질의 영상 특성에 기반한 적응적 웨이티드 CBIR 알고리즘 개발

Abstract

As smartphones became popular recently, image service users increased, and service quality, including storing, sharing, and syncing data, became high quality and high capacity. Various attempts are needed to process and service large amounts of image data efficiently, as such, production and consumption of large amounts of the image data are rapidly increasing. Because of the development of computer technology and multimedia information, it is easy to acquire and store various forms of image information as well as textual information. Additionally, imaging information has skyrocketed in many areas, while the management is becoming increasingly difficult. When searching for a specific data, the image information is larger than the character information, and it is not easy to find fast and efficient retrievals. In this regard, a new retrieval method is needed to effectively manage the images. Therefore, in this paper, I propose a useful adaptive weighted CBIR algorithm based features of query image. Proposed algorithm is composed of four steps. First, it is a preprocessing to remove unnecessary background and trim object image information using Otsu method. Also erosion and dilation are used for removing unwanted information from the images. Second, it labels the images and rearranges them in order of size, normalizing the area, performing size determination process of size filtering. And it is passing through the object discrimination process where the background is classified into the main image and the image where the object is discriminated. If the background is the main image, weights are assigned to use only color information. And if the object is a discriminated image, feature vectors are constructed by extracting feature information of area, compactness, number of corner points, and color value for each labeling. Third, based on the extracted feature vector, area, number of corner points, and compactness of the object are analyzed and the weighted value is adaptively applied according to each feature vector value. And the color pixel values ​​are adaptively weighted using the distribution map of the color map of the index data matrix. Finally, similarity measurement is performed according to the weighted values applied adaptively according to the query image, and the image retrieval was performed. Since adaptive weighted CBIR algorithm is possible to retrieve according to various query images, proposed algorithm complements the selection limits of the query image that existing algorithms have. In order to evaluate the performance of the proposed algorithm in this paper, I used Recall and Precision, which are widely used in CBIR. The precision is 0.08 lower than the color histogram statistical method, but the recall is 0.16 higher. Also the precision is 0.05 and the recall is 0.03 higher than the shape template matching method. More over the precision is 0.02 and the recall is 0.07 higher than HSV color and uniform local binary pattern method. The results confirmed that the proposed algorithm has better retrieval performance. Based on the above, I assume that the proposed algorithm will be able to extract the various features of multimedia data and utilize the optimal database as systems capable of constructing and retrieving.