KR20160032761A - The method for recovery of multimedia piece file - Google Patents

Abstract

The present invention relates to a method for recovering video or audio fragment files cut by a bit torrent or the like, and an apparatus using the same. More particularly, the method of the present invention firstly identifies a file format by a feature pattern of the fragment files and puts a maximum likelihood header corresponding to the identified format in the front of the fragment files for decoding in order to recover the cut video or audio fragment files.

Description

[0001] The present invention relates to a multimedia piece file recovery method,

The present invention relates to a method for restoring a video or audio fragment file that is cut by a video sharing program such as a bit torrent, and more particularly, to a method for restoring a cut video or audio fragment file by identifying a file format And attaching a maximum similarity header corresponding to the identified format to a fragment file before decoding.

With the development of computer and internet technologies, the amount of multimedia data is increasing and the distribution method is also becoming diverse. In recent years, there have been a lot of distributions using a file sharing system represented by BitTorrent, which is a method of cutting and sharing files in pieces. Most of the video files are distributed through AVI, MKV, MP4, RMVB, And most of the audio files are distributed through MP3, AAC, FLAC, and other formats.

BitTorrent is the name of the peer-to-peer file transfer protocol and the name of the application software that uses it. With BitTorrent, files are distributed on the Internet and stored, and files are imported simultaneously from multiple locations using multiple connections, thus speeding up the transfer. The BitTorrent protocol was invented by programmer Bram Cohen, and since April 2001, the first implementation was announced on July 2, 2001. Currently, BitTorrent Inc., a company founded by Cohen, Inc. Is maintaining this. Original BitTorrent programs are written in Python using distributed hash tables and are distributed under open source copyright. In addition, various bit-torrent clients compatible with the BitTorrent protocol are available.

Of these video files distributed in BitTorrent, the AVI (Audio Video Interleave) video file format is a multimedia container format for Windows operating system developed by Microsoft. The AVI file format is a Resource Interchange File Format (RIFF) file specification used for applications that capture, edit, and play audio / video sequences.

The MKV file format conforms to the open standard Matroska Multimedia Container format. These file formats can contain an unlimited number of video, audio, picture and subtitle tracks in one file. The MKV file structure is constructed on the basis of Extensible Binary Meta Language (EBML), a data container method based on XML.

MP4 (MPEG-4 Part 14) is a multimedia container format standard, commonly used for storing digital video and audio streams, and can also be used to store caption and other data such as still images. MP4 files are composed of boxes or atoms.

RealMedia Variable Bitrate (RMVB) is a variable bit rate extension of the Real Media Multimedia Container format developed by RealNetworks and is commonly used for multimedia content for storage. RMVB accounts for the highest percentage of illegal distribution video file formats.

MP3, the MPEG-1 audio compression standard, is one of the most widely used audio compression standards and accounts for the highest percentage of audio files distributed in BitTorrent. MP3 is a kind of lossy compression method, which uses a compression method to discard parts of audio data that are not important for human listening.

Advanced Audio Coding (AAC) also compresses audio data with a lossy compression method. AAC is a successful successor to MP3, an audio format used in Apple's iTunes, iPod, and iPhone. AAC provides good sound quality even for small-sized audio files and has high sound resolution.

FLAC (Free Lossless Audio Codec) is a file format that compresses audio data without loss. In other words, unlike MP3, AAC and other formats, there is no loss of audio data. FLAC can save storage space, bandwidth, transfer rate and so on. FLAC compresses and decodes audio data to reduce it to 50% of its original size, so that it can be accurately preserved in its original form.

In many studies on existing file format classification and fragment file identification, the main analysis formats are document format, image format, compression format, execution and library file formats.

However, since copyright infringement occurs mainly in illegal distribution of video or audio files, there is a desperate method of classifying professional video or audio file formats, identifying fragmented file formats, and recovering stream of fragmented files. However, There are various difficulties in recovering due to lack of information.

Histogram, Mean, Hamming weight, N-gram, Compressed length, Shannon entropy, Standard deviation, etc. are used for classification and identification of existing formats. However, video or audio streams of all formats have strong random entropy compression and stream data distribution all have similar random properties, so these features do not have a good effect on video or audio fragment file identification. In addition, when analyzing fragmented files distributed in bit torrents, it takes a long time to download file headers and tail, so it is also not efficient to use headers and tail features.

[Patent Document 1] Korean Patent Registration No. 10-1407166 "Method for restoring a database in a digital video recorder ", 2014.06.05. [Patent Document 2] Korean Patent Laid-Open Publication No. 10-2007-0093389 "Recovery of Real Time Video Data After Power Loss", 2007.09.18. [Patent Literature 3] Korean Patent Registration No. 10-0470844 "Data Recording Method and Data Recovery Method ", January 31, 2005.

In order to solve the above problem, in the present invention, a unique sync pattern of each video or audio file format is used as a feature to classify a format and identify a fragmented file.

First, the cut video or audio fragment file does not have the header information necessary to reproduce. Therefore, in order to recover it, first, a header analyzed by statistical method is added before the fragment file, and the fragment file format is first identified to add a header of the correct format I want to provide a way to recover fragmented files.

To this end, a method of classifying a video or audio file format in order to improve the identification accuracy of the fragmented file format and the speed of the identification calculation is provided.

The present invention also provides a method for recovering a video or audio stream in a fragmented file by adding a statistically analyzed maximum similarity header before a truncated video or audio fragment file for the purpose of recovering a video or audio fragment file.

In order to solve the above problems, a file format classification DB in which a class distribution for classifying a file format is stored and a maximum similarity header DB in which a maximum similarity header is stored are constructed, and a system including a video processing unit, a voice processing unit, A method for recovering video or audio fragment files on a network, comprising the steps of: (A) collecting fragmented files on a network; (B) analyzing a class distribution of the file format classification DB to identify a file format of the fragmented file; (C) extracting a maximum similarity header corresponding to the file format of the fragmented file from the maximum similarity header DB and combining the extracted maximum similarity header at the end of the fragmented file; (D) receiving and decoding a fragmented file in which the maximum similarity header is combined from the central processing unit by the image processing unit or the audio processing unit.

The present invention has the following effects.

1. The present invention identifies the format of a video or audio fragment file and restores it by decompressing it according to the format. Therefore, it is possible to reconstruct a new header by adding insufficient information to an incomplete video or audio file, Not only does it improve the accuracy of recovering video or audio data from fragmented files, it also improves recovery speed.

2. Also, the present invention has the effect of shortening the time required for downloading and restoring video or audio whole fragment files distributed on-line, such as a bittorrent network.

Particularly, the present invention is advantageous in that it is difficult to recover video or audio data when a video or audio fragment file downloaded by a bittorrent client is deleted or insufficient.

[Fig. 1] is a structure diagram of an AVI among video file formats.
[Fig. 2] is a structural view of MKV in a video file format.
[Fig. 3] is a structure diagram of MP4 in a video file format.
4 is a structural view of RMVB in a video file format.
[Fig. 5] is a structure diagram of MP3 in an audio file format.
6 is a structure diagram of an AAC among audio file formats.
[Fig. 7] is a structural diagram of FLAC among audio file formats.
8 is a flowchart of a sync pattern counting procedure.
9 is a flowchart of a process of classifying a video or audio file format and constructing a file format classification DB.
10 is a flowchart of a process of identifying a format of a video or audio fragment file.
11 is a flowchart of a process of configuring a maximum similarity header.
12 is a flowchart of a process of recovering a fragmented file of video or audio.

The present invention includes a file format classification DB in which a class distribution for classifying a file format is stored and a maximum similarity header DB in which a maximum similarity header is stored. The system includes a video processing unit, a voice processing unit, A method for recovering a fragmented file, comprising the steps of: (A) collecting fragmented files on a network; (B) analyzing a class distribution of the file format classification DB to identify a file format of the fragmented file; (C) extracting a maximum similarity header corresponding to the file format of the fragmented file from the maximum similarity header DB and combining the extracted maximum similarity header at the end of the fragmented file; And (D) receiving and decoding a fragmented file in which the maximum similarity header is combined from the central processing unit or the image processing unit or the audio processing unit; The method comprising the steps of:

The central processing unit may further include: (1) cutting out data having a length L ₁ randomly in a range of data excluding a header and a tail in the collected N video or audio sample files, (2) generating a sync pattern counter value for each file format according to a sync pattern counting procedure for each cut-out data; (3) constructing a high dimensional vector set by combining sync pattern counter values of each data generated for each file format; And (4) storing the generated N high-dimensional vector sets into M file formats using a cluster algorithm or SVM (Support Vector Machine) to store M class distributions in the file format classification DB And a file format classification DB is constructed.

According to another aspect of the present invention, in the step (B), the central processing unit may include: (B-1) generating a sync pattern counter value included in the fragment file according to a sync pattern counting procedure; (B-2) constructing a set of high dimensional vectors from sync pattern counter values corresponding to each format; (B-3) generating a reduced high-dimensional vector set by performing vector normalization on the high-dimensional vector set to fit the size of the fragmented file to L ₁ ; And (B-4) comparing the class distribution of the reduced high-dimensional vector set and the file format classification DB with a KNN technique (K-Nearest Neighbor technique, a technique for measuring a minimum euclidean distance) And storing the selected fragmented file in a format of the fragmented file, wherein the format of the fragmented file is identified.

The central processing unit may further comprise: (a) collecting video or audio file headers distributed in a bit-torrent network; (B) rearranging the header according to a probability value by calculating a probability of a video or audio attribute appearing in the header by a statistical method; And (c) storing a header having the highest probability value in the maximum similarity header DB as a maximum similarity header; A maximum similarity header DB is constructed by a process including a step of extracting a maximum similarity header DB.

In addition, the present invention provides a method for recovering a multimedia fragment file, wherein the clustering algorithm is an Optimized Discriminative Subspace Clustering (ODiSC) algorithm.

In the present specification, general terms that are widely used at present are used, but some terms arbitrarily set by the applicant have been used. For arbitrarily chosen terms, the meaning may be interpreted as described in the Detailed Description section of the invention.

Hereinafter, the present invention will be described with reference to specific examples, but the present invention is not limited to or limited by the following examples.

A file format classification DB in which a class distribution for classifying a file format is stored and a maximum similarity header DB in which a maximum similarity header is stored are constructed and a video fragment file on the network is recovered through a system including an image processing unit and a central operation unit (B) analyzing a class distribution of the file format classification DB to identify a file format of the fragmented file; (C) Extracting a maximum similarity header corresponding to a file format of the fragmented file from the maximum similarity header DB and combining the extracted maximum similarity header at the end of the fragmented file; And (D) receiving and decoding the fragmented file in which the maximum similarity header is combined from the central processing unit.

The present invention relates to a method for recovering video data from a video fragment file distributed in a network. In order to accomplish the present invention, a file format is classified from a plurality of sets of high dimensional vectors extracted from a plurality of sample files, It is preferable that a maximum similarity header DB in which the stored file format classification DB and the maximum similarity header are stored is built in advance. It is preferable that the central processing unit is constructed in a system including a central operation unit and an image processing unit. The central operation unit constructs the file format classification DB and the maximum similarity header DB and becomes a central axis in the process of restoring the fragmented file to be restored . The image processing unit or the image processing unit is responsible for decoding the fragmented file in which the maximum similarity header is combined by analyzing in the central operation unit. The central operation unit and the image processing unit are not conceptual or relative.

Such a video and audio restoration method can be confirmed in the flowchart of the process of restoring video and audio according to the present invention of FIG. 12, when the format identification 1220 of the input piece file 1210 ends, the maximum similar header 1250 of the corresponding format is added to the beginning of the fragment file to obtain the reconstructed fragment file, (1230), and inputting the reconstructed fragment file to the video and audio decoder to reproduce video and audio (1240).

First, steps (A) and (C) are performed in the central operation unit, and step (D) is decoded in the image processing unit.

In step (A), the central operation unit collects fragmented files on the network. A video sculpture file distributed on the network such as a bit torrent is generated by dividing a video file corresponding to the original into a plurality of sculpture files and generating a plurality of sculpting files and downloading the sculpting files for each sculpting file. However, Decoding may be difficult until the header is down because the format is not identified. Accordingly, it is a process of preparing for decoding and identifying the format of the fragmented files to be downloaded, respectively.

In the step (B), the central operating unit analyzes the class distribution of the file format classification DB to identify the file format of the fragmented file. In order to decode the fragmented file, it is necessary to confirm what the file format of the fragmented file is.

To this end, the file format classification DB is applied. First, M classifications are classified into file format classification DBs for each file format, and a plurality of sets of high dimensional vectors classified and stored in each class are stored. And the comparison and analysis with the high dimensional vector set is performed.

First, since there are many video files or audio files on the network, that is, multimedia file formats, video files through avi, mkv, mp3, RMVB, and audio files through MP3, AAC and FLAC are mostly distributed. The embodiment will be described.

1 is a structural diagram of an AVI among video file formats to which the present invention is applied. Referring to FIG. 1, an AVI file is composed of two data containers, chunk and list. A list can contain another list or chunk, and a chunk can contain header, video, or audio data. RIFF AVI has file size information, hdrl has size information of LIST hdrl, avih mainly includes time length of one frame, total number of frames, number of streams, video width and video height, There is size information of strl. If the fccType of strh is 'vids', it specifies the information of the video stream and if it is 'auds' it specifies the information of the audio stream. strh mainly includes a codec ID, a frame rate, a sample rate, a start position of a frame, a length of a stream, a quality parameter used for decoding a stream, and a sample size. If strh contains video information, strf is composed of a BITMAPINFOHEADER structure, whereas if strh contains audio information, strf is composed of a WAVEFORMATEX structure.

The BITMAPINFOHEADER structure mainly contains the structure size, the number of bitmap width pixels, the number of bitmap height pixels, the number of bits per pixel, the compression method, and the image size. The WAVEFORMATEX structure mainly includes the audio format type, the number of channels included in the audio data, the sample rate, the average data transmission rate, the block alignment, and the number of bits per sample.

LIST Stores the video or audio stream data in chunks in movi, and idxl stores the start position and length of each chunk. The video data chunk begins with 4 bytes of FourCC (Four Character Code), ie, 30 30 64 63 in hexadecimal, and the audio data chunk begins with 30 31 77 62. FourCC follows the size of the next 4-byte chunk data. In any video file, the audio and video streams are stored alternately. If you set FourCC to sync, the AVI file has a sync pattern that indicates that there is another sync at that position if the 4 bytes value after the sync, 30 30 64 63 or 30 31 77 62, is moved by the specified length.

2 is a structure diagram of an MKV of a video file format to which the present invention is applied. Referring to FIG. 2, an MKV file can be divided into two parts: a header and a segment, and a segment can be further divided into eight parts, so that the MKV file has nine parts in total. Header mainly includes version information of EBML, file version information, ID maximum length and maximum length. Meta Seek Information stores the ID and location information of other EMBL elements to access data more quickly. Segment information mainly contains information for identifying a file (Title, Segment unique ID, etc.) and the duration of the segment. Tracks mainly include audio and video codec information, video width and height, audio sample rate, and number of channels. Chapters contain all editions and chapter lists. One edition has one set of chapters. If there is a chapter, the video can be played one chapter at a time. Cueing data mainly contains time and position information of key frames. Attachment contains CD cover photo, caption font, or other attachment information. Tagging contains additional information for Tracks, Chapters or Albums. Clusters contain audio and video data in BlockGroup or SimpleBlock units. The difference between Block and SimpleBlock in BlockGroup is that the keyframe flag and the unnecessary frame flag are only in SimpleBlock and not in Block of BlockGroup. The EBML ID of the BlockGroup begins with hex A0 followed by the size, the EBML ID of SimpleBlock begins with A3 followed by the size. If you set the BlockGroup and SimpleBlock to sync, the MKV file has a sync pattern where there is another sync at that position when the size of the sync after A0 or A3 moves by the specified length.

3 is a structural diagram of MP4 among video file formats to which the present invention is applied. Referring to FIG. 3, an MP4 file has been formed in various box (or atom) structures. ftyp is a file type box that verifies file compatibility. moov, trak, mdia, minf, stbl etc. Boxes are containers only, but there is no actual data. moov is a box that contains all the metadata boxes for the media. mvhd usually contains the time the file was created, the time it was modified, and the duration. Trak contains all the meta information boxes of a particular media, audio or video. tkhd mainly contains the time, modified time and duration of the track. mdhd also contains the time that media was created, the time it was modified, and so on. hdlr specifies the type of trak. stsd contains the codec information needed to decode the track, stts contains the time information of the samples, stsc contains the number of samples in each chunk, stsz contains the size of the samples in each chunk, stss contains It contains the key frame sample number, and stco contains the location information of each chunk. mdat is a data box that stores actual media data in units of samples. The video data samples start with a size of 4 bytes and the next two bytes are set to sync. We obtained 11 sync values which appeared frequently in statistical methods. Since the audio data samples start with 21 in hexadecimal, this is also set to sync. Therefore, the MP4 file has a sync pattern that if one of the 11 video samples sync has a 4-byte value moved by the specified length, there will be one of 12 sync at that position.

4 is a structural view of RMVB in a video file format to which the present invention is applied. Referring to FIG. 4, the RMVB file is composed of six chunks including RMF, PROP, MDPR, CONT, DATA, and INDX. The RMF chunk usually specifies the number of headers, and the PROP chunk contains the bit rate, the data packet size, the number of data packets, the duration of the file, the beginning of the first INDX chunk, the beginning of the first DATA chunk, And the MDPR chunk mainly includes the start time of the stream, the duration of the stream, the codec information, the frame size, the sample rate, and the sample size. The CONT chunk includes the title, The audio data is stored, and the INDX chunk includes the stream number, the position of the next INDX chunk, and the start point of the packet. All data packets start with a 2-byte packet version followed by a 2-byte packet size. There are two packet versions, 00 00 or 00 01 in hexadecimal. If you set this packet version to sync, the RMVB file has the following sync pattern. That is, if a 2 byte value after the sync of 00 00 or 00 01 is moved by the specified length, there is another sync at that position.

The file formats of audio files existing on the following network are mostly the distribution of audio files through MP3, AAC, FLAC, etc. Therefore, these three file formats will be described.

5 is a structural diagram of each parameter in a frame header of an MP3 file to which the present invention is applied. Referring to FIG. 5, a frame header includes 12 sync bits, a version ID, a layer, a CRC protection bit, a bit rate, a frequency, a padding bit, a private bit, a channel mode, And the like. Among them, version ID, layer, bit rate, frequency, padding bit, and channel mode are the only six parameters necessary for audio data recovery. The length of the MP3 frame is calculated using the following three parameters: bit rate, frequency, and padding bit.

Therefore, when 12 sync bits 111111111111 are set to sync, an MP3 file has the following sync pattern. That is, if a value is extracted at a position corresponding to a bit rate, a frequency, and a padding bit after 111111111111, the frame length is calculated by substituting the value into the formula, and if the length is shifted by that length, there is another sync at that position.

6 is a structural diagram of each parameter in a frame header of an AAC file to which the present invention is applied. The frame header of the AAC file has an ADTS (Audio Data Transport Stream) structure. 6, the frame header includes 12 synchronization bits, a version ID, a layer, a CRC protection bit, a profile, a sampling frequency, a private bit, a channel, an original bit, a home, a copyright protection bit, , Number of AAC frames, and CRC.

Therefore, when 12 sync bits 111111111111 are set to sync, the AAC file has the following sync pattern. In other words, if you extract a number where 111111111111 back frame length parameter is occupied and then move that length, there is another sync at that position.

7 is a structural diagram of a FLAC file to which the present invention is applied. Referring to FIG. 7, the FLAC file stream begins with fLaC followed by one or more metadata blocks and one or more frames. The metadata block contains the header and the data. The metadata block header is followed by a metadata block number, followed by the corresponding metadata block. The information in the metadata block includes the minimum or maximum block size, the minimum or maximum frame size, the sample rate, the number of channels, the number of samples per bit, the total number of samples in the stream, and the MD5 signal. The frame includes a frame header, a sub-frame, and a frame footer. The frame header starts with the synchronization bit '11111111111110' and has the most basic information such as the sample rate and the number of channels. Since the FLAC file does not store frame size information separately, it moves to the end of one frame by decoding all parameters and audio data in the header and moving backward.

Therefore, if you set 14 sync bits 11111111111110 to sync, the FLAC file has the following sync pattern. That is, it decodes parameters and audio data after 11111111111110, and when arriving at the end of the frame, there is a sync at the next position.

A method of constructing the file format classification DB according to the present invention will be described with reference to the embodiments of the four video file formats and three audio file formats.

The step (1) is a step of generating N data by cutting out data having a length L ₁ randomly in a range of data excluding a header and a tail in N video or audio sample files collected on the network . Generally, since header and tail are low in probability of dropping out of a plurality of sample files, excluding the header and tail is suitable for recovering a fragmented file in which general data is stored.

In other words, downloading applications such as bit-torrents transmit fragmented files in a random order, and therefore the probability of receiving a video or audio stream fragment file, which occupies most of the fragmented files, is high. Therefore, feature points are extracted from data except for the header and tail will be. In addition, since the method of identifying the file format using the information of the header and the tail is already common, and it is not efficient in identifying the original file format of the fragmented file distributed in the bittorrent, Method.

Step (2) is a step of generating a sync pattern counter value for each file format according to a sync pattern counting procedure for each cut-out data.

8 is a flowchart of a sync pattern counting procedure according to the present invention. Referring to FIG. 8, data of a fragmented file is received (810) and a sync value corresponding to each file format is initialized (820). For example, if the format is RMVB, the sync is 00 00 or 00 01, and in the case of AVI the sync is 30 30 64 63 or 30 31 77 62. All positions where sync appears are retrieved and stored in the offset variable (830). In each offset variable, the size information is moved 840 by a designated length and then it is determined 850 whether the piece reaches the end of the piece, and it is determined whether there is a sync at the position 860. The feature quantity of each file format included in the fragment file is calculated by a method (870) of recording the order in which the sync corresponding to each file format in the fragment file is recorded in the sync_counter. If applied to M file formats, M sync pattern counter values are output (880).

Step (3) is a step of constructing a high dimensional vector set by combining sync pattern counter values of each data generated for each file format.

In order to more efficiently and quickly identify the file format of the video and audio fragmented files distributed on the network, the result of classifying the pattern of each file format in the database is stored in advance. FIG. 9 is a flowchart of a process of building a file format classification DB according to the present invention. Referring to FIG. 9, a header and a tail of each sample are first cut 920 in a sample space 910 composed of N video and audio files. Downloading applications such as BitTorrent send fragments in random order, so extracting feature points from data excluding header and tail is more likely to receive video and audio stream fragments that occupy most of the number of fragments. In addition, the method of identifying the file format using the information of the header and the tail is already common, and it is not effective in identifying the original file format of the fragmented file distributed in the bittorrent.

Video fragment size is typically in 256KB and between 4MB L _I = 1MB of pieces, because the file size using the probability of the best high-cut, 1MB be random in the stream of data is between the audio fragment size is typically 1MB from 128KB L Since the probability of using the fragmented file size with _I = 256 KB is highest, the stream data is randomly truncated 256 KB (930). In the next cut data, each format sync pattern counting procedure 940 is executed to obtain a sync pattern counter value corresponding to M formats, and these counters are combined to construct an M-dimensional vector, that is, a high dimensional vector set (950).

In step (4), the generated N high-dimensional vector sets are classified into M file formats using a cluster algorithm or SVM (Support Vector Machine), and M class distributions are stored in the file format classification DB.

The clustering algorithm may be applied to a video file and the SVM may be applied to an audio file. The SVM is a conventional discriminant analysis method that focuses on the data at the boundary of the group, so its explanation is omitted.

Particularly, the clustering algorithm can classify the classes by Optimized Discriminative Subspace Clustering (ODiSC) algorithm.

In general, it is often not accurate enough to judge which format corresponds to a high-dimensional vector set only. For example, when analyzing four formats (ie M = 4), we analyzed one MP4 fragment file. The AVI sync counter is 10, the RMVB sync counter is 10, the MP4 sync counter is 15 and the MKV sync counter is 15. v = (10, 10, 15, 15). Likewise, when analyzing the second MP4 fragment file, if the result v = (5, 15, 15, 20) is found, it is not accurate to decide which format belongs to only the vector number. And so on.

[Reference Figure 1]

In [Reference figure 1] above, triangles, squares, circles represent vectors for three formats. X is a sculpted file that is downloaded from the Internet and does not know the format. The distance between X and each vector is calculated to identify that it belongs to the nearest format. In fact, X is the same format as a rectangle, but in [Reference Figure 1], it is likely to be mistaken for a format such as a triangle.

Therefore, high dimensional vector sets for a plurality of sample fragment files are generated as shown in [Reference Figure 2] below.

[Reference Figure 2]

[Reference Figure 3]

[Reference Figure 3] is a result of classifying a plurality of sets of high dimensional vectors generated as shown in [Reference Figure 2] using the ODiSC cluster algorithm.

The Optimized Discriminative Subspace Clustering (ODiSC) classifies N M-dimensional vectors into M classes (960). The ODiSC improves existing DiSCs according to the video sync pattern characteristics, It is a cluster algorithm that is further optimized by using the algorithm. Classified class distribution information is stored in the file format classification DB (970) and used to identify the original file format of the fragmented file downloaded later from the bit torrent client.

[Reference Figure 4]

The red triangles, squares, circles in [Figure 4] are the center of gravity. If X is a sculpture file that is downloaded from the Internet and is unknown to the format, X calculates the distance between it and each center of gravity, that is, calculates the distance using KNN and identifies it as belonging to the format of class distribution information of the nearest distance.

A method of identifying a file format by comparing and analyzing with the file format classification DB constructed as described above in the step (B) is as follows.

In step (B-1), a sync pattern counter value included in the fragment file is generated through a sync pattern counter procedure. In step (B-2), a high-dimensional vector set is obtained from a sync pattern counter value corresponding to each format Respectively.

10 is a flowchart of a process of identifying a format of video and audio fragmented files according to the present invention. 10, when the video and audio fragmented files distributed in the bit torrent BT are down (1010), the format sync pattern counting procedure 1020 is executed in the same manner as the video and audio file format classification procedure, A sync pattern counter value corresponding to the format is obtained, and an M-dimensional vector is formed by combining these counter values (1030). If the sync counter value is sc and the number of dimensions of the high dimensional vector is M dimensions, then the high dimensional vector v can be expressed as:

(B-3) is a step of performing vector normalization on the set of higher dimensional vectors to fit the size of the fragment file to data of L _I.

Since the size of the downloaded video fragment file is between 256 KB and 4 MB and the audio fragment file size is between 128 KB and 1 MB, the M dimension vector is normalized (1040) in order to match the fragment file size L _I defined at the time of classification at the beginning.

(B-4) is a method of comparing the KNN with a K-Nearest Neighbor (KNN) method of measuring a minimum euclidean distance between a set of a high-dimensional vector reduced by the vector normalization and a class of a file format classification DB, .

The method for measuring the Euclidian distance dist between two vectors is as follows.

When a class closest to the set of high dimensional vectors of the fragmented file is found in class distribution information 1070 stored in the database using K nearest neighbor (KNN) technique in operation 1050, It determines that the format corresponding to the class is identified in the identified video or audio file format (1060).

The step of extracting the class distribution information corresponding to the high dimensional vector set of the file format classification DB and identifying the format by the step of combining the header corresponding to the file format of the fragmented file, (C).

In step (C), the central operation unit extracts a maximum similar header corresponding to the file format of the fragmented file from the maximum similarity header DB, and combines the maximum similarity header at the end of the fragmented file.

In particular, the maximum similarity header is stored in the maximum similarity header DB, and the method of constructing the maximum similarity header DB is as follows.

Step (a) is a step of collecting video or audio file headers distributed in the network.

11 is a flowchart of a process of configuring a maximum similarity header according to the present invention. Referring to FIG. 11, first, headers of video and audio files distributed on a network are collected (1110). The probability that the video and audio attributes 1120 appear in the header 1120 in the statistical method is calculated 1130 and the headers are rearranged according to the magnitude order of the probability 1140.

The step (b) is a step of calculating the probability of a video or audio attribute appearing in the header by a statistical method and rearranging the header according to the probability value.

이다. 마찬가지로 Frame Rate=β 인 헤더의 확률은

이며, 가령 S개 헤더들의 Frame Rate가 α,β,γ 등 세 가지 값을 갖고 있으며 그 중 확률

가 가장 높으면 Frame Rate=γ인 헤더들을 선택하도록 하여 해당 헤더가 배열의 앞부분으로 이동하도록 한다. 이후 다음에 선택된 헤더 중 파라미터 Stream Length(AVI 포맷인 경우)가 α,β,γ 등 세 가지 값을 갖고 있으며 그 중 확률

가 가장 높으면 Stream Length=γ인 헤더들을 선택하여 배열의 앞부분으로 이동한다. The statistical method for calculating the probability that video and audio attributes appear in the header is as follows. There are several parameters in the header. If the number of headers with a parameter of Frame Rate = α is FR _α , the probability that these headers occupy is

to be. Similarly, the probability of a header with Frame Rate =

For example, if Frame Rate of S headers has three values such as α, β, γ,

The header having the highest Frame Rate =? Is selected so that the corresponding header moves to the beginning of the array. Then, parameter stream length (in case of AVI format) of the next selected header has three values such as α, β and γ,

The headers having Stream Length = y are selected and moved to the beginning of the array.

In this way, the probability of another parameter is calculated, and the last remaining header becomes the maximum similar header, which is the header having the highest probability.

In the step (c), the header having the highest probability value is stored in the maximum similarity header DB as the maximum similarity header, and the maximum similarity header extracted by the above method is stored in the maximum similarity header DB.

For example, since a header having a video attribute of a frame rate of 24 is the most common, it is positioned at the rear of the array because the headers are located at the beginning of the array and have a frame rate of 30 This is repeated to determine the header with the highest probability as the maximum similar header. When M types of video or audio file formats are recruited, M maximum similar headers can be obtained. The maximum similarity headers are stored in the maximum similarity header DB (1150).

In the step (D), the image processing unit or the audio processing unit decodes the fragmented file in which the maximum similarity header is combined. And decoding accuracy of the video and audio restoration method according to the present invention is improved when the fragmented file in which the maximum similarity header is combined is decoded.

Such a video and audio restoration method can be confirmed in the flowchart of the process of restoring video and audio according to the present invention of FIG. 12, when the format identification 1220 of the input piece file 1210 ends, the maximum similarity header 1250 of the corresponding format is added to the beginning of the fragment file to obtain the reconstructed fragment file 1230 ). The next reconstructed fragment file is input to the video and audio decoder to reproduce video and audio (1240).

Hereinafter, a multimedia fragment file recovery method according to the present invention will be described in conjunction with a specific embodiment. However, modifications and variations falling within the scope of the present invention are possible in addition to the above embodiments. Accordingly, the claims of the present invention include modifications and variations that fall within the true scope of the present invention.

none

Claims (5)

A file format classification DB in which a class distribution for classifying a file format is stored, and a maximum similarity header DB in which a maximum similarity header is stored. The system includes a video processing unit, a voice processing unit, and a central operation unit. As a method of recovering,
(A) collecting fragmented files on the network;
(B) analyzing a class distribution of the file format classification DB to identify a file format of the fragmented file;
(C) extracting a maximum similarity header corresponding to the file format of the fragmented file from the maximum similarity header DB and combining the extracted maximum similarity header at the end of the fragmented file; And
(D) receiving and decoding a fragmented file in which the maximum similarity header is combined from the central processing unit, the image processing unit or the audio processing unit;
And restoring the multimedia video fragment file.
The method of claim 1,
Wherein the central operation unit constructs a file format classification DB by a method including the following steps.
(1) randomly cutting out data of a length L _I among data in a range excluding a header and a tail from the collected N video or audio sample files to generate a total of N pieces of data;
(2) generating a sync pattern counter value for each file format according to a sync pattern counting procedure for each cut data;
(3) constructing a high dimensional vector set by combining sync pattern counter values of each data generated for each file format; And
(4) storing the generated N high-dimensional vector sets into M file classifications using a cluster algorithm or SVM (Support Vector Machine) to store M class distributions in the file format classification DB;
The method of claim 1,
In the step (B)
The central operation unit,
(B-1) generating a sync pattern counter value included in the fragment file according to a sync pattern counting procedure;
(B-2) constructing a set of high dimensional vectors from sync pattern counter values corresponding to each format;
(B-3) generating a reduced high-dimensional vector set by performing vector normalization on the set of higher dimensional vectors to fit the size of the fragmented file to L _I ; And
(B-4) Comparing each class distribution of the reduced high-dimensional vector set and the file format classification DB with a KNN technique (K-Nearest Neighbor technique, a technique of measuring a minimum Euclidean distance) Storing the fragmented file in a format of the fragmented file;
Wherein the format of the fragmented file is identified through a process of including the fragmented file.
The method of claim 1,
Wherein the central operation unit constructs the maximum similarity header DB by a method including the following steps.
(A) collecting video or audio file headers distributed in a network;
(B) rearranging the header according to a probability value by calculating a probability of a video or audio attribute appearing in the header by a statistical method; And
(C) storing a header having the highest probability value in the maximum similarity header DB as a maximum similarity header;
3. The method of claim 2,
The clustering algorithm includes:
Wherein the optimization algorithm is an Optimized Discriminative Subspace Clustering (ODiSC) algorithm.

Images