Re-ranking for web image search results
Date of Issue2015
School of Computer Engineering
The Internet has become a place where massive amounts of information are stored. Growing stream of digital data in the form of images are sent through and uploaded to the Internet every second with the increasing popularity of digital cameras. Web images can be retrieved with the query of text words from sources such as Flickr.com. The topic of text-based image retrieval (TBIR) has become trending in the field of machine learning. Methods to improve the performance of TBIR are tested out by researchers. This project is to carry out the task of re-ranking the relevant image results given the textual query. The project aims to improve the web image search results with the newly proposed bag-based image re-ranking framework. The scope of the project covers the preprocessing the images data using the surrounding contextual and visual features, initial ranking of images based on text query given, weak bag annotation process, implementation of multiple SVM learning algorithms and the experiments with different combinations of settings. This project implements Single Instance Learning SVM, Multi-Instance Learning SVM including mi-SVM, MI-SVM and sparse Multi-Instance Learning SVM. The experiments prove the improvement in retrieval performance for all re-ranking algorithms. Specifically, MI-SVM and sparse Multi-Instance Learning SVM demonstrate outstanding performance enhancements.
Final Year Project (FYP)
Nanyang Technological University