ENTROPY FEATURES TRAINED SUPPORT VECTOR MACHINE BASED LOGO

					         INTERNATIONAL and Multimedia (IJGM), ISSN 0976 – 6448(Print),
  International Journal of Graphics JOURNAL OF GRAPHICS AND
  ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME
                               MULTIMEDIA (IJGM)

ISSN 0976 - 6448 (Print)
ISSN 0976 -6456 (Online)                                                     IJGM
Volume 4, Issue 1, January - April 2013, pp. 20-30
© IAEME: www.iaeme.com/ijgm.asp
Journal Impact Factor (2013): 4.1089 (Calculated by GISI)
                                                                         ©IAEME
www.jifactor.com



      ENTROPY FEATURES TRAINED SUPPORT VECTOR MACHINE
      BASED LOGO DETECTION METHOD FOR REPLAY DETECTION
              AND EXTRACTION FROM SPORTS VIDEOS

                               Vilas Naik1, Raghavendra Havin2
         1 Department of CSE, Basaveshwar Engineering College, Bagalkot, India
         2 Department of CSE, Basaveshwar Engineering College, Bagalkot, India


  ABSTRACT

         In many sports, the majority of highlights are confined to relatively short durations of
  intense action. In some sense these segments capture the essence of a game and summarize
  the moments of important action. Automatic detection of these highlights could provide an
  important browsing mechanism in a video library of sports games. These replays often cor-
  respond to highlights in a game and can be used as indices of a sports video. The proposed
  mechanism employs support vector machine (SVM) for detection of logos that are flashed at
  beginning and end of every reply action. The algorithm is composed of logos detection and
  replay segment extraction. First SVM is trained with features of all possible logos normally
  used in various sports videos. Then the SVM is used for detection of logos that sandwich
  replay segment further that segment is automatically extracted to produce replay clip. The
  SVM classifier is trained with histogram features of logos is utilized. Experiments conducted
  on IPL and Soccer videos demonstrate the effectiveness of this method. Moreover, algorithm
  can be easily applied to other sports videos where replay is sandwiched by pair of logos.

  Keywords: Logo-based detection, Support vector machine (SVM learning (ML), Sports
  replay detection, Video summarization.

 1.       INTRODUCTION

         In a sports video library, one may find thousands of hours of recordings, representing
  many hundreds of individual games. Fortunately, in any given game, not every minute com-
  mands the same interest. A majority of the excitement is contained in a small fraction of the
  coverage of the game. These highlights are the essence of the game and present a succinct


                                                 20
 International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
 ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

 summary of the significant actions and events. For many viewers of sports video these high-
 lights are all they are really interested in seeing. Detection of these highlights is therefore of
 interest in itself, but might also serve as an important indexing mechanism for searching for
 more substantial sections of material from large archives. With the rapid increase of media
 data, it is in urgent need of an efficient and effective method for information management and
 retrieval. Semantics based event detection and indexing has been attracted much attention
 recently for interesting events are highly useful for video browsing, indexing and highlights
 generation. Replay is an important video editing way in broadcasting programs. Commonly,
 an important or interesting segment will be played with a slow-motion pattern in order to let
 audiences explore and enjoy the detail. Replay is a significant indication for highlights and
 often taken as a key factor in event detection. The sports video summarization is widely in-
 vestigated since sports holds a large amounts of audiences worldwide. As an important com-
 piled clue, the replay shows the details of an important video segment with a slower speed.
 Thus it is widely employed in sports video analysis especially for event detection, highlight
 creation and summarization etc.
          Replays are related to the sports, which are shown more than once when any impor-
 tant event in sports takes place. To get more information from replay, generally it is shown
 with slow motion. Hence it is also called slow motion replay (SMR). Replay is which is ne-
 cessary for highlights generation. Highlight is a shorter version of sports video that is also
 called as sports video summarization. Replay is important for all spectators and referees to
 judge and take final decision in most of the sports. When an important event happen like goal
 in football or hockey cricket. When referee is in doubtful condition, or strong appeal is made
 by players, putting all such events in an original temporal order we can have highlights of any
 sports video. Replay contains all-important events. Replays are used for highlight making.
 Replays can be used as a summary of game for searching it in digital video library.
          Replay is an important video editing way in broadcasting programs. Commonly, an
 important or interesting segment will be played with a slow-motion pattern in order to let au-
 diences explore and enjoy the detail. Replay is a significant indication for highlights and of-
 ten taken as a key factor in event detection.
          In recent years, researchers have reported many approaches of replay detection in lite-
 ratures. In this paper, an automatic and effective logo based replay detection method is pro-
 posed. It is well known that in most cases there exists a special transition at both start and end
 of a replay, in which a highlighted logo comes in and out gradually. This transition is called
 as “logo-transition”. Logos often keep the same during the whole program. Our motivation is
 to use the logos to assist replay segments detection. Firstly some logo-transitions are de-
 tected and further extract logo-samples from them at the beginning of a program. After that,
 the logo-template is extracted from these samples. Further, this template is employed to
 detect the other logos. After all logos are obtained, the video can be partitioned into segments
 with taking logos as boundaries. Then, shot and motion features are extracted in each seg-
 ment. Finally, based on these cues, an SVM classifier is used to replay segment identification.

2.      RELATED WORK

         Slow motion replays in broadcast sports video are important and reliable clues for
 highlights and key events, since they are usually chosen by human experts. Therefore, detec-
 tion of replays could facilitate semantic-based high level video processing tasks such as high-
 light generation [1], event detection [2], video summarization [3], etc.

                                                  21
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

         Some slow motion replay detection methods have been proposed in the literature.
They can be broadly classified into three categories. The first category of method directly
analyzes the inherent attributes of replay video segments and tries to use those attributes to
differentiate the slow motion replays from the normal plays. Before high speed cameras are
widely used, slow motion replays were usually produced by repeating frames recorded from
standard cameras. A few detection methods are based on this assumption. A replay detection
method based on the macroblock type, motion vector and bit rate information is presentd in
[4]. The detection is performed in the MPEG compressed domain to find the locations of re-
peating frames. The algorithm in [1] calculates the image differences between adjacent
frames and search fluctuations in the differences to identify replays. These methods usually
get low performance on replays that captured by high speed cameras and they can not provide
accurate boundaries of replays.
         The second category of method employs statistical techniques to perform shot classi-
fication for the purpose of detecting replay shots. The author of [5] presented a method based
on the difference of motions between replays and normal plays. A support vector machine
(SVM) was trained with features of color ratio, shot length, mean color value and motion re-
lated feature to detect replay shots. The proposal in. [6] applied scene transition structure
analysis on the shot classification results. Replays were extracted based on the generated shot
label sequences by using some predefined rules. A HMM to classify replay and non-replay
shots is employed in [7]. The performance of this category is usually not satisfied and results
are still not precise enough.
         The third category of method tries to find the differences between replays and normal
plays in terms of specific production behaviors, such as special video effects or logo patterns.
Replays are then determined according to the existence of such behaviors. The work pre-
sented in [8] captures a digital video effect (DVE) l interactively for given videos. It is de-
scribed by color and motion of the gradually changing boundary between adjacent shots. A
replays are located by two DVEs detected with the model. Based on the observation that
more and more replays that are sandwiched in between two certain logo transition sequences,
some researchers try to use logo transition locations to identify replays. The mean-square dif-
ferences of intensity and color histogram as frame similarity measure to detect logo template
and to search logos in videos is discussed in[9]. A method in [10] employed the spatiotem-
poral mode seeking on selected logo sequences to capture the dominant color mode. This
color mode is then used to perform EMD-based similarity matching for logo searching pur-
pose. Tong et al. [11] used frame-to-frame difference with MSD to select candidate logo se-
quence. They calculated the average image of those candidate logo frames near to the center
to obtain logo template. Template matching was performed based on color and shape fea-
tures. Huang et al. [12] first extracted gradual transitions and employed motion features to
learn the logo template. A color representation was then computed from the logo template
and was used to search logos. Dang et al. [13] semiautomatically extracted logo template se-
quences and detected replays based on template sequence matching. Han et al. [14] learned
gradual transition patterns based on motion vector reliability classification from a database by
SVM. Replay detection was performed by fusing the pattern matching results and slow mo-
tion detection results based on motion vector information. The third category of method can
usually obtain high accuracy on replay boundaries and the performance is better than above
two categories. However, these methods often make assumptions on the logo transition se-
quences such as logo locations, moving patterns and speed, types of digital effects, etc. And



                                                22
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

nearly all of them assume that the two logo transition sequences before and after the replay
are identical.
        The Proposed solution is a methodology of determining the replay detection in a
sports video using support vector machine for classifying frames as Logo frame and non logo
frame. In most of sports video a logo swaps at beginning and end of the replay, the frames in-
between the start logo frame to its corresponding end logo frame of a videos construct a rep-
lay event. Applying same procedure for total video it results into collection of replays of
sports video. These replays can be used to form a summary of video.

3.     PROPOSED ALGORITHM FOR REPLAY DETECTION

         The proposed method employs a Support Vector Machine based classifier for detect-
ing the logos that will swap across the screen in broadcasted video at start and end of the rep-
lay action.
The proposed algorithm consist of mainly TWO stages
       i. Train the SVM classifier with logo patterns and non logo patterns for classifying
            the frames in the video as logo frames and non logo frames.
      ii. Detecting logos at start and end of the replay action and extracting the frames be-
            tween them.
The flow charts for two phases of proposed algorithm are presented in Fig. 1.

                                                                        Start
                            Start


                                                                  Read the Input Video
           Extract Logo frames from the Input
           Video and take some Non logo frames

                                                           Extract the Entropy Features of each
             Extract the Entropy Features of a                            frames
                          frame

                                                          Test the frame on trained SVM Clas-
             Train the SVM Classifier with Entropy                        sifier
            Features of Logo frame and Non logo
                           frame
                                                          Detection of Two logos at beginning
                            END                              and end of each replay event


                                                            Extraction of frames between two
                                                                           logos



                                                                         END


Fig. 1 Flow diagram of proposed Algorithm, (a) Flow diagram for Training SVM Clas-
             sifier (b) Flow diagram for Replay Detection and Extraction.

                                                     23
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

In most of the sports video typical logo-transition lasts 0.5-0.8 seconds or 5-24 frames. In
logo-transition, there is an image frame that contains the clearest and the most complete logo.
Usually, the logo is highlighted and located at the middle part of a frame. With these prior
knowledge, the logo-samples from video can be detected and extracted.

3.1 Logo-transition Detection
        A typical logo-transition lasts 0.5-0.8 seconds or 15-24 frames. The frame-to-frame
difference is measured by intensity mean square difference (MSD). The difference sequence
after median filtering is a plateau-like shaped pattern (Fig. 3). Therefore, the logo-transition
detection can be performed through plateau-like shape detection over the frame-to-frame dif-
ference sequence. The proposed algorithm is
         i.   Compute frame-to-frame difference with MSD.
        ii.   Check the frame difference. Proceed to step 3 When the difference exceeds a
              threshold, otherwise go back to step 1) for next frame.
       iii.   Count the number of consecutive frame-differences that all exceed the frame-
              difference-threshold until encounter several consecutive frame-differences that
              drop below this threshold. If the counter exceeds a certain threshold, a wipe
              transition can be determined. Otherwise go back to step 1) for next frame.




                     Fig. 2 A logo-transition (8 frames are displayed)




                           Fig. 3 Frame difference sequences.
                      Original(left) and after median filtering(right)


3.2    Logo-sample Determination
       In a logo-transition, there is an image frame that contains the clearest and the most
complete logo. Usually, the logo is highlighted and located at the middle part of a frame.
With these prior knowledge, the logo-samples can be extracted.


                                                24
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

3.3      Feature Extraction
         The color, texture, shape and motion estimation have been commonly used features in
the literature, It is observed that color and motion features play a dominant role in the extrac-
tion of characteristics from the videos and hence the color information is adopted in this
work. Normally color distribution is estimated in the form of histogram, the proposed work
estimates the information of color content at different color intensity planes of images in
terms of entropy values. The entropy computed for a color plane in an image gives the aver-
age information conveyed by an image. In the pr oposed work the entropy of three layers of
RGB are computed and feature vector is written as in equation (1)

                                                                                  (1)

       Where
                  : Feature vector of   frame from video sequence
                 : Entropy of Red plane
                 : Entropy of green plane
                 : Entropy of blue plane
                   of an RGB frame from video sequence.
       The feature vectors as above are constructed for Logo templates extracted from vari-
ous videos and some Non logo frames and are used for training SVM.

3.4     Support Vector Machine (SVM)
        The support vector machine (SVM) algorithm seeks to maximize the margin around a
hyperplane that separates a positive class from a negative class [20]. Given a training dataset
with n samples (x1, y1); (x2,y2)……….(xn, yn), where xi is a feature vector in a v-
dimensional feature space and with labels yi {-1,1}belonging to either of two linearly se-
parable classes C1 and C2. Geometrically, the SVM modeling algorithm finds an optimal
hyper plane with the maximal margin to separate two classes, which requires to solve the op-
timization problem, as shown in equations (2) and (3).

                                                                            (2)


                                                                            (3)


       where, αi is the weight assigned to the training sample xi. If αi > 0, xi is called a sup-
port vector. C is a regulation parameter used to balance the training accuracy and the model
complexity so that a superior generalization capability can be achieved. K is a kernel func-
tion, which is used to measure the similarity between two samples. Different choices of ker-
nel functions have been proposed and extensively used in the past and the most popular are
the gaussian radial basis function (RBF), polynomial of a given degree, and multi layer per-
ceptron. These kernels are in general used, independently of the problem, for both discrete
and continuous data.




                                                25
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

3.5 SVM classifier based Replay detection Algorithm
        The proposed algorithm model uses priorly retrieved logos of sports video for the
training, then its features i.e its entropy values are extracted and trained to SVM classifier.
SVM is used as a classifier due to its clear classification method. Some randomly selected
frames without logo are also trained to SVM classifier as Non Logo templates.It classify two
Class vectors by a linear hyper plane , by which classifier can easily classify Logo and Non
Logo frames.


                                             Start


                                READ Video file (.avi,.mpeg) frame wise
                                        LOGO_COUNT =0




                               Extract Entropy features of current frame




                                  Test the Frame with trained SVM for
                                       LOGO or NON LOGO frame
                                  For LOGO frame LOGO_COUNT =1 for
                                      NON LOGO LOGO_COUNT =2




                                                Is LOGO
                                             frame detected




                                Extract all the frames between Current
                                LOGO detected and previous detection
                                write them to file which gives a replay
                                event.




                                                     End




                      Fig. 4 Flow chart of replay detection Algorithm




                                                      26
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

     Algorithm:
     Step 1: Extract the Logo frames from the input videos as explained in above sections.
     Step 2: Extract the entropy features of a frames as given in section 3.3.
     Step 3: Store these entropy features into vectors as feature patterns.
     Step 4: Train the SVM Classifier with the feature patterns.
     Step 5: Include non logo frames features in SVM training set.
     Step 5: The pair of logos can be detected as described in flowchart of Fig. 4.
     Step 6: Replay is Detected between a pair of logos.

Flow Diagram of Testing Phase to Identify an Logo Image in a Sports Video          and Replay
Detection Phase is shown in Fig. 4

4.       EXPERIMENTAL RESULTS AND DISCUSSIONS

        In the proposed work experimentation is conducted on data set of ten videos compris-
ing of soccer and IPL videos. The dataset of all videos logo frames are extracted from all ten
videos and are shown in next page in Fig.5, the features of which are used to train support
vector machine. The SVM trained with these features is employed for detection of pairs of
logos at the beginning and end of each replay action in soccer and cricket videos. The sum-
mary of experimentation is presented in Table 1.

          .




                                                27
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME




                            Fig. 5 Logo frames trained to SVM classifier

         The Table 1 summarizes the performance of proposed algorithm for replay detection
over ten sample sports video out of which 3 are soccer video and 7 are cricket video. The proposed
mechanism for replay detects the two logos that sandwich the replay action. The logos are detected
by SVM classifier by testing every frame of the video. The video segment between two consecu-
tive logos is picked as replay. The average values of performance parameters recall and precision
up to 94% and 97% for the algorithm indicates that the proposed mechanism is suitable for ex-
traction of replay actions in sports video and construct summary of sports videos.

                                                 28
 International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
 ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

                   TABLE1 Results of proposed algorithm for 10 test video samples

     Video File     Actual Num-     Total Num-      Undetected    False     Recall    Precision
       Name        ber of Replays   ber of replay    replays     detected
                                      correctly                  replays
                                      detected
                                       replays
     Video1.avi         10                9              1          1       90.00%    90.00%
     Video2.avi         60               56              4          2       93.33%    96.55%
     Video3.avi         20               20              0          0       100.00%   100.00%
     Video4.avi         34               31              3          1       91.17%    96.85%
     Video5.avi         8                 7              1          0       87.50%    100.00%
     Video6.avi         15               14              1          1       93.33%    93.33%
     Video7.avi         19               19              0          0       100.00%   100.00%
     Video8.avi         10               10              0          1       100.00%   90.90%
     Video9.avi         25               23              2          0       92.00%    100.00%
     Video10.avi        12               12              0          0       100.00%   100.00%

5.        CONCLUSION

         The proposed algorithm for SVM based method for replay detection and extraction
 from sports video is designed and experimented with sufficient number of soccer and cricket
 clips. The algorithm is implemented in Matlab 2010b and executed on Pentium® P6200 pro-
 cessor with 2GB RAM memory. The replay detection and extraction in soccer and cricket
 games are tested well with a sufficient number of video clips. The Proposed algorithms iden-
 tify a replay by finding a pair of logos and hence extract the replay by extracting all frames
 in-between the pair of logos. The proposed method identify a pair of logos by matching a
 entropy features of trained logos with the each frames histogram values of input videos. The
 proposed work uses support vector machine to classify between true logo and non logos. The
 replays are used as highlights of a game, and by which we formed a summary of sports video
 by these highlights.

 REFERENCES

 [1] H. Pan, P. van Beek, and M.I. Sezan, “Detection of slowmotion replay segments in sports
 video for highlights generation,”in Proc. of ICASSP'01, vol. 3, pp. 1649–1652, 2001.
 [2] X. F. Tong, H. Q. Lu, and Q. S. Liu, “A three-layer event detection framework and its
 application in soccer video,” in Proc.of ICME'04, vol. 3, pp. 1551–1554, 2004.
 [3] A. Ekin, A. M. Tekalp, and R. Mehrotra, “Automatic soccer video analysis and summari-
 zation,” IEEE Transactions on Image Processing, vol. 12, no. 7, pp. 796–807, 2003.
 [4] V. Kobla, D. DeMenthon, and D. Doermann, “Detection of slow-motion replays for iden-
 tifying sports videos,” in Proc. of MMSP'99, pp.135–140, 1999.
 [5] L. Wang, X. Liu, S. Lin, G. Y. Xu, and H.-Y. Shum, “Generic slow-motion replay detec-
 tion in sports video,” in Proc. of ICIP'04,vol.3, pp. 1585–1588, 2004.
 [6] J. J. Wang, E. Chng, and C. S. Xu, “Soccer replay detection using scene transition struc-
 ture analysis,” in Proc. of ICASSP'05,vol.2, pp. 433–436, 2005.
 [7] Y. Yang, S. X. Lin, Y. D. Zhang, and S. Tang, “A statistical framework for replay detec-
 tion in soccer video,” in Proc. Of ISCAS'08, pp. 3538–3541, 2008.

                                                    29
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print),
ISSN 0976 – 6456(Online) Volume 4, Issue 1, January - April 2013, © IAEME

[8] N. Babaguchi, Y. Kawai, Y. Yasugi, and T. Kitahashi, “Linking live and replay scenes in
broadcasted sports video,” in Proc. of MULTIMEDIA'00, pp. 205–208, 2000.
[9] H. Pan, B. Li, and M. I. Sezan, “Automatic detection of replay segments in broadcast
sports programs by detection of logos in scene transitions,” in Proc. of ICASSP'02, pp. 3385–
3388, 2002.
[10] L.-Y. Duan, M. Xu, Q. Tian, and C.-S. Xu, “Mean shift based video segment representa-
tion and applications to replay detection,” in Proc. of ICASSP'04, pp. 709–712, 2004.
[11] X. F. Tong, H. Q. Lu, Q. S. Liu, and H. L. Jin, “Replay detection in broadcasting sports
video,” in Proc. of ICIG'04, pp.337–340, 2004.
[12] Q. Huang, J. M. Hu, W. Hu, T. Wang, H. L. Bai, and Y. M. Zhang, “A reliable logo and
replay detector for sports video,” in Proc. of ICME'07, pp. 1695–1698, 2007.
[13] Z. H. Dang, J. Du, Q. M. Huang, and S. Q. Jiang, “Replay detection based on semi-
automatic logo template sequence extraction in sports video,” in Proc. of ICIG'07, pp. 839–
844, 2007.
[14] B. Han, Y. Yan, Z. H. Chen, C. Liu, and W. G. Wu, “A general framework for automatic
on-line replay detection in sports video,” In Proc. of MM'09, pp. 501–504, 2009.
[15] Reeja S R and Dr. N. P Kavya, “Motion Detection for Video Denoising – The State of
Art and the Challenges”, International Journal of Computer Engineering & Technology
(IJCET), Volume 3, Issue 2, 2012, pp. 518 - 525, ISSN Print: 0976 – 6367, ISSN Online:
0976 – 6375.




                                              30

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:1
posted:5/24/2013
language:
pages:11