Text extraction from name card images

Abstract

The purpose of this thesis is to extract text from complex designed name card images. Some conventional methods cannot solve the three problems mentioned above very well. This thesis solves the problems by modeling human visual perception theories. We model four features of texts: color, size, line segment and regularity, which preattentively distinguish text from non-text. Color and size are normal features captured by conventional methods. Some pictures or graphics can be distinguished from text by line segment modeling. By modeling regularity, some character-like logos can be filtered out. Two methods are proposed in this thesis. The first approach is a simple but efficient algorithm which uses some empirical thresholds to filter and combine the contours of characters to text lines. The second approach utilizes neural networks to get the result objectively. Similar features are modeled in both approaches.