A Modular Model of Active Vision

Modularity is the main characteristic of the biological systems and especially of the brain 2]. For example , Vision can be described as the interactions of diierent processing systems 1]. Physiologists relate two pathways from the retina. The rst conveys information from the retina to the cortical areas, where specialized areas analyse the visual data (e.g., colour vision, motion detection, object recognition and lo-calization). The second joins the retina to the superior colliculus, which sends motor signals to extraoc-ular neurons (e.g., saccade from one important zone in the visual world to another one). We described a biologically inspired connectionist model of the visual system. Neural networks, despite their ability to solve a lot of hard problems, are rarely modular. Connectionist models usually do not take into account eye movement informations underlying the attentional processes 5], except a recent study on reading 10]. On the contrary, our system involved three modules. The rst (Foveal Retina) received as input the central part of the retina (i.e., a high resolution image) and computed a local clue on this region. The second (Peripheral Retina) received information from the peripheral retina (i.e., a low resolution image) and computed movement parameters (i.e., angle and amplitude of the movement). The inputs of the Foveal Retina and Peripheral Retina simulated the activity of ganglion cells of small and large receptive elds respectively. These activities were implemented using two-dimensional wavelet transforms. Both Retina modules were feed-forward networks. The last module (Cortical Integration) integrated the foveal clues to recognize objects. Its architecture was the following: the input layer encoding one local clue, a fully recurrent hidden layer allowing the integration of the information, and an output layer encoding global decisions (e.g., labels of diierent alphanumeric characters). Thus, the nal decisions were taken when salient features were extracted from the visual world. The learning strategy for the Central Retina module was particularly original: we used the generalization abilities of the Cortical module as a positive (or negative) reward to modifying its connections. The learning phase of the Peripheral Retina module consisted in teaching predeened eye movements and updating connections using either the back-propagation algorithm or a pseudo-inverse rule 9]. Learning of the Cortical's connections was realized using a mod-iied version of the Alopex algorithm 11]. Preliminary results showed that our approach can easily perform the recognition of handwritten characters. We think that this modular active vision model could solve many image …