A Character Pattern Extraction Method Applicable to Various Types of Images

As the applications of the OCR are widely expanding recently, computers have to extract character patterns from various types of images such as scanned document images, scene images and video frames. Many methods for character pattern extraction have been proposed so far. They were designed and tuned for a special type of image and for a certain input device. However, we have many images which cannot be processed by the conventional framework. It is necessary for us to develop an allpurpose, unified processing which can extract character patterns from various types of images using various input devices. To realize the unified processing, we previously developed a character pattern extraction method based on the local multilevel processing and the region growing. The performance of the method was evaluated using scanned document images, however, no detailed evaluation has been performed for scene images. By the new experiments using scene images, we have verified that the method works well even for scene images. The rate of text line extraction is 75%, which is almost same as that of a conventional method. Although the rate is not very high, the remarkable property of our method is that it is applicable to various types of images.