Fuzzy Classifier with Convolution for Classification of Handwritten Digits

Abstract

Traditional fuzzy classifier is an important part of artificial intelligence. It achieves classification based on membership function and fuzzy rules which can deal with the uncertainty of data and has semantics. However, the definition of fuzzy rules requires prior knowledge. And fuzzy rules is too sample to achieve high accuracy of classification for classification of handwritten digits. The classifier proposed in this paper combines convolution with fuzzy classifier to classify handwritten digits. The classifier can be divided into two parts: convolution feature extraction part and Gauss membership calculation part. Using back propagation algorithm, the classifier parameters are trained by a large number of labeled data. It can independently extract useful features of handwritten digits to build handwriting feature prototypes, and establish membership functions according to feature prototypes. Experiments on MNIST datasets show that, compared with traditional fuzzy classifiers, the proposed fuzzy classifier can greatly improve the accuracy with less raised time complexity. For MNIST datasets, the proposed fuzzy classifier with convolution can reach higher classification accuracy.