Video Copy Detection Using F- SIFT and SURF Descriptor Algorithm

Transcription

1 IJMEIT// Vol. 2 Issue 10//October //Page No: //e-ISSN x 2014 Video Copy Detection Using F- SIFT and SURF Descriptor Algorithm Authors Shabeeba Ibrahim 1, Arun Kumar. M 2 1 Dept. of CSE, Ilahia College Of Engineering and Technology, Ernakulam, Kerala, India shabeeba1990@gmail.com 2 Assistant Professor, Ilahia College Of Engineering and Technology, Ernakulam, Kerala, India ABSTRACT arunpvmn@gmail.com This paper proposes Video Copy Detection (VCD) System and is done by using two keypoint descriptor and detector namely Flip- Invariant Scale Invariant Feature Transform (or F- SIFT) and Speeded Up Robust Feature (or SURF). Flip operation creates the mirror of an image and incorporate flip invariance properties to the existing descriptor algorithm called SIFT to get a new descriptor namely Flip- Invariant SIFT. Hence in a video copy detection system, flip operation need to be identified. VCD System also contain RANSAC algorithm, Hamming Distance Calculation and Forged Region Localisation. The time computation of F- SIFT is very high. To speed up the system, we presents a novel scale- and rotation- invariant detector and descriptor, coined SURF (Speeded-Up Robust Features) for video copy detection. SURF approximates or even outperforms previously proposed schemes with respect to repeatability, distinctiveness, and robustness, yet can be computed and compared much faster. Keywords- F- SIFT, Hamming Distance Calculation, Keypoint descriptor, RANSAC algorithm, SIFT, Speeded Up Robust Feature, video copy detection 1. INTRODUCTION In image processing and computer vision application, image matching is a fundamental aspect of many problems, including object or scene recognition, solving for 3D structure from multiple images, stereo correspondence, and motion tracking. Various local keypoint detector and descriptor [1], [2], [3], are proposed to finding the matching between two images or videos. Due to the success of SIFT [1], image local feature have been extensively employed in a variety of application such as computer vision, image processing, etc. The main characteristics of SIFT features are invariance to various image transformations such as rotation, scaling, lighting changes, displacements of pixels in a local region etc. It provides robust matching across a substantial range of affine distortion, change in 3D viewpoint, addition of noise, and change in illumination. The SIFT features are highly distinctive, that means a single feature can be correctly matched with high probability against a large database of features from many images. The shortcomings to SIFT is impervious to color, lighting and small pixel displacement, due to its spatial partitioning and 2D directional gradient binning. Despite these desirable properties, SIFT is not flip invariant. As a result, the descriptors extracted from two identical but flipped region could be completely different in feature space. In video copy detection, the effectiveness of feature point matching [4] is degraded and introduced extra Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 786

2 computational overhead [5] [7] for applications such as video copy detection. Flip is a transformation in which an image or video is flipped or reflected, creating a mirror of the original one. Flip or flip- like transformations are happen in real- world application. It has been one of the frequently used tricks [8], [9], in copyright infringement. Horizontal flipping is more commonly observed in images or videos, because this operation visually will not result in any apparent loss of image or video content. Generally, flip is a common operation used in creating nearduplicate videos. The main advantage of this operation is that it will not cause a change in the video content. Only the direction of information flow will get changed. So it is easy to create the copy of a video without much change in content. In short, the feature invariant descriptor used in video copy detection system must be invariant to flip transformation to identify flip. Flip property is added to the Scale- Invariant Feature Transform result to get a new descriptor called Flip- Invariant Scale- Invariant Feature Transform (F-SIFT) [3]. In the literature, large scale video copy detection tasks require a compact and computational-efficient descriptor that is robust to various transformations. Video copy detection [5] is the process of detecting illegally copied videos by analyzing them and comparing them to the original content. The main goal of this process is to protect a video creator s intellectual property. The shortcoming to F-SIFT is replaced by using another descriptor algorithm called SURF (Speeded Up Robust Feature) [2]. It is both scale- and rotationinvariant interest point detector and descriptor. It approximates or even outperforms previously proposed schemes with respect to repeatability, distinctiveness, and robustness, yet can be computed and compared much faster. The remaining of this paper is organized as follows. Section 2 reviews variants of local descriptors and their utilization for video copy detection. Section 3 describes the proposed system and Section 4 describes the solution methodology which includes video acquisition method, extraction of descriptors from local regions by using F-SIFT and SURF, RANSAC algorithm, Hamming Distance Calculation, and Forged Region Localization. Finally Section 5 concludes this paper. 2. RELATED WORKS Flip invariance property is not considered in existing descriptor algorithm namely SIFT [1]. The SIFT descriptor differ by the partitioning scheme of local region, which divides a region into 4 4 blocks and describes each grid with an 8 directional gradient histogram. It generates the feature by concatenating the histograms in row major order from left to right and the histogram bins in right to left or in clockwise manner. So the flip transformation of the region will disorder the placement of bins and blocks. Due to the predefined order of feature scanning, the different version of descriptor was developed. The solutions for this problem are, altering the partitioning scheme or scanning order [10], [13], and feature transformation [12]. There are many flip invariant descriptors which includes RIFT [10], SPIN [10], MI-SIFT [12] and FIND [13]. RIFT (Rotation invariant SIFT) descriptor is a rotation-invariant generalization of SIFT proposed by S. Lazebnik in [10]. One of the earliest flip invariant SIFT descriptor. This descriptor is somewhat sensitive to scale changes and is less discriminative than original SIFT.A circular normalized patch is considered around each detected interest points. This circular patch is divided into concentric rings of equal width. And within each ring, from each division an 8 directional gradient orientation histogram are computed. The concatenation of gradient orientation histogram about all the identified key points yields the RIFT descriptor. To preserve the rotation in-variance property, at each key point, the orientation is measured relative to the direction pointing outward from the center. RIFT uses four rings and eight histogram orientations, yielding 32-dimensional descriptors. Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 787

3 SPIN is also proposed by S. Lazebnik in [10]. It is also a flip invariant descriptor. SPIN preserves flip invariance property by taking into account the spatial information. Here, it en-codes a region using 2D histogram of pixel intensity and also considers the distance from region center. SPIN as well as RIFT are outperformed by SIFT. Mirror and Inversion invariant SIFT (MI- SIFT) descriptor was proposed by R.Ma in [12]. This descriptor improves the SIFT descriptor by enhancing the invariance to mirror reflections and grayscale inversions. Mirror reflection and Inversion invariance is achieved by combining the SIFT histogram bins of both the image and its mirror reflected image. This approach provides a unified descriptor for the original, mirror reflected and grayscale inverted images with an additional cost of computation. MI-SIFT perform better than SIFT in finding transformations including mirror reflection and inversions. It also achieves comparable performance in image matching with ordinary images. X. Guo [13] introduced a novel flip invariant descriptor through a new cell ordering scheme namely FIND. As like SIFT, it also adopts the DoG detector. FIND employs an overlap extension strategy to obtain the feature descriptor. For each detected key points, it reads the 8 directional gradient histograms by following an S order. Here, for a given image, the descriptor obtained before and after a flip operation are mirror of each other. FIND exhibits stable performance in both flip and non-flip case. FIND is also tolerant to scaling, rotation and affine transformations. Flip operation have been viewed as in widely used copyright infringement, TRECVID [9], [14]. The TREC Video Retrieval Evaluation (TRECVID) 2008 is a Text Retrieval Conference (TREC)-style video analysis and retrieval evaluation, the goal of which remains to promote progress in content-based exploitation of digital video via open, metrics based evaluation. In TRECVID copy detection task (CCD) [9], [14], for instance, video copies as a result of flips are regarded as one of the major testing items. Flipped copy detection is occurred in it with some problems. The problem of flipped copy detection is engineered by indexing two SIFT descriptors for each region [6], [15], of which one of them is computed by simulating flip operation. This results in significant increase in both indexing time and memory consumptions. Other descriptor algorithms are PCA- SIFT [16], GLOH [17], MIFT [16], BRIEF [16], ORB [16] etc. PCA- SIFT is an improved SIFT descriptor proposed by Ke and Suk-thankar in [16]. This descriptor is more compact, highly distinctive and as robust as SIFT. PCA-SIFT, like SIFT encode the characteristics of image by considering the feature point s neighborhood. Here, SIFT descriptor is improved by applying Principal Component Analysis for dimensionality reduction. Gradient Location Orientation Histogram (GLOH) [17] is pro-posed as an extension of the standard SIFT descriptor. It is a 64 dimensional descriptor. Unlike SIFT, in GLOH histogram representation considers more spatial regions. The GLOH feature descriptor is constructed using Histogram of location and Orientation of pixels in a window around the interest point. MIFT (Mirror reflection Invariant Feature Transform) [18] is a framework for providing feature descriptor and is a local feature descriptor providing mirror reflection invariance while preserving existing merits of SIFT descriptor. BRIEF [19] is a general-purpose feature point descriptor that can be combined with arbitrary detectors. It is robust to typical classes of photometric and geometric image transformations. BRIEF is targeting real-time applications. ORB (Oriented and Rotated BRIEF) is an efficient, fast binary descriptor based on BRIEF descriptor. It is introduced by E. Rublee, V. R Rabaud, K. Konolige, and G. Bradski in [20]. ORB feature descriptor achieves rotation invariance and noise resistance. ORB is a 32 bit binary feature descriptor. In ORB, for detecting key points, it uses the FAST key point detector. To the best of our knowledge, no work has yet seriously addressed the issue of detection Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 788

4 performance by contrasting features with and without incorporating flip invariance property. 3. VIDEO COPY DETECTION (VCD) The proposed system describes Video Copy Detection (VCD). Video copy detection is the process of detecting illegally copied videos by analyzing them and comparing them to original content. The goal of this process is to find out the copy detection in video and to protect a video creator's intellectual property. The copy move attack can be mainly divided into two; splicing and multiple cloning.. In splicing, a part of the image or video is copied and pasted it into the same one with the intend to cover an important image or video feature. In multiple cloning, a part of the image or video is copied and pasted somewhere else with the intend to cover an important feature. To demonstrate the use of F-SIFT and SURF for copy detection, we adopt our framework originally developed for near-duplicate video detection [15]. Modifications to the framework are made considering the new features introduced by F-SIFT. SURF is more faster than F- SIFT, so the limitations of F- SIFT is overcome by using SURF. SURF is a scale- and rotation-invariant interest point detector and descriptor. 4. SOLUTION METHODOLOGY We proposed a method to find the copy detection in video which consist of five steps: 1. Video acquisition 2. Descriptor point calculation using F- SIFT and SURF. 3. RANSAC Algorithm 4. Hamming distance calculation 5. Forged region localization 4.1 Video Acquisition Video Acquisition or Video capture is the process of converting an analog video signal such as that produced by a video camera or DVD player to digital video. The resulting digital data are computer files referred to as a digital video stream, or more often, simply video stream. This is in contrast with screen casting, in which previously digitized video is captured while displayed on a digital monitor. 4.2 Descriptor Point Calculation Using F- SIFT F- SIFT is a new descriptor that incorporates the flip invariance property to SIFT, while preserving its original property. Flip invariant detectors are capable of locating regions under various transformations. But the main problem is due to the feature descriptor because of flip operation. So the solution is to propose F- SIFT Descriptor which revises SIFT to be flip invariant. a. F-SIFT Descriptor The aim is to enrich SIFT to be flip invariant while preserving its original properties including the grid-based quantization. Flip transformation can happen along arbitrary axis. We considered flip as a decomposition of flip along a predefined axis followed by a rotation. To make a descriptor flip invariant is done by normalizing a local region before feature extraction through rotating the region to a predefined axis and then flipping it along the axis. F-SIFT perform a selective flipping on the image regions based on the dominant curl associated with the regions. Flip invariant descriptor is obtained by first rotating the region patches to its dominant orientation and estimating the Curl associated with the regions. The direction of the curl indicates the direction of rotation. if a region has been rotated to its dominant orientation which is the case for regions identified by key point detectors, the normalization can be simply done by flipping the region horizontally (or vertically). In other words, a prominent solution for flip invariance is to determine whether flip should be performed before extracting local feature from the region. If the curl is negative it denotes that the region patch is flipped and is needed to be rotated anti clockwise. Then SIFT descriptors are extracted from these Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 789

5 normalized regions. In short, in F-SIFT, the regions are normalized geometrically by flipping horizontally or vertically and complementing their dominant orientations. For flipped images F-SIFT out- performs SIFT. F- SIFT generate descriptors as follows: Given a region rotated to its dominant orientation, first the curl is computed to estimate the flow direction of image i.e. either clockwise or counter clockwise. Definition of curl is, mathematically as a vector operator that describes the infinitesimal rotation of a vector field. The direction of curl is the axis of rotation determined by the right-hand rule. Given a vector field F (x, y, z) defined in R 3 which is differentiable in a region, the curl of F is given by, Generally, the dominant curl C have only two possible direction, either clockwise or anticlockwise. The direction for C is represented by its sign. If the vector field has been flipped along an arbitrary axis (horizontal or vertical), then sign is changed. In short, the normalization can be done by flipping the region along an arbitrary axis. If the sign of flow is clockwise, the normalization is performed by flipping the region whose sign are anti- clockwise. In other words, flipping the region is based on the sign of C. For robustness, Eqn. 3 can be further enhanced by assigning higher weights to vectors closer to region center as following,. (1) (4) According to Stokes theorem, the integration of curl in a vector field can be expressed by, ϵ R 3. (2) But in this case, the curl is defined in a 2D discrete vector field I. The curl at a point is the cross product on the first order partial derivatives along x and y directions respectively. The flow (or dominant curl) along the tangent direction can be defined by, where (3) and θ is the angle from direction of the gradient vector to the tangent of the circle passing through (x, y). where the flow is weighted by a Gaussian kernel G of size σ equal to the radius of local region. To summarize, F-SIFT generates descriptors as following. Given a region rotated to its dominant orientation, Eqn. 4 is computed to estimate the flow direction of either clockwise or counter clockwise. F-SIFT ensures flip invariance property by enforcing that the flows of all regions should follow a predefined direction indicated by the sign of C in Eqn. 4. For regions whose flows are opposite of the predefined direction, flipping the regions along the horizontal (or vertical) axis as well as complementing their dominant orientations are performed to geometrically normalize the regions. SIFT descriptors are then extracted from the normalized regions. If the curl is negative it denotes that the region patch is flipped and is needed to be rotated anti clockwise. In other words, F SIFT operates directly on SIFT and preserves its original property. Selective flipping based on dominant curl analysis is performed prior to extracting flip invariant descriptor. Compared to SIFT, the overhead involved in F-SIFT is merely the computation of Eqn. 4 which is cheap to calculate. During flip, F- SIFT exhibit significantly stronger performance than SIFT. The numbers of matching pairs recovered by F- SIFT is Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 790

6 much more than SIFT. F-SIFT descriptor computation involves high computational overhead. Performance degrades if there are errors during finding the curl and it cause to finding fewer matching pairs. The extraction of F-SIFT descriptors is approximately one third slower than SIFT Using SURF SURF (Speeded Up Robust Features) is a scale- and rotation invariant interest point detector and descriptor. SURF is a robust local feature descriptor and is based on sums of 2D Haar wavelet responses. SURF has proven its efficiency in various computer vision tasks. SURF descriptors are more commonly used for applications like object recognition and 3D reconstruction. The extraction of such a r obust descriptor is inspired from the efficient SIFT descriptor. SURF is a combination of novel detection, description, and matching algorithm. It is a detector-descriptor scheme. The detector is based on the Hessian matrix, but uses a very basic approximation, just as DoG is a very basic Laplacian- based detector. It relies on integral images to reduce the computation time and we therefore call it the Fast-Hessian detector. The descriptor, on the other hand, describes a distribution of Haar wavelet responses within the interest point neighborhood. a. Fast-Hessian Detector Fast- Hessian Detector is based on the Hessian matrix because of its good performance in computation time and accuracy. However, rather than using a different measure for selecting the location and the scale (as was done in the Hessian- Laplace detector [11]), we rely on the determinant of the Hessian for both. Given a point x = (x, y) in an image I, the Hessian matrix H (x, σ) in x at scale σ is defined as follows, where σ is the convolution of the Gaussian second order derivative with the image I in point x, and similarly for σ and σ. Gaussians are optimal for scale-space analysis. The Hessian matrix contains 2nd order derivatives. However, in practice, the Gaussian needs to be discretised and cropped and is shown in Figure.1(a). So the Computation costs increase as filter size increases. As Gaussian filters are nonideal in any case, and given Lowe s success with LoG approximations, we push the approximation even further with box filters (Figure.1(b) ). These approximate second order Gaussian derivatives, and can be evaluated very fast using integral images, independently of size. The performance is comparable to the one using the discretised and cropped Gaussians. (a) (b) Fig. 3. The (discretised and cropped) Gaussian second order partial derivatives in y- direction and xy- direction, and our approximations thereof using box filters. The grey regions are equal to zero. H(x,σ)= σ σ σ σ (5) The 9 9 box filters in Fig. 3 are approximations for Gaussian second order derivatives with σ = 1.2 and represent our lowest scale (i.e. highest spatial resolution). We denote Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 791

7 approximations by D xx, D yy, and D xy. The weights applied to the rectangular regions are kept simple for computational efficiency, but we need to further balance the relative weights in the expression for the Hessian s determinant with, where x F is the Frobenius norm. This yields, (6) Furthermore, the filter responses are normalized with respect to the mask size. This guarantees a constant Frobenius norm for any filter size. b. SURF Descriptor The good performance of SIFT compared to other descriptor is remarkable. It s mixing of crudely localised information and the distribution of gradient related features seems to yield good distinctive power while fending off the effects of localisation errors in terms of scale or space. Using relative strengths and orientations of gradients reduces the effect of photometric changes. The proposed SURF descriptor is based on similar properties, with a complexity stripped down even further. The first step consists of fixing a reproducible orientation based on information from a circular region around the interest point. Then, we construct a square region aligned to the selected orientation, and extract the SURF descriptor from it. These two steps are now explained in turn. Step 1. Orientation Assignment In order to be invariant to rotation, we identify a reproducible orientation for the interest points. For that purpose, we first calculate the Haarwavelet responses in x and y direction, and this in a circular neighborhood of radius 6s around the interest point, with s the scale at which the interest point was detected. Also the sampling step is scale dependent and chosen to be s. In keeping with the rest, also the wavelet responses are computed at that current scale s. Accordingly, at high scales the size of the wavelets is big. Therefore, we use again integral images for fast filtering. Only six operations are needed to compute the response in x or y direction at any scale. The side length of the wavelets is 4s. Once the wavelet responses are calculated and weighted with a Gaussian (σ = 2.5s) centered at the interest point, the responses are represented as vectors in a space with the horizontal response strength along the abscissa and the vertical response strength along the ordinate. The dominant orientation is estimated by calculating the sum of all responses within a sliding orientation window covering an angle of. The horizontal and vertical responses within the window are summed. The two summed responses then yield a new vector. The longest such vector lends its orientation to the interest point. The size of the sliding window is a parameter, which has been chosen experimentally. Small sizes fire on single dominating wavelet responses, large sizes yield maxima in vector length that are not outspoken. Both result in an unstable orientation of the interest region. Step 2. Descriptor Components For the extraction of the descriptor, the first step consists of constructing a square region centered around the interest point, and oriented along the orientation selected in the previous section. For the upright version, this transformation is not necessary. The size of this window is 20s. The region is split up regularly into smaller 4 4 square sub-regions. This keeps important spatial information in. For each sub-region, we compute a few simple features at 5 5 regularly spaced sample points. For reasons of simplicity, we call d x the Haar wavelet response in horizontal direction and d Y the Haar wavelet response in vertical direction (filter size 2s). Horizontal and vertical here is defined in relation to the selected interest point orientation. To increase the robustness towards geometric deformations and localisation errors, the responses Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 792

8 d X and d Y are first weighted with a Gaussian (σ = 3.3s) centered at the interest point. Then, the wavelet responses d x and d Y are summed up over each sub region and form a first set of entries to the feature vector. In order to bring in information about the polarity of the intensity changes, we also extract the sum of the absolute values of the responses, d x and d Y. Hence, each sub-region has a four-dimensional descriptor vector v for its underlying intensity structure. v (7) This results in a descriptor vector for all 4 4 sub-regions of length 64. The wavelet responses are invariant to a bias in illumination (offset). Invariance to contrast (a scale factor) is achieved by turning the descriptor into a unit vector. In the matching stage, we only compare features if they have the same type of contrast. Hence, this minimal information allows for faster matching and gives a slight increase in performance. SURF is proven to be several times faster than SIFT. SURF is more robust to image transformations like rotation, scaling and noise. This is a fast and performant interest point detection description scheme which outperforms the current state-of-the art, both in speed and accuracy. The descriptor is easily extendable for the description of affine invariant regions. Future work will aim at optimising the code for additional speed up. It approximates or even outperforms previously proposed schemes with respect to repeatability, distinctiveness, and robustness, yet can be computed and compared much faster. smallest set possible and proceeds to enlarge this set with consistent data points. The RANSAC Algorithm is shown in below The number of iterations, N, is chosen high enough to ensure that the probability p (usually set to 0.99) that at least one of the sets of random samples does not include an outlier. Let u represent the probability that any selected data point is an inlier and v = 1 u the probability of observing an outlier. N iterations of the minimum number of points denoted m are required, where 1 p = (1 u m ) N (8) and thus with some manipulation, Algorithm 1: RANSAC Algorithm (9) 1: Select randomly the minimum number of points required to determine the model parameters. 2: Solve for the parameters of the model. 3: Determine how many points from the set of all points fit with a predefined tolerance ϵ. 4: If the fraction of the number of inliers over the total number points in the set exceeds a predefined threshold, re-estimate the model parameters using all the identified inliers and terminate. 5: Otherwise, repeat steps 1 through (maximum of N times). 4.3 RANSAC Algorithm The RANdom SAmple Consensus (RANSAC) algorithm proposed by Fischler and Bolles is a general parameter estimation approach. RANSAC is a resampling technique that generates candidate solutions by using the minimum number observations (data points) required to estimate the underlying model parameters. RANSAC uses the 4.4 Hamming distance calculation Next step is to find out the minimum distance between the two points. For that we use Hamming distance calculation. It is used to compare the two images. A hamming signature is used to reduce the quantization loss. The Hamming distance d(x, y) between two vectors x, y ϵ F (n) is the number of coefficients in which they differ. d satisfies the following conditions for a metric: Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 793

9 a. d (x, y) 0 and d (x, y) = 0 if and only if x = y b. d(x, y) = d(y, x) c. d (x, z) d (x, y) + d (y, z) for any x, y, z ϵ F (n) ). 4.5 Forged Region Localization The last step is the forged region localization. Here check the similarity between the query image Q and reference image R and is given by, (10) where h(q, p) is the distance [16] between Hamming signatures of q and p. The notation BoW (Q) denotes the bag-of- words of Q. Each key frame extracted from videos is then represented as a bag-of- visualwords (BoW). Similarity between two key frames is revised by weighting the significance of matched words based on their Hamming distance and matching confidence. 5 CONCLUSIONS Copy move detection can be finding out by using F- SIFT and SURF. These are two descriptor algorithms to find out the key points in each frame of video. The time of computation of F-SIFT is high. We have presented a fast and performant interest point detection-description scheme which outperforms the current state-of-the art, both in speed and accuracy. This descriptor is known as SURF. The descriptor is easily extendable for the description of affine invariant regions. The VCD system also contain RANSAC algorithm, Hamming Distance Calculation and Forged Region Localisation for finding the copy move attack in video. ACKNOWLEDGEMENT The authors would like to thank the Management and Principal and Head of the Department (CSE) of Ilahia College of Engineering and Technology for their support and help in completing this work. REFERENCES 1. D. Lowe, Distinctive image features from scale-invariant keypoints, Int. J. Comput. Vis., vol. 60, no. 2, pp , W. L. Zhao and C. Ngo, Flip-Invariant SIFT for Copy and Object Detection, IEEE Trans. Image Processing, vol. 22, no. 3, MARCH 2013, pp H. Bay, A. Ess, T. Tuytelaars, and L. V. Gool, SURF: Speeded up robust features, Comput. Vis. Image Understand., vol. 110, no. 3, pp , M.-C. Yeh and K.-T. Cheng, A compact, effective descriptor for video copy detection, in Proc. Int. Conf. Multimedia, 2009, pp Z. Liu, T. Liu, D. Gibbon, and B. Shahraray, Effective and scalable video copy detection, in Proc. Int. Conf. Multimedia Inf. Retr., 2010, pp M. Douze, H. Jégou, and C. Schmid, An image-based approach to video copy detection with spatio-temporal post filtering, IEEE Trans. Multimedia, vol. 12, no. 4, pp , Jun Y.-D. Zhang, K. Gao, X. Wu, H. Xie, W. Zhang, and Z.-D. Mao, TRECVID 2009 of MCG-ICT-CAS, in Proc. NIST TREVCID Workshop,2009, pp J. Law-To, A. Joly, and N. Boujemaa. (2007). Muscle-VCD-2007: A Live Benchmark for Video Copy Detection [Online]. Available: inria.fr/imedia/civr-bench/ 9. TRECVID. (2008) [Online]. Available: projects/trecvid/ 10. Lazebnik, C. Schmid, and J. Ponce, A sparse texture representation using local affine regions, IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 8, pp , Aug K. Mikolajczyk and C. Schmid, A performance evaluation of local descriptors, IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 10, pp , Oct Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 794

10 12. R. Ma, J. Chen, and Z. Su, MI-SIFT: Mirror and inversion invariant generalization for SIFT descriptor, in Proc. Int. Conf. Image Video Retr., 2010, pp X. Guo and X. Cao, FIND: A neat flip invariant descriptor, in Proc. Int. Conf. Pattern Recognit., Aug. 2010, pp A. F. Smeaton, P. Over, and W. Kraaij, Evaluation campaigns and TRECVid, in Proc. Int. Conf. Multimedia Inf Retr., 2006, pp W.-L. Zhao, X. Wu, and C.-W. Ngo, On the annotation of web videos by efficient near-duplicate search, IEEE Trans. Multimedia, vol. 12, no. 5, pp , Aug H. Jégou, M. Douze, and C. Schmid, Hamming embedding and weak geometric consistency for large scale image search, in Proc. Eur. Conf. Comput. Vis., 2008, pp Shabeeba Ibrahim, Arun Kumar. M IJMEIT Volume 2 Issue 10 October 2014 Page 795