Objectives. In the framework of biological models of vision, we assume that the outputs of complex cells of the visual cortex are suitable for analyzing local and global properties of the retinal image. These outputs sample the local Fourier amplitude spectrum of images.
Method. We propose a model of the first stages of human visual system that provides efficient image description features. These features are insensitive to a large number of variations: Illumination (solved by the retinal preprocessing), Spatial position (solved by local Fourier amplitude spectrum), Image scale and rotation (solved by biologically plausible Log-Normal filters in lieu of the classically used Gabor filters). The model output is compared to psychophysics experiment results. This comparison both provides explanations of perceptual mechanisms in terms of the structure of the model; and reciprocally suggests model improvements, eg. by choosing more diagnostic filters to suit images and/or tasks.
Results. We are using this model in various image analysis applications such as categorization, extraction of saliency maps, or local estimation of perspective. The model predictions are confirmed by the behavior of observers, either when categorizing natural scenes under various conditions: rotation, image size and screen distance, or when estimating 3D texture orientations.