RGB 최대 주파수 인덱싱과 BW 클러스터링을 이용한 콘텐츠 기반 영상검색

Abstract

This thesis presents content based image retrieval (CBIR) using RGB maximum frequency indexing and BW clustering in order to improve the precision error of image retrieval algorithm that is using histograms. In first stage, RGB color image is segmented into R, G, and B components and their respective histograms are extracted. Each of these histograms are divided into 32 bins. The maximum pixel occurrence values are recorded in each bin's histogram, subsequently which are analyzed and compared. Hence, image retrieval using the maximum pixel value occurrences could be performed, and top 100 output results gathered. In second stage, the extracted 100 images are stored in a database, and are filtered once again using the number and the distributional rate of the clusters. After converting the query image from RGB to grayscale, this histogram is equalized in order to level the biased distributional rate. This equalized histogram is divided into 8 bins, and converted to binary. Then, total number of clusters and percent area of clusters with respect to image is computed. This result is compared with the similar features of the Top 100 matches drawn through the first step. Therefore, the number of bins is changed by applying the existing grayscale histogram algorithm in this research. Similar bin divisions are also performed over color histogram, that involves each of R, G and B components. Finally, The proposed algorithm is analyzed through performance tests and compared against existing methods. The tests verify that proposed algorithm demonstrates better results, and improves precision in the rate of search and priority. Hence, the proposed algorithm is better than the existing grayscale methods.