Brain model recognizes images

During visual perception and recognition, the human eyes move and fixate successively at the most informative points of an image. "Because of this," says Ilya Rybak of E. I. du Pont de Nemours Neural Computation Program (Wilmington, DE), "perception and recognition must be treated as behavioral processes that include both the sequential image scanning by an `attention window` and parallel processing of image fragments within the "attention window."

Mar 1st, 1997

Brain model recognizes images

During visual perception and recognition, the human eyes move and fixate successively at the most informative points of an image. "Because of this," says Ilya Rybak of E. I. du Pont de Nemours Neural Computation Program (Wilmington, DE), "perception and recognition must be treated as behavioral processes that include both the sequential image scanning by an `attention window` and parallel processing of image fragments within the "attention window."

Pattern recognition consists of consecutive eye movements and verification of expected image fragments recalled from memory. To process these data, the human visual system contains two pathways: one that leads to the parietal cortex, which is involved in processing and representation of spatial information, and one that leads to the inferior temporal cortex, which deals with processing and representation of object features.

Using this as a basis, Rybak and his collaborators from A. B. Kogan Research Institute for Neurocybernetics (Rostov-on-Don, Russia) have built a model of the way humans recognize images. In operation, the attention window transforms the image into a retinal image--one that simulates the decrease in resolution from the fovea to the retinal periphery. This image is then used for primary feature extraction--a function of the primary visual cortex.

To perform pattern recognition, the image is processed at successive fixation points. At each point, a set of edges is extracted and transformed into invariant second-order vectors that are stored in sensory memory. Each eye movement is memorized in motor memory. "These two memory traces alternate as a chain of events that can be considered a behavioral program for image recognition," says Rybak. For recognition to occur, a series of successful matches must occur; otherwise, the model returns to an image search mode. The developed model can recognize complex images (for example, faces) invariantly in respect to shift, rotation, and scale. A demo is available on the World Wide Web at http://www. voic enet.com/~rybak/bmv.zip.

For more information, contact Ilya Rybak at rybaki@eplrx7.gifs.duPont. com or on the World Wide Web at http://www.voicenet.com/~rybak/.

More in Scientific & Industrial Research