A recurrent neural architecture mimicking cortical preattentive vision systems

Low- and intermediate-level tasks of visual perception are based on multiple and cooperating simple agent architectures, made up of networks of formal neurons organized into a three-layer neural network locally connected. In each layer, there is a specific intermediate representation of the image: the first layer extracts oriented feature elements; the second and third ones are used to control and coordinate interactions among features. Inter- and intra-layer recurrent connections are responsible for the integration of the different computational tasks performed by each neuron into a non-local percept. The output of cortical neurons also depends on the excitation of neighboring neurons. Applications to texture discrimination in both natural and artificial textures are presented.