Re-ranking for web image search

• 主要作者: Ngoh, Him Lim.
• 其他作者: Xu Dong
• 格式: Final Year Project
• 語言: English
• 出版: 2012
• 主題: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• 在線閱讀: http://hdl.handle.net/10356/49096
• 機構: Nanyang Technological University

The utilization of Internet as a tool for social interaction has been the latest trend in the recent years. Social networking websites have created a new way to interact and stay in contact with people. Users can now easily share information online by adding texts or uploading pictures to these sites. The ease in sharing has led to the increase of images on the Internet. As such, the retrieval of relevant images from a large collection is now an important topic. One of such retrieval systems is the well-known Text-Based Image Retrieval. Its retrieval performance is however dependent on the textual features provided which gives poor performance when the textual features of the images are sparse and noisy. The aim of this project is to develop a new image re-ranking framework for large scale TBIR. The development of this framework can be divided into 3 phases, Initial ranking, Weak bag annotation and mi-SVM. Based on the given textual query in conventional TBIR, relevant images are to be re-ranked using visual features after the initial text-based search. The re-ranking framework incorporates multi-instance (MI) learning methods such as mi-SVM. It involves the clustering of relevant images using both textual and visual features, treating each cluster as a “bag” and the images in the bag as “instances”. Experiments are carried out on the challenging real-world data set NUS-WIDE to illustrate that the image re-ranking framework can provide better retrieval performance when compared to the conventional text-based search.