papera full reference algorithm for dropped frames identification in uncompressed video using...

Upload: mthakurjiit

Post on 03-Apr-2018

222 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    1/12

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    1Manish K Thakur, 2Vikas Saxena, 3J P Gupta*1,Jaypee Institute of Information Technology, Noida, India, [email protected]

    2,Jaypee Institute of Information Technology, Noida, India, [email protected],Sharda University, Greater Noida, India, [email protected]

    AbstractDropped frames identification in a given video stream is always a challenging task for research

    community due to required heavy computation. To preserve the quality of service during any

    processing over visual information, drop of frame is least desired. Many contemporary work identifies

    the frame drop in terms of full reference algorithm (where reference and distorted video streams are

    available for comparison), or reduced reference algorithm (where some information about reference

    video are available) or no reference algorithm (where information about reference video are not

    available). This paper presents a novel full reference heuristic approach using genetic algorithm which

    identifies the dropped frame indices in a given distorted video stream with respect to original video

    stream. The proposed algorithm efficiently identifies dropped frame indices even if reference video

    stream contains repeated frames and spatially distorted too with low or high spatial distortions. The

    proposed algorithm is simulated and tested with 12 video streams. Simulation results suggested that it

    is more efficient with a video stream having lesser repeated frames.

    Keywords: Frame Drop; Temporal Distortion; Spatial Distortion; Peak Signal To Noise Ratio(PSNR); Longest Increasing Sequence; Genetic Algorithm

    1. Introduction

    In the last decade, enormous and unprecedented growth of information in various domainsacross the world demanded the significant application of high ended multimedia based

    technology to cater the effective data flow communication, storage, retrieval, compression, andediting at various platform. Due to various processing over visual information, it may get

    distorted when reaches to end user [1-4]. Apart from accidental distortions during processing, itincludes some malafied intentions/video tampering where one might distort pixels of a video

    frame, or drop some video frames, or insert average frame into video [5-8]. Consequently thesedistortions lead to loss of visual information and if tampered intentionally or accidentally must

    be identified.Distortions in a video stream appear either in intra frame or at inter-frame level. Intra frame

    distortions or spatial distortions (SD) arises due to change of pixel bits of a reference video

    frame; this might be multifactorial governed viz due to introduced bit error during

    communication, or other manipulation over pixel bits of a video frame like compression, bitinversion, acquisition, watermarking, editing, and storage while, inter frame distortions or

    temporal distortions (TD) arises with respect to time like, drop of video frames, swapping ofvideo frames, and frame averaging [8-10].

    Figure 1 depicts an example of SD where, figure 1a is an original video frame i.e. without

    distortion and figure 1b is the distorted video frame which is spatially distorted due tointroduced bit error as highlighted by blue and orange color. Figure 2 depicts an example of TD

    where figure 2a is the original video stream of 5 frames, whereas figure 2b, figure 2c, and figure

    2d are the temporally distorted video streams due to frame drop, frame swapping, and frameaveraging respectively. Due to frame drop or frame averaging the order of unaffected frames in

    distorted video is not disturbed while frame order will be changed in distorted video stream due

    to frame swapping.

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    International Journal of Digital Content Technology and its Applications(JDCTA)

    Volume6,Number20,November 2012

    doi:10.4156/jdcta.vol6.issue20.61

    562

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    2/12

    Figure 1. (a) is original video frame and (b) is spatially distorted video frame

    Figure 2. (a) original video frames, (b) 3rd frame as dropped frame, (c) video frames 3 and 4

    are swapped, and (d) average frame inserted at frame number 4

    As frame drop is least desired, dropped frames identification in a video stream is always a

    challenging task for research community. There are significant work in terms of quality metricswhich identify the dropped frames under full reference (FR), reduced reference (RR), and no

    reference (RR). S. Wolf in his work identified the dropped frames identification as a problemwhich requires heavy computation [11].

    Therefore to reduce the required computational time, we present in this paper a novel fullreference algorithm which identifies the dropped frame indices in a distorted video stream with

    respect to given reference (or original) video stream. As shown in section 4, presented heuristicefficiently identifies the dropped frames indices. Here, we have used genetic algorithms

    mutation operator to resolve issues which arises after applying the heuristic.This paper is organized as follows: apart from introduction in section 1, section 2 describes

    the problem, section 3 deals with proposed solution, simulation and analysis of solution

    approach is presented in section 4 followed by conclusion and references.

    2. The Problem

    Many challenges arise during processing of multimedia data which are voluminous andcontinuous in nature. One of such processing is its communication over multimedia network,

    where due to packet loss; video frames may not be received at users end and resulting in drop ofvideo frames. The drop of frames in multimedia communication network is accidental and a

    simple error recovery is to replay the last received frame. However this remedial approach is on

    the cost of received streams quality which is undesirable. Further, video frames in a video

    stream may be dropped with malafied intentions. In either case it is required to identify thedropped frames and analyse its impact over the video quality.

    Researchers are continuously addressing and resolving the mentioned problem. One of thecontemporary works presented by S. Wolf proposed NR and RR metrics to identify dropped

    frames by maintaining motion energy time history of a video clip [11]. As it is NR or RR metric

    6060606060606060

    8080808080808080

    1010101010101010

    1212121212121212

    1414141414141414

    1616161616161616

    1818181818181818

    2020202020202020

    6060606060606060

    3550808080808080

    4530101212121010

    3228121314131212

    4434201615161414

    3836201717181616

    1818181818181818

    2020202020202020

    (a) (b)

    (d)

    Avg

    frame

    1 23

    5

    12 4

    35

    (c)

    1 23

    45Frame #

    (a) (b)

    12 3

    4 5

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    563

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    3/12

    where we do not (or little information) have the information about reference video stream, other

    distortions (SD) which may be introduced during processing were unhandled.

    In another contemporary work being carried out by Graphics and media lab of Moscow State

    University (MSU) presented dropped frame metric which identifies the dropped frames bycalculating each frame difference with the previous one [12]. As it identifies the dropped frames

    by computing the difference of each frame with previous frame; it might restrict the metric toexplore other frames for similarity.

    A work being carried out by Nishikawa K et. al. presented a no reference method forestimating the degradation in video stream due to packet loss by estimating peak signal to noise

    ratio (PSNR) between frames affected by packet loss and the originally encoded frames [13].

    Like MSUs metric, this method also utilizes the similarity between adjacent frames and thus

    restricting the method to explore other frames (non-adjacent) for similarity.In subsequent paragraph we are defining the problem of dropped frames identification in a

    full reference (FR) mode by addressing unhandled issues in contemporarys work like spatiallydistorted video stream and comparing a video frame with all possible frames for similarity.

    Say, we have given a reference video stream (VR) with m video frames as VR1, VR2, VR3 .. VRm.

    The reference video VR is distorted (intentionally or accidentally) by one of the TD i.e. frame

    drop to give distorted video stream (VD) with n video frames as VD1, VD2 .. VDn, where m>n.

    Further, some video frames of the distorted video stream VD are spatially distorted too byintroducing bit errors at different pixel bits. The introduced SD may be high or low. Apart from

    dropped frames in VD, following cases are possible related with introduced SD in video frames:C2.1: There is no SD in video frames ofVD.C2.2: Some video frames ofVD are distorted with low SD (we call change of 5 to 10 % in at

    least 10% pixel bits as low SD, as highlighted by orange color in figure 1b)C2.3: Some video frames ofVD are distorted with high SD (we call change of more than 10%

    in at least 10% bits as high SD, as highlighted by blue color in figure 1b)

    It is desired to identify the dropped frames in distorted video stream VD with respect tooriginal video stream VR which we call as reference video stream.

    3. Proposed Solution

    As a solution to the problem defined in previous section, this section presents a heuristicapproach which identifies dropped frame indices by maximizing the overall similarity betweenvideo frames ofVR and VD.

    Considering that we have only to deal with frame drop (i.e. frame order ofVR is retained inVD), a frame VDi ofVDmight present at the indices ranging from i to i + dropped frames count

    in VR. Therefore unlike previous approaches (section 2) where a frame was compared with

    adjacent frames, it is now required to compare it with range of frames (as mentioned) for the

    similarity.In the example of figure 3, where VR is with m frames and VD is with n = m-1 frames, first

    frame ofVD will be either first frame or second frame ofVR. Thus, it is required to compare firstframe ofVD only with first two frames ofVR for the similarity and so on.

    There are several quality metrics which through different indexes identifies the introduced

    noise or similarity between two given images or video frames. PSNR, Mean square error (MSE),

    and Structural similarity index (SSIM) are few of them [14-18]. PSNR of 100 dB between twovideo frames indicates that there is no noise in both video frames, i.e. both are identical whereas

    PSNR of less than 100 dB indicates noise in compared frames. Due to simplest computationrequired for computation of PSNR, we have used the PSNR index to identify the dissimilarity

    between two video frames.

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    564

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    4/12

    Figure 3. An example of compared frames ofVR (m frames) and VD(n = m1 frames)

    Subsequent paragraph describes the steps of our solution approach:Step 3.1: Input reference video VR = {VR1, VR2, .. VRm} with m frames and distorted video VD =

    {VD1, VD2, .VRn} with n frames, where m > n.Step 3.2: Compute PSNR of each frame ofVD with respect to m-n+1 frames ofVR(i

    th frame

    of VD is compared with ith to (i+m-n+1) th frame of VR) and store these indices into a two

    dimensional matrix having n rows corresponding to length of VD and m+2 columns

    corresponding to length ofVR with two additional fields Max and Index to store highest PSNR ina row and its index respectively i.e. index will be the column number corresponding to highest

    PSNR in a row. If more than one field in a row have same highest PSNR then store the smallerindex into theIndex field. We call this matrix as PSNR difference matrixDiffMat.

    Following cases can arise in Index field ofDiffMat as elaborated by examples in figure 4,figure 5, and figure 6 where DiffMat has been created with 12 reference video frames and 6

    distorted video frames. As there are drop of 6 video frames, i th frame ofVD (for all i = 1 to n)

    has been compared with ith to (i+6)th frames of VR and computed PSNR has been stored inrespective field ofDiffMat.

    C3.2.1: If there are only TD or frame drop errors (i.e. no SD) in VD, then, at least one field in

    each row of matrix DiffMat will contain PSNR as 100dB which indicate the matching frame ofVD in VR. Since there is no SD, indices in Index field of each row should be in increasing order.

    This case has been presented by an example in figure 4.C3.2.2: Along with TD, if some or all frames ofVD are distorted spatially too, then Max field

    of a row ofDiffMatmight be or might not be 100dB. Since, there is SD in some or all frames of

    VD, these indices in Index field (a) might be in increasing order or (b) in random order. CasesC3.2.2a and C3.2.2b have been presented by examples in figure 5 and figure 6 respectively.

    Step 3.3: Find out longest increasing sequence (LIS) [19-22] in the Index field ofDiffMat.

    Following cases (C3.3.1 and C3.3.2) will arise while computing length of identified LIS.

    VR1 VR2 VR3 VR4 VR5 VR6 VR7 VR8 VR9 VR10 VR11 VR12 Max Index

    VD1 100 60 50 45 35 33 29 100 1

    VD2 100 48 44 41 38 30 26 100 2

    VD3 31 32 36 100 58 50 38 100 6

    VD4 31 34 35 100 55 51 32 100 7

    VD5 36 37 39 44 100 61 55 100 9

    VD6 26 27 29 31 32 58 100 100 12

    Figure 4. An example of case C3.2.1 whereDiffMat is created with m = 12 and n = 6

    VR1 VR2 VR3 VR4 VR5 VR6 VR7 VR8 VR9 VR10 VR11 VR12 Max Index

    VD1 65 60 50 45 35 33 29 65 1

    VD2 57 48 44 41 38 30 26 57 2

    VD3 31 32 36 59 58 50 38 59 6

    VD4 31 34 35 63 55 51 32 63 7

    VD5 36 37 39 44 80 61 55 80 9

    VD6 26 27 29 31 32 58 64 64 12

    Figure 5. An example of case C3.2.2a whereDiffMatis created with m = 12 and n = 6

    VD with n = m-1 frames

    VR with m frames

    .

    .

    VR1 VR2 VR3 VRm-1 VRm

    VD1 VD2 VD3 VDn

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    565

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    5/12

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    6/12

    As indicated in figure 7b, apart from entries as 1 in PopArr corresponding to LIS in

    preceding example, we randomly filled only one 1 (because in this example only 6 entr ies as 1

    is allowed) at index number 11 into the PopArrand other fields as 0.Step 3.3.3: Create prefix sum array PfxSum which will store the prefix sum [23] of the array

    PopArr (figure 7c). Compute fitness score of the population by averaging the PSNR ofDiffMat[PfxSum[i]][i] for all i = 1 to n where PopArr[i] = 1. If fitness score of currentpopulat ion is greater than MinFit then assign current fitness score into MinFitand make a copy

    ofPopArrinto Final.Generate new population into PopArrusing mutation operator. Use bit string mutation [24]

    to flip randomly any two bits of current population by ensuring that one of the flipped bit is 0

    and another one is 1. As count of 1 is unchanged while performing bit string mutation, there

    will be n entries of 1s in newly generated population placed at random places in PopArr.In the preceding example, fitness score for current population is 63dB which is greater than

    MinFit= 54dB. Thus modifiedMinFitis 63dB and Final as {1, 1, 0, 0, 0, 1, 0, 0, 1, 0, 1, 1}.Step 3.3.4: Repeat step 3.3.3 until kth population orMinFit=MaxFit.Step 3.3.5: We say that a frame ofVR is present into VD if that index in Final is 1 i.e. Final[i]

    = 1 for all i from 1 to m. All i from 1 to m will be called as dropped frame indices of VR if

    Final[i] = 0.

    4. Simulation and Analysis

    4.1. Analysis

    The proposed solution approach for identification of dropped frames in previous sectionmainly performs three operations: (a) Creation ofDiffMat (b) Identification of LIS (c) Creation

    of population list and generation of next population using mutation operator if required.To create DiffMat the required computation is of the order O((m-n+1) n) whereas to

    identify LIS in Index field ofDiffMat, it is of the orderO(n log n). In best case if there is no or

    nominal frame drop (i.e.mn), n computational steps are required to create DiffMat, whereas

    in worst case if almost all frames are dropped (i.e.n 1), m2 computations are required.Length of identified LIS will be approximately n for the problem of nature mentioned in

    cases C2.1 and C 2.2. It will be n if all the video frames of VR are unique (i.e. no repeated

    frames in VR) and there is no SD (which is case C2.1) i.e. there will be single appearance ofPSNR as 100dB in each row of DiffMat. It will be approximately n if some of the video frames

    of VR are repeated (i.e. two identical frames in VR) or frames ofVR are spatially distorted with

    low SD. Under worst case, length of LIS will be much less than n.Creation of population list and generation of next population are required only when the

    problem scenario is according to case C3.3.2. Best case for creation of population list will beencountered when length of identified LIS is n-1 (orn). Problems of nature mentioned in case

    C2.3 (where along with frame drop, frames ofVR are spatially distorted with high SD) there will

    be more possibility of getting the scenario like in case C3.3.2.

    4.2. Simulation

    We simulated the scheme presented in section 3 and analysed the performance in terms ofpercentage accuracy for the cases mentioned in section 2. As a video stream may containrepeated frames, cases mentioned in section 2 (C2.1, C2.2, and C2.3) have been extended by

    introducing a new scenario of repeated frames to each case. Updated cases are listed insubsequent paragraphs.

    To analyse the performance, we conducted experiments under following cases. These cases

    are based upon introduced SD (no SD, low SD, and high SD) and repeated frames in each video

    stream.C4.1: There is no SD in frames ofVD and all frames in VR are unique.

    C4.2: There is no SD in frames ofVD and there are (a) 1% or (b) 5% repeated frames in VR.C4.3: Some frames ofVD are distorted by low SD and all frames in VR are unique.

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    567

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    7/12

    C4.4: Low SD in some frames ofVD and there are (a) 1% or (b) 5% repeated frames in VR.

    C4.5: High SD in some frames ofVD and all frames in VR are unique.C4.6: High SD in some frames ofVD and there are (a) 1% or (b) 5% repeated frames in VR.

    To conduct experiments in all listed cases, we required three types of reference videostream; (a) set of video stream in which video frames of each video stream are unique, (b) set of

    video stream in which 1% frames are repeated, and (c) set of video stream in which 5% framesare repeated.

    We used 12 video streams (figure 8) to conduct these experiments. These video streams(bus, coastguard, galleon, football, Stefan, container, garden, hall_monitor,

    tt, Akiyo, carphone, and foreman) are publicly available at

    http://media.xiph.org/video/derf/ in compressed format. We converted all video streams into

    raw video streams and conducted experiments with these raw videos.

    Figure 8. Videos (http://media.xiph.org/video/derf/) used as test video streams

    As we do not have the information about repeated frames in all listed video streams, we

    assumed that video frames in each video stream are unique. Therefore to conduct experiments in

    mentioned cases we intentionally dropped some (1% or 5%) video frames and inserted repeatedframe such that length of video stream is unchanged. Now for our experiments we have three

    sets of reference video stream, one is the original video stream with n unique frames (figure 8),

    second one is video stream with 1% repeated frames and last one is with 5% repeated frames.

    To conduct experiments, we also created set of distorted video stream from these referencevideo streams. We created distorted streams in two phases; first we introduced SD (low or high)

    and then introduced TD by dropping video frames.We used least significant bit (LSB) watermarking scheme [25] to introduce SD into frames of

    a video stream. We embedded watermark at (a) LSB and (b) 6 th LSB bit of a pixel in a videoframe. When watermark is embedded into LSB, there will be least change in pixel bit of a video

    frame, thus we called it as low SD, whereas embedding watermark at sixth LSB will be resulted

    in major change (10 to 25 %) in pixel bit of a video frame, thus we called it as high SD.

    Finally we randomly dropped frames (1% to 5%) of these spatially distorted video streamssuch that we have set of distorted video streams VD having lesser frames. Now it is required to

    identify all the dropped frame indices.

    Performance of proposed algorithm has been analysed in terms of percentage accuracy to

    identify the dropped frames in each case (C4.1 to C4.6). Percentage accuracy has beencomputed according to equation 1 for each case (C4.1 to C4.6) and each percentage drop of

    frames (1% to 5%).

    arden hall monitor

    tt Akiyo car hone foreman

    us coast uard alleon

    Stefan container

    football

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    568

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    8/12

    n

    i i

    i

    b

    a

    1

    100 (1)

    where, i is the experiment # of the conducted experiment for that case ( C 4.1 to C4.6) with

    m% (1% to 5%) drop of frame, a is the count of successfully identified dropped frames indices,and b is the count of dropped frames.

    Obtained accuracy while identifying the dropped frame indices for each video stream under

    different experiment cases have been analysed using following cases:A4.1 : Analysis based upon repeated frames (a) No repeated frames, (b) 1% repeated frames,

    and (c) 5% repeated frames.A4.2 : Analysis based upon SD in video frames (a) No SD, (b) Low SD, and (c) High SD

    We analysed the performance of each video stream under above categories (figure 11 and

    figure 12) but as an example we separately present next the simulation steps and analysis of ourapproach for video stream bus (figure 9 and figure 10).

    There are 126 video frames in uncompressed video bus. We assumed all frames in bus as

    unique. As discussed, we need three sets of reference video of each video stream, therefore wecreated two more video streams bus1 and bus2 (having 126 frames) from bus.

    To create bus1, we introduced 1% repeated frames (2 frames) in bus by dropping 2frames and inserting 2 repeated frames. Similarly bus2 is created by introducing 5% repeated

    frames (7 frames) and dropping 7 random frames in bus. Video stream bus has been used as

    reference video stream in cases C4.1, C4.3, and C4.5 whereas bus1 in cases C4.2a, C4.4a, andC4.6a and bus2 for cases C4.2b, C4.4b, and C4.6b.

    To create distorted video streams, first we introduced SD by embedding watermark (at LSB

    and 6th LSB pixel bits) into video frames of each reference video stream (bus, bus1, and

    bus2) and then introduced TD. After inserting SD, we have in total 9 video streams of bus(three each spatially distorted video stream of bus, bus1, and bus2).

    We introduced TD by dropping random video frames 1 to 5% (2 to 7 frames) of each

    spatially distorted video streams of bus. For each percentage of dropped frames in a particularcase (C4.1 to C4.6) we created 5 distorted video streams. i.e. For drop of 1% frames in case

    C4.1, we created 5 distorted video streams, similarly for drop of 2% frames in case C4.1, we

    have created 5 distorted video streams and so on.Altogether, to conduct experiments for case C4.1 with reference video stream bus we have

    25 distorted video streams (5 distorted videos for each percentage of dropped frames). Similarly

    we have 25 distorted video streams of bus in each cases C4.2a, C4.2b etc.

    (a) (a)

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    569

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    9/12

    Figure 9. Analysis of video stream bus having

    (a) No, (b) 1%, and (c) 5% repeated frames

    Figure 10. Analysis of video stream bus

    having (a) No, (b) Low, and (c) High SD

    (b)

    (c)

    (b)

    (c)

    (a) (a)

    (b) (b)

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    570

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    10/12

    Figure 11. Analysis of video streams (includingbus) having (a) No, (b) 1%, and (c) 5%

    repeated frames

    Figure 12. Analysis of video streams (includingbus) having (a) No, (b) Low, and (c) High SD

    After creating reference and distorted video streams, we computed drop frame indices usingour proposed algorithm (section 3). It is simulated for maximum 100 generation of population

    (ifMinFit = MaxFit, next population is not generated) using mutation operator. Subsequent

    paragraphs present performance analysis of proposed scheme for video stream bus:Our approach performs best when there are no or low SD and all frames are unique in bus

    (figure 9a, figure 10a, and figure 10b). Here (C4.1 and C4.3) our algorithm successfullyidentified all dropped frame indices. Similar observations have been made for other video

    streams (figure 11a, figure 12a, and figure 12b), which infers that if there are no repeated

    frames with no or low SD, our algorithm successfully identify all dropped frame indices. Butobvious reason behind this is the single matching index for each frame in distorted video stream

    with respect to reference video stream and length of LIS is equal to length of distorted video.

    As observed from figure 9c and figure 10c, our algorithms performance is worst when framesare distorted with high SD and there are many repeated frames in reference video stream.

    Similar observations have been made for other video streams in figure 11c and figure 12c. Butobvious reason behind this is the conflicts regarding selection of index due to repeated frames.

    As observed from figure 9, our algorithm performance decreases when number of repeated

    frames increases in video stream bus. Similarly, we observed (figure 10) that our algorithmperformance decreases when SD into frames of video stream bus increases. Similar

    observations have been made in figure 11 and figure 12 for other video streams.

    4.3. Comparison with Other Schemes

    As discussed in section 2, some of the contemporary works are being presented by S. Wolf;

    MSUs dropped frame metric; and Nishikawa K et. al. Wolfs [11] method is based upon

    identification of dropped frames when there is no or reduced information about reference videostream therefore SD in video frames are unhandled. Whereas we presented a full reference

    algorithm and it is able to efficiently identify the dropped frames even though frames are

    spatially distorted with low or high SD which might occur during processing over video frames.

    Comparing a frame with its adjacent frame is the main step of MSUs [12] dropped frame

    metric or Nishikawa K et. al [13] which might restrict the approach to explore other frame in avideo stream for the similarity. As presented in section 3, our algorithm explores the

    possibilities of dropped frame indices into entire video stream and therefore giving an edge overprevious schemes to efficiently (section 4.2) identify the dropped frame indices into a distorted

    video stream with respect to reference video stream.

    5. Conclusion

    In this paper we have presented a full reference algorithm which efficiently identifies the

    dropped frame indices in distorted video stream with respect to reference video stream. Analysis

    section shows that the minimum computations required by our algorithm is n + n log n, i.e.n

    (c) (c)

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    571

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    11/12

    computations for creatingDiffMat (when there is nominal frame drop) and n log n computations

    to identify LIS. In worst case, the computations required by our algorithm is m2 + n log n + k,i.e. m

    2 computations for creating DiffMat (when almost all frames are dropped), n log n

    computations to identify LIS, and kis the required computational steps (if required) to generatekpopulation by mutation operator along with its fitness score computation. If length of LIS is

    much less than n, then to identify dropped frames it is required to explore all combination suchthat best match can be selected which is an exponential time problem and requires heavy

    computation. Use of genetic operator reduces it to k steps by compromising a little bit inaccuracy (section 4.2) to identify the dropped frame indices. Simulation results based upon 50

    experiments in each case show that the proposed algorithm successfully identifies all the

    dropped frame indices when there is no or low SD in distorted video stream and all frames in

    reference video stream are unique, whereas its accuracy is approximately in between 76% to90% (worst obtained accuracy) for the cases where there is high SD in distorted video stream

    and many repeated frames in reference video stream.

    6. References

    [1] Moorthy A K, Seshadrinathan K, Soundararajan R, Bovik A C, Wireless video qualityassessment: A study of Subjective scores and Objective algorithms, IEEE Transactions onCircuits and Systems for Video Technology, vol. 20, no. 4, pp. 587-599, 2010.

    [2] Seshadrinathan K, Soundararajan R, Bovik A C, Cormack L K, Study of Subjective andObjective quality assessment of video, IEEE Transactions on Image Processing, vol. 19, no. 6, pp.1427-1441, 2010.

    [3] Seshadrinathan K, Bovik A C, Motion tuned spatio-temporal quality assessment of naturalvideos, IEEE Transactions on Image Processing, vol. 19, no. 2, pp. 335-350, 2010.

    [4] Chikkerur S, Sundaram V, Reisslein M, Karam L J, Objective Video Quality AssessmentMethods: A Classification, Review, and Performance Comparison, IEEE Transactions onBroadcasting, vol. 57, no. 2, pp. 165-182, 2011.

    [5] http://www.cyberlawsindia.net/index1.html Accessed 26 April 2012.[6] http://cyber.law.harvard.edu/metaschool/fisher/integrity/Links/Articles/winick.html Accessed 26

    April 2012.[7] Li Y, Gao X, Ji H, A 3D Wavelet Based Spatial-Temporal Approach for Video Watermarking,

    in Proceedings of Fifth International Conference on Computational Intelligence and Multimedia

    Applications, ICCIMA03, 2003.[8] Gulliver S R, Ghinea G, The Perceptual and Attentive Impact of Delay and Jitter in Multimedia

    Delivery, IEEE Transactions on Broadcasting, vol. 53, no. 2, pp. 449-458, 2007.

    [9] Yim C, Bovik A C, Evaluation of temporal variation of video quality in packet loss networks,Signal Processing: Image Communication 26 (2011), pp. 24-38, 2011.

    [10] Pinson M H, Wolf S, and Cermak G, HDTV Subjective Quality of H.264 vs. MPEG-2, With andWithout Packet Loss, IEEE Transactions on Broadcasting, vol. 56, no. 1, pp. 86-91, 2010.

    [11] Wolf S, A No Reference (NR) and Reduced Reference (RR) Metric for Detecting Dropped VideoFrames, Fourth International Workshop on Video Processing and Quality Metrics for ConsumerElectronics, VPQM 2009.

    [12] http://compression.ru/video/quality_measure/metric_plugins/dfm_en.htm Accessed 26 April 2012.[13]Nishikawa K, Munadi K, Kiya H, No-Reference PSNR Estimation for Quality Monitoring of

    Motion JPEG2000 Video Over Lossy Packet Networks, IEEE Transactions on Multimedi, vol. 10,no. 4, pp. 637-645, 2008.

    [14] Winkler S, Mohandas P, The Evolution of Video Quality Measurement: From PSNR to HybridMetrics, IEEE Transactions on Broadcasting, vol. 54, no. 3, pp. 1-9, 2008.

    [15] Wang Z, Sheikh H R, Bovik A C, Objective video quality assessment, in The Handbook of VideoDatabases: Design and Applications, B. Furht and O. Marques, ed., CRC Press, pp 1041-1078,2003.

    [16] Girod B, Whats wrong with mean-squared error, in Digital Images and Human Vision. A. B.Watson, ed., MIT Press, pp. 207220, 1993.

    [17] Yuanjiang LI and Yuehua LI, "Passive Millimeter-wave Image Denoising Based on ImprovedAlgorithm of Non-local mean", IJACT, vol. 4, no. 10, pp. 158-164, 2012.

    [18] H. Duan and G. Chen, "A New Digital Halftoning Algorithm by Integrating Modified Pulse-Coupled Neural Network with Random Number Generator", JDCTA, vol. 6, no. 12, pp. 29-37,2012.

    A Full Reference Algorithm for Dropped Frames Identification in

    Uncompressed Video Using Genetic Algorithm

    Manish K Thakur, Vikas Saxena, J P Gupta

    572

  • 7/28/2019 PaperA Full Reference Algorithm for Dropped Frames Identification in Uncompressed Video Using Genetic Algorithm

    12/12