Image Partial Blur Detection and Classification
∗
Renting Liu Zhaorong Li Jiaya Jia
Department of Computer Science and Engineering
The Chinese University of Hong Kong
{rtliu,zrli,leojia}@cse.cuhk.edu.hk
Abstract
In this paper, we propose a partially-blurred-image clas-
sification and analysis framework for automatically detect-
ing images containing blurred regions and recognizing the
blur types for those regions without needing to perform blur
kernel estimation and image deblurring. We develop several
blur features modeled by image color, gradient, and spec-
trum information, and use feature parameter training to ro-
bustly classify blurred images. Our blur detection is based
on image patches, making region-wise training and classi-
fication in one image efficient. Extensive experiments show
that our method works satisfactorily on challenging image
data, which establishes a technical foundation for solving
several computer vision problems, such as motion analysis
and image restoration, using the blur information.
1. Introduction
In this paper, we focus on detecting and analyzing par-
tially blurred images and propose a novel method to au-
tomatically detect blurred images, extract possible blurred
regions, and further classify them into two categories, i.e.,
near-isotropic blur and directional motion blur.
Our method attempts to tackle two major problems. One
is blur detection with simultaneous extraction of blurred re-
gions. The result in this step provides useful high-level
regional information, facilitating a variety of region-based
image applications, such as content-based image retrieval,
object-based image compression, video object extraction,
image enhancement, and image segmentation. It can also
serve as one of the criteria of measuring the quality of cap-
tured images.
The second objective of our method is to automatically
classify the detected blur regions into two types: near-
isotropic blur (including out-of-focus blur) and directional
∗
The work described in this paper was fully supported by a grant from
the Research Grants Council of the Hong Kong Special Administrative
Region, China (Project No. 412307).
(a) (b)
Figure 1. Two image examples with (a) motion blurred regions and
(b) out-of-focus blurred regions.
motion blur. We classify image blur into these two classes
because they are most commonly studied in image restora-
tion. The blur classified images also easily find applications
in motion analysis and image restoration. Two partial-blur-
image examples are illustrated in Figure 1.
Although the topics of image blur analysis have attracted
much attention in recent years, most previous work focuses
on solving the deblurring problem. General blur detection,
on the contrary, is seldom explored and is still far from prac-
tical. Rugna et al. [5] introduced a learning method to clas-
sify blurry or non-blurry regions in one input image. This
method is based on an observation that blurry regions are
more invariant to low pass filtering. In our experiments, we
find that only using this information is not sufficient for de-
signing a reliable classifier. Different blur measures should
be combined in order to achieve high-quality blur detection.
This method also does not distinguish blur types.
Utilizing the statistics of gradient information along dif-
ferent directions, the method in [18] builds an energy func-
tion based on the inferred blur kernel in order to segment
image into blur/nonblur layers. This method only discovers
motion blurred regions by inferring directional blur kernels.
Other blur estimation methods, such as the one proposed
in [7], only provide a measure of blur extent, which cannot
be directly used to discriminate blurry against non-blurry
regions.
In this paper, we present a new blur detection and anal-
ysis method for automatically extracting blurry regions by
combining specifically designed blur features represented
978-1-4244-2243-2/08/$25.00 ©2008 IEEE