Concept Detection for Content-Based Image Retrieval

• Main Authors: Yen-Ju Hsu, 許雁茹
• Other Authors: Shyi-Chyi Cheng
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/78844734844143013624

碩士 === 國立臺灣海洋大學 === 資訊工程學系 === 97 === In this thesis, a new semantic learning method to detect semantic region for image retrieval from a given amount of labeling effort is proposed. In our approach, the database images are classified into two classes –the labeled class and the unlabeled class. Form images in the labeled class, we construct a concept detection to detect the important regions in each image based on the statistical information of a semantic class. All the images in the database are segmented into multiple disjoint regions, each of them is represented by three type of low-level visual features ( i.e. color, shape, and texture). With this representation a region weighting model based on the statistical information of low-level visual features is predicted to analyze semantic concepts hidden in the database. One key obstacle in applying statistical methods to discover the hidden semantic concepts for annotating images in the amount of manually-labeled images is normally insufficient. For images that have not been annotated, the learning algorithm estimates their important regions whose low-level features are then extracted to retrieve semantic all similar image s form the test data base. Experimental results show that the performance of the proposed method is excellent as compared with that of simulated traditional content-based image retrieval.