Dynamic selection of primitives in the metric model
This thesis has three main objectives. First, to initiate the investigation of the new degrees of freedom offered by the recently proposed pattern recognition model, specifically to the process of selection of the low-level primitives for structural pattern representation. Second, to suggest that not only should the latter process not be separated from the (high-level) recognition process, as has been practised so far, but, moreover, it should be driven by it. Third, to continue the investigation of the primitives selection model proposed by Muchnik (1974). The metric model for pattern recognition recently proposed by Goldfarb (1990) makes use of the concept of distance to classify the patterns during the learning stage. The optimal value of the objective function used in the learning process can also be used in a feedback loop to measure the appropriateness of the low level primitives chosen by the Muchnik model. An implementation of this model for the handwritten digits is presented. The results of the experiments together with the suggestions for future work are also presented.