitivity to image translations. Sampling the frequency spectrum on a log-polar basis and by means of Log-Normal filters allows to easily process zooms and rotations and also to estimate local perspective. * In the cortical area V4, a further Fourier Transform of the log-polar spectrum provides insensitivity to zooms and rotations, as well as to perspective transformations. We can consider images as represented by high-dimensional vectors, the components of which are the image's energy extracted by each frequency filter. Taking advantage of all these preprocessing stages, data variability can be expected to be optimally reduced for comparison purposes. In the last part of this paper, we give an example of image categorization by means of CCA, a self-organizing neural network, which, after noise reduction, reduces dimension and unfolds the manifold where the data are embedded. It is shown that the network, taught by human subjects performing the same task, is able to improve its behavior by providing a better separation of classes