Frequency-tuned Salient Region Detection
Radhakrishna Achanta
†
, Sheila Hemami
‡
, Francisco Estrada
†
, and Sabine S
¨
usstrunk
†
†
School of Computer and Communication Sciences (IC)
Ecole Polytechnique F
´
ed
´
erale de Lausanne (EPFL), CH-1015, Switzerland.
[radhakrishna.achanta,francisco.estrada,sabine.susstrunk]@epfl.ch
‡
School of Electrical and Computer Engineering
Cornell University, Ithaca, NY 14853, U.S.A.
Abstract
Detection of visually salient image regions is useful for
applications like object segmentation, adaptive compres-
sion, and object recognition. In this paper, we introduce
a method for salient region detection that outputs full reso-
lution saliency maps with well-defined boundaries of salient
objects. These boundaries are preserved by retaining sub-
stantially more frequency content from the original image
than other existing techniques. Our method exploits fea-
tures of color and luminance, is simple to implement, and is
computationally efficient. We compare our algorithm to five
state-of-the-art salient region detection methods with a fre-
quency domain analysis, ground truth, and a salient object
segmentation application. Our method outperforms the five
algorithms both on the ground-truth evaluation and on the
segmentation task by achieving both higher precision and
better recall.
1. Introduction
Visual saliency is the perceptual quality that makes an
object, person, or pixel stand out relative to its neighbors
and thus capture our attention. Visual attention results both
from fast, pre-attentive, bottom-up visual saliency of the
retinal input, as well as from slower, top-down memory and
volition based processing that is task-dependent [24].
The focus of this paper is the automatic detection of
visually salient regions in images, which is useful in ap-
plications such as adaptive content delivery [22], adap-
tive region-of-interest based image compression [4], image
segmentation [18, 9], object recognition [26], and content
aware image resizing [2]. Our algorithm finds low-level,
pre-attentive, bottom-up saliency. It is inspired by the bio-
logical concept of center-surround contrast, but is not based
on any biological model.
Figure 1. Original images and their saliency maps using our algo-
rithm.
Current methods of saliency detection generate regions
that have low resolution, poorly defined borders, or are ex-
pensive to compute. Additionally, some methods produce
higher saliency values at object edges instead of generat-
ing maps that uniformly cover the whole object, which re-
sults from failing to exploit all the spatial frequency content
of the original image. We analyze the spatial frequencies
in the original image that are retained by five state-of-the-
art techniques, and visually illustrate that these techniques
primarily operate using extremely low-frequency content
in the image. We introduce a frequency-tuned approach
to estimate center-surround contrast using color and lumi-
nance features that offers three advantages over existing
methods: uniformly highlighted salient regions with well-
defined boundaries, full resolution, and computational effi-
ciency. The saliency map generated can be more effectively
used in many applications, and here we present results for
object segmentation. We provide an objective comparison
of the accuracy of the saliency maps against five state-of-
the-art methods using a ground truth of a 1000 images. Our
method outperforms all of these methods in terms of preci-
sion and recall.
