Matching of characters in scene images by using local shape feature vectors

The paper describes a new method for locating and recognizing colored patterns of characters in complex scene images where translation, rotation, scale and contrast are unknown. A model of local shape feature vectors is presented. It consists of three vectors and represents some identifiable features in a pattern of characters. Based on this model, potential search points are first found from an unknown target image with this model matched to its edge image. Then, a template matching technique is employed on these candidate points, and the results are classified by a simple nearest neighborhood method and a best match is finally picked out in each cluster. Thus, multiple instances of a pattern of characters are matched and recognized. Experimental results demonstrate the effectiveness of this method.