Learning in correlators based on projections onto constraint sets Academic Article uri icon

abstract

  • The problem of teaching a correlator to classify many patterns can be described as an optimization problem.' In the case of shift-invariant pattern recognition, we sometimes deal with the optimization of an error function with a huge number of variables, and therefore we must find efficient algorithms that can handle a problem with reasonable complexity and in a relatively short period of time. Recently a method2 to calculate a synthetic discriminant function3 (SDF) from a given training set of objects was proposed based on the well-known projections- onto-constraint-sets (POCS) algorithm. 4 The learn- ing procedure has been employed on a simulated joint-transform correlator in order to find a reference function that can then distinguish between two object classes. In this Letter the learning method and its tasks are modified. The main goal is to introduce a general procedure for synthesizing SDF's from a given train- ing set of any size. At the …

publication date

  • January 1, 1993