The classification of handwritten digits through an analog feature extractor chip and a neural classifier is discussed in this paper. The chip implements a feature extraction algorithm onto analog circuits; it extracts a set of 112 features from the input character (32×24 binary pixel matrix). The features, coded by current signals, are given in input to a neural classifier which performs the recognition task. The chip validation results are reported: a set of handwritten digits have been classified by a neural network implemented by a software simulator. The resulting classification error rate has been successfully compared with the ones obtained by a high level model of the chip and to those obtained with other techniques reported in the literature.
展开▼
