|M.Sc Student||Gefen Gal|
|Subject||Biologically Based Model of Pattern Recognition: Computation|
|Department||Department of Biomedical Engineering||Supervisor||ASSOCIATE PROF. Moshe Gur|
Identification of transformed objects (rotation, shift, scale, illumination etc.) is one of the human visual system’s basic capabilities. It is known that this capability requires the combined activity of several areas of the visual cortex, but there is no model that successfully describes the process.
The generally accepted model is the Hierarchical Processing (HP) Paradigm. It describes a sequential recognition process, in which each area of the cortex produces data based on the output of a preceding area. According to this paradigm, cortical area V1 encodes only basic features and is at the base of the hierarchy. Other cortical area’s neurons represent more elaborate objects, until, at the summit, elaborate objects are represented by small groups of neurons. This model is attractive, but is partially contradicted by biological data.
An alternative approach is the NOrmalized V1 Ensemble (NOVE) Hypothesis. This theory claims that the recognition process is based on a parallel and distributed system. While V1 is the main area detecting the objects features, other areas operate in parallel on the visual information to acquire knowledge on the transformations the object has been subjected to. This data is used to normalize the activity of V1, thus creating a unique pattern of activity for each object. V1 activity itself then becomes the basis for recognition.
In my research I have implemented a computer simulation of the NOVE model. A method for achieving orientation, shift and illumination invariance was used. The method is based on finding an object-centered main axis and normalizing the object by aligning this axis with the viewer-centered axes. Then object features are extracted and compared to prototypes in memory. The method was tested on both identification and categorization tasks, using several sets of handwritten characters. The results indicate that good identification can be achieved using such a method. The applicability of the method to categorization task is, however, uncertain and requires further studies.