형태 정보 검색을 위한 재배열 체인 코드 기반 영상 검색 알고리즘

Abstract

In our current Internet-based society, as networks have been developing, the amount of audio-video information is increasing. Such multimedia information is represented in diverse forms such as photos, video, graphics, three-dimensional models, sounds and voices. As well, since digital cameras are widely available, it is easily possible for anyone to make and distribute digital photos. Therefore, the public can be suppliers as well as consumers of images. Also, with the development of advanced personal computers, the universal use of large-scaled data storage system, and computer networks of the world-wide web, it is easier to produce, transmit, and manufacture digital videos. Early methods of retrieving video data using text-based retrieval systems with query languages have some problems. Since text-based retrieval systems require key words that can express each image clearly, and key words may be decided subjectively, there may be cases when intrinsic key words will not be effectively decided. Therefore, instead of simple comments or a title-based retrieval system, we need a system that can express and store data more effectively. In order to have the advantage of ease of use by people, machines should recognize multi-media data in a similar way to people. Content Based Image Retrieval(CBIR) which expresses and stores image data itself to use for retrieval is one of the fields which have been intensively studied in multi-media databases). To develop effective CBIR, what images mean should be appropriately expressed. For effective expression of image data, the system extracts and stores several features per image automatically, and then the images required by users are retrieved with the uses of such features. The main features used in CBIR include colors, textures, and shapes. To manage and retrieve images from a large-scale database based on such multi-media content elements, features should have a high discriminative ability and be sensitive to the expansion, reduction, movement, and rotation of the objects. To reduce processing speed, the retrieval algorithm should be simple. Information on features are widely used to meet such requests. Information on features is more useful in that only outlines are used and the visual structures are easier to notice. Also, to obtain information on features: outlines, corner points, moments, motion, and segmentary features are used. In particular, in using an outlined feature, a chain code system is widely used. Chain codes have 4 or 8 directions and are based on information on the directions of the edges. However, as their directionality causes lots of problems in feature matching, there have been extensive efforts to improve them. In other words, the retrieval algorithm should be simple while chain codes are good at noticing the expansion, reduction, movement and rotation of objects. Therefore, this study proposed a new algorithm that can strengthen the noise invariant in the previous chain codes that are used for feature recognition. First, chain codes are extracted from images, the differences between values of the chain codes are obtained, and then the chain codes are rearranged. Second, the chain codes are extracted from images and are rearranged using their typical values. Then the correlogram is applied. Third, chain codes are extracted from images and rearranged in descending order of frequency. Then the frequency is extracted and normalized to make a feature vector. Fourth, chain codes are extracted from images and then a maximum histogram to the extracted chain codes is obtained. The second chain codes are obtained through subtraction from the original chain code value and the chain codes are rearranged. This study tested and analysed such rearrangement of the chain code algorithms using images of varying size. As a result, it was demonstrated that they overcame the disadvantages of normal chain code algorithms which are sensitive to rotation, expansion, and reduction of images and it is expected that they can be widely used for a feature retrieval system.