Reduced-reference image quality assessment through SIFT intensity ratio
- PDF / 1,040,909 Bytes
- 9 Pages / 595.276 x 790.866 pts Page_size
- 87 Downloads / 174 Views
ORIGINAL ARTICLE
Reduced-reference image quality assessment through SIFT intensity ratio Tongfeng Sun • Shifei Ding • Wei Chen
Received: 27 June 2013 / Accepted: 25 January 2014 Ó Springer-Verlag Berlin Heidelberg 2014
Abstract Scale invariant feature transform (SIFT) points are scale-space extreme points, representing local minutiae features in the Gaussian scale space. SIFT intensity ratio (SIR), as a novel reduced-reference metric, is feasible to assess various common distortions without the prior knowledge of distortion types. It describes relative changes in the number of SIFT points between a test image and its corresponding reference image. SIFT points in the metric are detected in the first octave of the difference-of-Gaussian scale space under certain preprocessings: neighborhood enhancement through a Laplacian operator to sharpen isolated points and thin edges, reducing false SIFT points; double-size image magnification through linear interpolation to amplify distortion effects, improving its sensitivity to image distortions. Experimental results demonstrate that SIR is superior to existing classic reduced-reference metrics, and can be used to assess different distortions. Keywords Scale invariant feature transform Gaussian scale space SIFT intensity ratio Neighborhood enhancement
1 Introduction Image quality assessment is widely used in the fields of image preprocessing, fusion, transmission, etc. Subjective metrics are simple and accurate. But they cannot work without human’s participations. Therefore, it is indispensable to attach more importance to objective metrics. T. Sun (&) S. Ding W. Chen School of Computer Science and Technology, China University of Mining and Technology, Xuzhou Jiangsu, China e-mail: [email protected]
Objective metrics are classified into three categories: fullreference (FR) metrics, reduced-reference (RR) metrics and no-reference (NR) metrics. FR metrics are based on the following assumptions: the reference image is an image with the highest quality and is available to access. Image quality is assessed by calculating the similarity between a test image and its reference image [1–3]. NR metrics predict image quality based on a test image itself without any access to the reference image [4–6]. But it is difficult for NR metrics to assess image quality degraded by different distortions. RR metrics provide compromising solutions that lie between FR and NR metrics. They only need partial reference information instead of the full reference image [7–9, 17–26]. For RR metrics, it is vital to choose the types of features which should be useful to depict image quality. And the features should occupy small memory space or transmission bandwidth since the chosen features need to be embedded or transmitted to a receiver side for quality analysis. RR metrics are classified into two groups: distortion-specific RR metrics and undistortion-specific (general) RR metrics. Distortion-specific RR metrics are available for specific distortions [10–16], such as noise, blurring, blocking a
Data Loading...
