Watching pattern distribution via massive character recognition

摘要

The purpose of this paper is to analyze how image patterns distribute inside their feature space. For this purpose, 832,612 manually ground-truthed handwritten digit patterns are used. Use of character patterns instead of general visual object patterns is very essential for our purpose. First, since there are only 10 classes for digits, it is possible to have an enough number of patterns per class. Second, since the feature space of small binary character images is rather compact, it is easier to observe the precise pattern distribution with a fixed number of patterns. Third, the classes of character patterns can be defined far more clearly than visual objects. Through nearest neighbor analysis on 832, 612 patterns, their distribution in the 32 × 32 binary feature space is observed quantitatively and qualitatively. For example, the visual similarity of nearest neighbors and the existence of outliers, which are surrounded by patterns from different classes, are observed.