Score Metrics For Reference-Based And Blind Blur Estimation
Loading...
Date
2014-07
Authors
Bong, David Boon Liang
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Blur artifact is a type of common distortion for digital images. The presence of
blur artifact in image is an annoyance to image viewers, and affects the perceived
quality of the image. Telecommunication service providers and imaging product
manufacturers are interested in knowing this perceived quality for the purpose of
process and product improvement. However, quality feedback from the image viewers
is tedious, expensive and has to be done in compliance with the standards for subjective
evaluation such as the ITU-R BT. 500 standard. On the other hand, automatic
assessment does not need to comply with any laboratory standard because the
assessment is performed by computation method without using human observers. The
automatic assessment is an objective estimation to predict the degree of blurriness of an
image. Objective estimation can be categorized to the reference-based and blind
methods. Reference-based estimation uses partial or full information of the original
image, and compares it to the blur image. Blur score is derived from the difference
between the original and test images. On the other hand, blind estimation predicts the
degree of blurriness from the test image alone. Classical methods for blur estimation
such as mean squared errors do not produce high correlation, especially for blur
distortion. Hence, there is a need to have new assessment method to measure blurriness
effectively. In this research, a method for reference-based estimation is proposed by
using the contraction of the local contrast masses. Local contrast is determined from the
luminance histogram, and the differences of the local contrast masses are used for the
derivation of blur scores. The proposed method is further improved as a blind method
by using mathematical knowledge of double Gaussian convolution. The minimum of
the Gaussian convolution is used to create valid Gaussian reblur image. From this
reblur process, new blind blur score is created together with the contraction value of
local contrast masses. Both of the proposed methods are performed in the spatial
domain without data transformation, conversion or filtering. In addition, prior training
is not required at all for both methods. Simulations were carried out to check the
performances of both methods. Simulation results show high correlation of 0.9476 ~
0.9888 between the proposed blur scores with respect to image blurriness.
Comprehensive applications and correlation analysis were performed by using three
image databases. Application results show Pearson correlation of 0.8429 ~ 0.9652 for
the new blur scores. For most applications, the proposed methods are also shown to be
statistically more significant than other methods from the same category.
Description
Keywords
Blur artifact is a type of , common distortion for digital images