A NOVEL IMAGE TAG SALIENCY RANKING ALGORITHM BASED ON SPARSE
REPRESENTATION
caixia Wang, Beijing Jiaotong University, 11120480@bjtu.edu.cn
zehai Song, Beijing Jiaotong University, zhsong@bjtu.edu.cn
songhe Feng, Beijing Jiaotong University, shfeng@bjtu.edu.cn
congyan Lang, Beijing Jiaotong University, cylang@bjtu.edu.cn
shuicheng Yan, National University of Singapore, eleyans@nus.edu.cn
ABSTRACT
As the explosive growth of the web image data, image tag
ranking used for image retrieval accurately from mass
images is becoming an active research topic. However, the
existing ranking approaches are not very ideal, which
remains to be improved. This paper proposed a new image
tag saliency ranking algorithm based on sparse
representation. we firstly propagate labels from image-level
to region-level via Multi-instance Learning driven by sparse
representation, which means reconstructing the target
instance from positive bag via the sparse linear combination
of all the instances from training set, instances with nonzero
reconstruction coefficients are considered to be similar to
the target instance; then visual attention model is used for
tag saliency analysis. Comparing with the existing
approaches, the proposed method achieves a better effect
and shows a better performance.
Index Terms — tag saliency ranking, sparse
representation, multi-instance learning, Diverse Density,
visual attention model
1. INTRODUCTION
In recent year, as the rapid development of multimedia
information and internet technology, there appears many
online image sharing sites, such as Flickr, Facebook etc,
which not only allow users to upload their own photos, but
allow users to add tags manually to explain and describe the
content of the image. However, because of the difference of
the network user’s cultural background, knowledge
structure and concerns, different people will add very
different tags on the same or similar images. So image label
listing on the network is often inaccurate and disorder and
far from satisfied with the quality of image retrieving.
Consequently, image tag ranking is becoming an active
research field of multimedia information in recent years. In
this paper, for reordering the original tag list more
effectively, we improve the conventional tag saliency
ranking algorithm mentioned by S.Feng et al. in [1] via
proposing a new tag saliency ranking algorithm based on
sparse representation.
Existing main image tag ranking algorithm is roughly
divided into two categories, including the tag relevance
ranking algorithm and tag saliency ranking algorithm. As to
tag relevance ranking, it mainly reorder labels list based on
the relevancy between label and image. Li et al.
[2]proposed it based on neighbor voting, we firstly find k
nearest neighbor images set based on low-level features;
then calculate the frequency of the tag list of the given
image appear in the k nearest images, finally reorder the
given label list according to the frequency. In addition, Liu
et al. also implements the tag relevance ranking through the
fusion of Kernel Density Estimation (KDE) [3] and random
walk [3, 10] algorithm, firstly, we estimate the initial
correlation score of the each label through the probability
model established for it; then modify the relevance score
through random-walk algorithm, finally reorder the given
label list according to the relevance score. As to tag saliency
ranking, S.Feng et al. proposed it in [1]. Because of the
saliency degree of the tag is highly relies on the saliency
information of the corresponding regions, how to build a
semantic mapping between the regions of a given image
and the associated tags is an important issue. However, in
most annotated image databases and photo sharing websites,
tags are usually associated with image instead of individual
regions, and MIL (multi-instance learning) as a
generalized-supervised learning algorithm can well deal
with the situation that tags spread from image-level to
region-level. So, given an image, firstly, Multi-instance
Learning (MIL) algorithm is used to propagate the tags
from image level to region level; then using visual attention
model to analyze saliency degree of each segment, finally
disorder it according to the saliency degree of the
corresponding area. However, the tag relevance ranking
algorithm is mainly for large-scale data, but not very ideal
for small data. Although the tag saliency ranking algorithm
manages to deal with small data, the efficiency and
accuracy of it is not very ideal. Since the tag saliency
ranking algorithm relies heavily on the instance prototypes
selected through MIL algorithm, we consider improving the
accuracy of this algorithm in choosing instance prototypes.
In this paper, we proposed a new image tag saliency
ranking algorithm based on sparse representation, which
improves in the generation details of instance prototype
compared with the existing tag saliency ranking algorithm.
During the instance prototypes selection, to avoid the
limitation from one-to-one similarity measure in the context