Abstract : In this work we are interested in the analysis and the extraction of the 3D information (orientation and shape) contained in natural scenes and homogeneous textures.
For this, we adopt a multidisciplinary approach of the modeling of the visual system.
We first present psychophysical experiments aiming at evaluating the relative contribution of the frequency variation and of the linear perspective cues involved in 3D perception.
To do so we have created purposely designed stimuli representing homogeneous textures made of Gabor patches displayed on a planar surface.
The plane is viewed under perspective projection with particular slant and tilt angles.
The frequency and the orientation of each Gabor patch are set according to the local frequency gradient and the local linear perspective defined by the projection.
We synthesise textures presenting a frequency variation alone or an orientation variation alone or both kind of variations (in combination or in conflict).
For each texture, a tilt and a slant discrimination task are performed.
The frequency variation cue appeared to dominate over the linear perspective cue for slant estimation.
However both cues are involved in the tilt estimation.
These results validate the use of our stimuli for 3D perception study and the decomposition of the texture cue into elementary components.
Based on this approach, we present a biologically plausible model of the frequency variation analysis in the cortical area V1.
We model the complex cells responses with log-normal filters which present different theoretical and practical advantages against the classical Gabor filters.
The algorithm is composed of a pre-treatment stage corresponding to a retinal filtering allowing to keep only the texture information and of a decomposition of the image into local patches similarly to the cortical cells receptive fields.
A robust technique aiming at estimating the local mean frequency, independently of the orientation information, and corresponding to a simple combination of the whole set of filters is applied to every patch.
The measure of the local frequency variation between each patch allows to estimate the tilt and slant angles of the studied surface and its shape.
The method is evaluated on different images and textures databases.
It appears to be comparable in precision with the best known techniques and can be applied to irregular textures with a lower computational cost.