The problem of “brittleness” of convolutional neural networks in recognition of digit patterns with different writing styles

Abstract

This paper describes the research on the problem of maintaining high accuracy of convolutional neural networks in recognition of digit patterns, writing style of which is different from style of training patterns. To overcome this problem we propose an approach based on the creation of experts that specialize in recognition of patterns of one of the basic types and integration of their knowledge in a committee. The used architecture of experts, methods of their training, techniques of pattern preprocessing and committee voting allow reaching higher average recognition accuracy on the used databases with the help of our approach rather than with similar ones.