Towards searching for historical social images

• Main Author: Chew, Min Min
• Other Authors: Sourav Saha Bhowmick
• Format: Final Year Project
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Computer science and engineering
• Online Access: http://hdl.handle.net/10356/59184
• Institution: Nanyang Technological University

The increase in ease of uploading images to social image sharing platforms and the convenience of digital photography devices has given rise to a great number of images being available online. With a large number of images online, there has to be efficient, effective and interesting ways to search and retrieve these images to suit the user needs. One such way is through Tag-Based Social Image Retrieval (TagIR). Social image sharing platforms allow users to annotate their uploaded images with tags. These tags allow the images to be organised and TagIR will utilise the tags to search and retrieve images relevant to the search queries. In this project, a program was implemented to search and retrieve images based on their historical relatedness. The images returned would allow the user to identify the images with historical objects. This project was built upon an existing system, the Concept-Aware Social Image Search (CASIS) system. The historical relatedness of an image was determined using the historical relatedness score of its top ten relevant tags. The program used the Wikipedia Miner Services to derive the historical relatedness of a tag by comparing the tag word with the term “history”. Wikipedia Miner is a program that calculates the sematic relations between two different terms based on the Wikipedia Hyperlink structure. For this project, the set of 113,726 images and their corresponding tags used were from Flickr, a photo-sharing website. Java programming language and MySQL database were used to implement the program. The program was implemented and tested with a fraction of the data. From the results, it showed that the program achieved the project objective. It was able to search and retrieve the images according to their historical relatedness scores. In addition, it could perform the task efficiently, with an average retrieval time of 0.3s. However, it was found that certain limitations of the project affected the accuracy of some of the results. The first was due to the noisy nature of tags and the unrestricted feature of tagging. The second was due to Wikipedia Miner being based on the Wikipedia Hyperlink structure, which meant it depended on whether the term was in Wikipedia. In addition, the results derived were not able to fully conclude the effectiveness of the program as the set of data used did not have enough images with historical relations. Hence, there is space for future improvements to the project.