Character detection based on multi-scale measurement

Abstract

The paper presents a new method for character string extraction that works efficiently even for a complicated geographical map. The concept of multi-scale measurement is introduced to achieve development of an efficient string detection technique. In this paper, scale means the size of the area where character candidates may exist. The proposed method first merges small black regions within a certain area into a mass. When the size of the area changes, the mass will change. The proposed method observes the change of a mass corresponding to the change of the size of the area, and searches for a stable mass as a character string. Multi-scale measurement enables the detection process to find the adequate size of an area to detect a string. Because a stable mass may include small figures, a test of the shape of a detected mass and a character recognition process follow to judge whether a mass forms a character string. If a mass is rejected, it is split into smaller masses according to the results of multi-scale measurement. These judgment and split processes are repeated to detect character strings from a pattern where several strings are written closely.