Graph-theoretical Clustering Based Text Extraction from Color Images
HU Qiao-fu · Microelectronics & Computer · 2003
In this paper,an approach for text extraction from color images is proposed. First,with a statistical color model,the original color space for images is greatly reduced. Second,an unsupervised graph-theoretical clustering is carried out on the reduced color histogram,which decomposes original image into multiple binary images,each of which corresponding to one cluster resulting from clustering. Then,connected component analysis is applied on each of these binary images,and candidatetext regions are located. And text identification is carried out to discard those non-text regions. Finally,text regions on each binary image are integrated to form the detection result on the original color image. With further processing,these located text regions can be fed to existing optical character recognition (OCR) systems for specific applications. Experimental results show the efficiency of the proposed approach.