A pyramid matching framework for near duplicate image retrieval

There is a growing interest in Content-based Image Retrieval (CBIR) due to the range of its potential uses, as well as the inherent limitations of the traditional metadata- based system. The most common technique for comparing two images in CBIR is to introduce a distance measure, which compares the...

Full description

Saved in:
Bibliographic Details
Main Author: Niu, Yiming.
Other Authors: Xu Dong
Format: Final Year Project
Language:English
Published: 2011
Subjects:
Online Access:http://hdl.handle.net/10356/43858
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:There is a growing interest in Content-based Image Retrieval (CBIR) due to the range of its potential uses, as well as the inherent limitations of the traditional metadata- based system. The most common technique for comparing two images in CBIR is to introduce a distance measure, which compares their similarities in various dimensions such as color, texture, shape, and others. A positive value implies some degree of dissimilarities between these images. The task of CBIR can therefore be classified as a ranking problem. Many CBIR systems have been developed so far. However, it remains a challenging task for most of them to cope with the spatial shifts and scale variations among testing images. A recently proposed framework, termed Spatially Aligned Pyramid Matching (SAPM) by Xu Dong et al., reveals its unmatched robustness against these problems in empirical observations. The author then launched a project to explore the nature of this method and investigate the appropriate configurations, by applying which the optimal tradeoff between the performance and speed of SAPM could be achieved. More than conducting extensive experiments on Xu’s original work, the author came up with some revised algorithms that improved both of its retrieval accuracy and efficiency to a certain extent. It was noticed that this system can report up to 90% correct matching results on two prescribed image databases, within a quite acceptable wait time. To allow reviewers to perceptually understand the advantages of SAPM, a Matlab GUI-based demo system was implemented. Users are free to switch among the kernel frameworks and make necessary settings for each.