Tactile pattern recognition using neural networks

This thesis presents a system for the capture and recognition of tactile images using neural networks. Neural Networks utilizing the backpropagation technique are used to provide a general purpose recognition engine for several classes of pattern recognition problems. Examples of successful networks are presented with discussion of the results and methodology for development of each. The system consists of the following: Image capture, training of network using an iterative approach and testing of the network against independent images not present during training.;Pattern capture is performed by scanning a force sensitive tactile sensor that interfaces to a general purpose computer. Following capture, examples of tactile patterns of desired types are stored in a training file and the training goal of the network set. The goal is determined by the "trainer" who when the patterns are captured indicates the pattern type. Related patterns are given the same class name. The Network is required to consist of as many output neurons as classification types. The goal is that an output neuron becomes "activated" when its pattern types are present and "de-activated" when another type is present.;Patterns in the training file are then recursively applied to the inputs of the Neural Network. Once the Network converges to the desired goal it is tested against a new set of patterns to determine if the network has learned to apply generalization in its recognition of the patterns. The training set and network topology may be modified in heuristic fashion until satisfactory results are achieved.