Home // International Journal On Advances in Software, volume 13, numbers 3 and 4, 2020 // View article

Finding Better Matches: Improving Image Retrieval with EFM-HOG

Authors:
Sugata Banerji
Ryan R. Zunker
Atreyee Sinha

Keywords: Histogram of Oriented Gradients; Enhanced Fisher Model; Content-Based Image Retrieval; Shape Matching; EFM-HOG

Abstract:
Retrieving images from a dataset, which are similar to a query image, is an important high-level vision problem. Different tasks define similarity based on various low-level features, such as shape, color, or texture. In this article, we focus on the problem of image retrieval of similarly shaped objects, with the query being an object selected from a test image at run-time. Towards that end, we propose a novel shape representation and associated similarity measure, which exploits the dimensionality reduction and feature extraction methods of Principal Component Analysis (PCA) and Enhanced Fisher Model (EFM). We demonstrate the effectiveness of this representation on three shape-matching problems using multiple large-scale image datasets and also compare its retrieval performance with the Histograms of Oriented Gradients (HOG). Furthermore, to test the performance of our presented descriptor on the non-trivial task of image-based geo-localization, we create a large-scale image dataset and conduct extensive experiments on it. Finally, we establish that our proposed EFM-HOG not only works well on this new dataset, but also significantly improves upon the conventional HOG results.

Pages: 116 to 128

Copyright: Copyright (c) to authors, 2020. Used with permission.

Publication date: December 30, 2020

Published in: journal

ISSN: 1942-2628