A Novel Method for Character Segmentation in Natural Scenes

Introduction Nowadays the applications of scene text recog-nition are rapidly expanding with the developmentof portable digital imaging devices1]. However,character extraction from scene images has alwaysbeen a challenging problem due to complex back-ground, uneven illumination, shadows and noise ofimages2]. Besides, languages also impose anotherlevel of variation in text. Characters in such lan-guages as Chinese, Japanese and Korean are usual-ly composed of several strokes, which do not n…

A Novel Method for Character Segmentation in Natural Scenes

A new approach to extract and segment characters in natural scenes was proposed in this paper. First, a set of intrinsic features were calculated based on connected components (CCs) extracted by a non-linear Nilblack algorithm. Then, feature propagation was conducted for feature enhancement, under the constraint of the layout relations. Next, candidate CCs were fed into classifiers with the enhanced feature vector. At last, a model-based hierarchical merging (MHM) procedure was presented to obtain understandable characters. The proposed merging algorithm utilized the constraint of text lines for specific languages and dynamically merges CCs into characters.The whole algorithm was evaluated at both pixel level and character level, experimental results showed that the proposed method is effective in detecting scene characters with significant geometric variations, uneven illumination, extremely low contrast and cluttered background.