Character extraction by integrating color into edge-based methods

Text recognition is difficult in e-commerce images, which contain a mixture of digital born text and natural scene text. To extract character regions in an image, a method using the consistency of the stroke widths of a character, called stroke width transform (SWT), is promising because of its simplicity and low computation. The method is to measure stroke widths after detecting edges. On the other hand, character extraction based on color clustering has been investigated separately. This paper presents algorithms for extending the SWT method by integrating color-clustering information. Our experimental results show the effectiveness of the proposed algorithms both with digital-born and natural-scene images.