(IJCST-V5I2P15) :aggie Varghese, Smitha M. Jasmine

International Journal of Computer Science Trends and Technology (IJCST) Volume 5 Issue 2, Mar Apr 2017
RESEARCH ARTICLE OPEN ACCESS
A Survey Paper on News Web Video Event Mining

Aggie Varghese [1], Smitha M. Jasmine [2]
Department of Computer Science and Engineering
India
ABSTRACT
Due to the explosive growth of web videos, it becomes great challenge of how to efficiently browse lakhs or even millions of
videos at a glance. Given a user query, social media web sites usually return a large number of videos that are diverse and not
relevant. Exploring such results will be time-consuming and thus degrades user experience. Based on these observations, an
event mining solution should provide the users a quick overview for each topic with a reduced browsing time. There are
different methods for news web video event mining and are described in this paper.
Keywords :- Cooccurrence, multiple correspondence analysis (MCA), near-duplicate keyframes (NDK), news web video event
mining, visual feature trajectory.
is a challenging mission to mine the events under the noisy

I. INTRODUCTION and diverse social web scenario.
News web videos are mainly composed of visual and
textual information. Visual information contains semantic gap J. Yao, B. Cui, Y. Huang, and Y. Zhou, in [2]
and user subjectivity problems, and therefore, using either Bursty event detection from collaborative tags, proposed a
visual or textual information alone for news web video event technique based on visual Near Duplicate Keyframe -level
mining may lead to unsatisfactory results. In order to clustering. In this paper, they proposed a new method to detect
overcome these shortcomings, both visual and textual features bursty tagging event, which captures the relations among a
are utilized for web news video event mining. For visual group of correlated tags. These tags are either bursty or
information, some important shots are frequently inserted into associated with bursty tag co-occurrence. This kind of bursty
videos as a support of viewpoints, which carry useful tagging event generally corresponds to a real life event. The
information. Since there is unique role of near-duplicate events are profiled with more comprehensible and
keyframes (NDK) in the news search, topic detection and representative clues. The proposed method is divided into
tracking (TDT) and copyright infringement detection, these three stages. As the first step,they exploit the sliding time
duplicate keyframes/shots are clustered to form different intervals to extract bursty features, and graph clustering
groups according to visual content. Such groups are similar to techniques is adopted to group bursty features into meaningful
the hot terms in the text field. Here, each cluster is called an bursty events. It helps to detect the relationship among tags
NDK group, which can be used to group videos with similar with bursty tag cooccurrence. In this method, the number of
content to the same events. NDKs among videos is limited. Each event is composed of
different kinds of visual scenes, while CC_V just groups one
II. RELATED WORKS scene.
The various methods used for news web video event
mining are described here. X. Wu, Y.-J. Lu, Q. Peng, and C.-W. Ngo, in [3]
Mining event structures from web videos, proposed T+V
Q. He, K. Chang, and E.-P. Lim, in [1] Analyzing which is a fusion method. This paper explores the issues of
feature trajectories for event detection, proposed a method mining event structures from Web video search results using
based on the feature trajectory using text words . In this burst detection, text analysis, clustering, and other techniques
paper, they 1) spectral analysis is applied first to categorize It first applies feature trajectory to the visual field. Then it
features for di erent event characteristics: periodic and tries to mine events on the basis of text cooccurrence and
aperiodic, important or less-reported. 2) modeled periodic feature trajectory. In this method the visual near-duplicate
features with Gaussian mixture densities and aperiodic feature trajectories of NDKs are not consistent. It misses those
features with Gaussian density and, and detected each low-frequency terms and NDKs which are common for a large
features burst by the truncated Gaussian approach; 3) number of videos.
proposed an unsupervised greedy event detection algorithm to C. Zhang, X. Wu, M.-L. Shyu, and Q. Peng, in [4]
detect both aperiodic and periodic events. The highly sets to A novel web video event mining framework with the
the orrelated word features are grouped to events by mapping integration of correlation and co-occurrence information,
the word video sets. The performance of FT_T is poor. This c proposed a method inorder to mine the correlation between
ISSN: 2347-8578 www.ijcstjournal.org Page 78

events and NDKs by using the distribution characteristics of Mining event Better than Better than
the terms. In this paper, in order to improve the performance structures from web FT_T Better than FT_T and
of web video event mining they propose a novel four-stage videos [T+V] FT_T and CC_V
framework. The first stage is data preprocessing. Multiple CC_V
Correspondence Analysis (MCA) is then applied to explore A novel web video Can be low
the correlation between terms and classes. It targets to bridge event mining low or high
the gap between NDKs and high-level semantic concepts. framework with the Sometimes compared to
Next, the similarity between NDKs and classes are detected integration of
high FT_T,CC_V
using co-occurrence information . Finally, through negative correlation and co-
occurrence and MCA
NDK pruning and positive NDK enhancement both of them
information [MCA]
are integrated for web video event mining. It can bridge the
Integration of visual Better than Most of the high
gap between NDKs and terms. In this method, multiple temporal FT_T, T+V time high
languages, synonyms, and the number of videos in each NDK information and and MCA
are problems. textual distribution
Chengde Zhang, Xiao Wu, Mei-Ling Shyu, and information for
Qiang Peng in [5] Integration of visual temporal information news web video
and textual distribution information for news web video event event mining
mining, proposed a method which uses both the neighbor
stabilization process and MCA similarity measure to
generate the textual distribution information. It can better
explore the degree of correlation between different terms and IV. CONCLUSION AND FUTURE WORKS
events. The visual near-duplicate feature trajectory, i.e., the
time distribution information of an NDK,is integrated with the
NDK-within-video information (cooccurrence) as the visual This paper compares different methods used for news web
temporal information to cluster more NDKs belonging to the video event mining. The method used in integration of visual
same event. temporal information and textual distribution information for
news web video event mining can accurately identify more
relevant news web videos. Since CC_V method focuses on
III. COMPARISON BETWEEN RELATED visual content, it might miss many news web videos without
WORKS he duplicate visual content. While applying the method of
FT_T to the visual NDK level, different scenes of the news
To evaluate the performance of the web event mining, we use web videos can be grouped together which indicates that the
the Precision (P), Recall (R), and F1 measure (F1), which are visual near duplicate feature trajectory can deal with news
dened in (1)(3), respectively web video event mining with diverse visual scenes. However,
the NDK detection error problem may lead to inaccurate
Precision = |B i + |/|Ai| (1) visual near-duplicate feature trajectories.Therefore, the
instability characteristics would mine unrelated news web
Recall = |B i + |/ |Bi| (2) videos pertaining to other events. MCA focuses on the textual
information. For some news web videos, in order to attract
F1=(2PrecisionRecall)/(Precision+Recall) (3) more attention, the users might add some words in the title/tag.
Hence, some unrelated keyframes are falsely considered as
where B i + is the number of correctly grouped positive videos relevant information. The visual information are used for
for cluster Ai, and Bi is the number of positive samples in the grouping the relevant videos into events. The texts/terms are
ground truth. Since F1 considers both Precision and Recall relatively general, broad, and noisy. Hence, the integration of
values, it is mainly used to evaluate the performance. both visual and textual information can lead to good results.
Finally, clustering with the textual and low frequent
information could yield bring more relevant videos. The text
Methods Precision Recall F1 information extracted from the keyframes of the videos can
Analyzing feature low also be taken as the textual information for the event mining
trajectories for event low process.
detection [FT_T] low .
Bursty event low

detection from high low
collaborative tags
[CC_V]

ACKNOWLEDGMENT
We thank members of Computer Science and Engineering
department of College of Engineering and Management,
Punnapra, for their valuable support and feedbacks.
REFERENCES
[1] Q. He, K. Chang, and E.-P. Lim, Analyzing feature
trajectories for event detection, in Proc. 30th ACM Int.
Conf. Res. Develop. Inform. Retrieval, 2007, pp. 207
214.
[2] J. Yao, B. Cui, Y. Huang, and Y. Zhou, Bursty event
detection from collaborative tags, World Wide Web,
vol. 15, no. 2, pp. 171195, 2012.
[3] X. Wu, Y.-J. Lu, Q. Peng, and C.-W. Ngo, Mining
event structures from web videos, IEEE Multimedia,
vol. 18, no. 1, pp. 3851, Jan. 2011.
[4] C. Zhang, X. Wu, M.-L. Shyu, and Q. Peng, A novel
web video event mining framework with the integration
of correlation and co-occurrence information, J.
Comput. Sci. Technol., vol. 28, no. 5, pp. 788796, 2013.
[5] Chengde Zhang, Xiao Wu, Mei-Ling Shyu, and Qiang
Peng in [5] Integration of visual temporal information
and textual distribution information for news web video
event mining
[6] Statistics. (2017). [Online]. Available:
http://www.youtube.com/yt/ press/statistics.html

(IJCST-V5I2P15) :aggie Varghese, Smitha M. Jasmine

Uploaded by

Copyright:

Available Formats

(IJCST-V5I2P15) :aggie Varghese, Smitha M. Jasmine

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(IJCST-V5I2P15) :aggie Varghese, Smitha M. Jasmine

Uploaded by

Copyright:

Available Formats

International Journal of Computer Science Trends and Technology (IJCST) Volume 5 Issue 2, Mar Apr 2017

RESEARCH ARTICLE OPEN ACCESS

A Survey Paper on News Web Video Event Mining

is a challenging mission to mine the events under the noisy

ISSN: 2347-8578 www.ijcstjournal.org Page 78

Bursty event low

ISSN: 2347-8578 www.ijcstjournal.org Page 79

ISSN: 2347-8578 www.ijcstjournal.org Page 80

You might also like