Concept-based Object Segmentation for Video Retrieval Using Text-Vision Inference

• Main Authors: Shi-xin Mai, 麥世昕
• Other Authors: Chung-Hsien Wu
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/68251815056181382758

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 97 === In the issue of video retrieval, text terms and image feature of videos are important information for video content understanding. In recent years, researchers understand video content by searching high-level video concepts for similarity evaluation between videos. Nevertheless, research related to video concept detection presently, has focused mainly on concept detection by concept detector directly, which neglects deletion of image background noise, result in inaccuracy of video concept detection. Therefore, the study exploits the news terms and news images to infer the concepts in the news image, and search concept-based objects by image segmentation. We hope to reduce the effect of image background noise, increase the accuracy of video retrieval system. The major research purpose of the thesis has four key points. 1.) Use news terms and news images for concept-based object segmentation which include that provide clue for image decomposition by text extension, search concept-based object by image decomposition, increase clue for image decomposition by concept extension. 2.) In text extension, provide more correlation clue for image decomposition by neighboring terms collection. 3.) In text extension, transform the clues into concepts by ontology mapping. 4.) In image decomposition, search the best segmentation result in image segmentation by Genetic Algorithm. In the experiment, we carry out the evaluation of image background noise, concept detector accuracy, image segmentation result, text extension and video retrieval system. According to the experimental results, we reduce the effect of image background noise for concept detection, and increase the accuracy of video retrieval system.