CN104050247A - Method for realizing quick retrieval of mass videos - Google Patents

Abstract

The invention relates to a method for realizing fast retrieval of massive videos, which includes extracting spatial feature vectors for each frame of video images in a video stream in a video library to obtain a video feature sequence; extracting key feature vectors from the spatial feature vectors; The key feature vectors of all video files establish a distributed storage index library; extract the key feature vector set of the video to be retrieved and extract the video index file of the video to be retrieved; according to the video index file of the video to be retrieved in the distributed storage index library Video similarity retrieval and output video retrieval results whose similarity is greater than the system preset value. Using this structure to achieve fast retrieval of massive videos, representative visual words are used instead of key frames to completely represent video information, without a lot of redundancy, and very compact, speeding up the retrieval speed, and having the ability to concurrently process massive data , has a wider range of applications.

Description

A Method for Fast Retrieval of Massive Videos

technical field

The invention relates to the technical field of multimedia information, in particular to the fields of multimedia information retrieval, data mining and video processing, and specifically refers to a method for realizing fast retrieval of massive videos. the

Background technique

With the rapid development of multimedia information technology and the emergence of video sharing websites, the number of Internet videos has grown rapidly, and has increased exponentially. Publishing, sharing and retrieving videos through the Internet has become a way of life for people. In the face of massive multimedia data, how to quickly retrieve the same or similar videos has become a research hotspot in the industry and academia. the

Traditional video retrieval methods based on key frames are mainly inclined to the accuracy of video retrieval, but the computational complexity is extremely high, and it takes several minutes to complete a retrieval task. In the face of massive online videos, traditional video comparison technology is no longer competent. The current Internet-oriented video retrieval technology draws on the core ideas of traditional text search engines, regards video features as visual words, constructs an inverted index of video files, and realizes fast indexing of massive video files. the

Successful matching depends on the richness of the retrieval video and the reference video's own information and the degree to which its own information is expressed and described. Internet-oriented video retrieval methods often do not follow the traditional method when extracting key frames, first segment the shot, and then extract the key frames of the shot, because the location of the key frame is affected by factors such as video frame rate and resolution. Keyframes cannot be extracted stably and reliably. A simpler and feasible method is to sample the video every 1 second as a key frame. In fact, it is equivalent to increasing the sampling frequency. The higher the sampling frequency, the more fully the original information can be expressed, but the greater the amount of calculation will be. By increasing the sampling frequency to increase the degree of information expression, this will lead to both information being over-expressed and resulting in redundancy, and information not being fully expressed, resulting in information loss. However, linear sampling will make the lost information random, because the video information is not expressed linearly. Randomly missing information reduces the accuracy and stability of retrieval. On the other hand, the traditional key frame extraction method generally extracts fewer key frames where the information changes less, and extracts more key frames where the video frame changes too much, resulting in a more compact and complete image. Express information, the degree of which depends on the clustering or segmentation threshold. Retrieval videos and reference videos are often disturbed by various noises, such as video resolution degradation, network packet loss, video frame loss, low frame rate, video insertion, video editing, etc., which will make the original video information mixed with noise, Or cause partial information to be lost and no longer complete. The traditional video key frame extraction method is too idealistic, a) the complexity of external interference is not considered, and appropriate redundancy is necessary, b) the features used to extract key frames are not constructed for massive retrieval tasks, related method is not suitable for directly extracting keyframes. How to select the appropriate retrieval features to minimize the number of frames in the constructed key frame sequence and to express relatively complete and redundant video shot information has become a key problem to be solved for mass data retrieval technology. the

Contents of the invention

The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art, and to provide a method that can realize the use of representative visual words instead of key frames, has neither a large amount of redundancy, but is also very compact, speeds up retrieval, and has massive data concurrent processing capabilities. , A method for realizing fast retrieval of massive videos with wider application scope. the

In order to achieve the above object, the method for realizing the fast retrieval of massive videos of the present invention has the following composition:

The method for realizing the fast retrieval of massive video, its main feature is that described method comprises the following steps:

(1) Extract the spatial feature vectors for each frame of video images in the video stream of the video library to obtain the video feature sequence;

(2) extract key feature vector in the spatial feature vector of described video feature sequence;

(3) set up the distributed storage index library of all video files according to the key feature vector of all video files in the video library;

(4) Extract the key feature vector set of the video to be retrieved and extract the video index file of the video to be retrieved;

(5) Perform video similarity retrieval in the distributed storage index library according to the video index file of the video to be retrieved, and output video retrieval results whose similarity is greater than the system preset value. the

Preferably, the spatial feature vector includes the grayscale spatial distribution feature and the texture spatial distribution feature of the corresponding frame image, and the extraction of the spatial feature vector for each frame video image in the video stream of the video library includes the following steps :

(11) calculate the grayscale image and the edge texture image of each frame video image in the video stream of the video storehouse;

(12) calculate the central space feature and the boundary space feature of the grayscale image of each frame video image and obtain the grayscale space distribution feature of this frame video image formed by the center space feature and the boundary space feature;

(13) Calculate the texture spatial distribution feature of the edge texture image of each frame of video image. the

Preferably, the grayscale image and the edge texture image of each frame video image in the video stream of the video storehouse are obtained by the calculation, comprising the following steps:

(111) divide each frame video image in the video stream of the video storehouse into several sub-images of the same size and calculate the gray value and the number of texture edge points of each sub-image;

(112) calculate the grayscale value of each sub-image of each frame video image to obtain the grayscale image of this frame video image;

(113) Calculate the number of texture edge points of each sub-image of each frame of video image to obtain the edge texture image of the frame of video image. the

More preferably, the central space feature and the boundary space feature of the grayscale image of each frame video image are calculated, specifically:

Calculate the center space feature and boundary space feature of the local binary pattern of the grayscale image of each frame video image;

The texture space distribution feature of the edge texture image of the described calculation each frame video image is specifically:

The texture spatial distribution feature of the local binary pattern of the edge texture image of each frame video image is calculated. the

Preferably, said spatial feature vector also includes a color histogram feature, and said extracting a spatial feature vector for each frame video image in the video stream of the video library also includes the following steps:

(14) Calculate the color histogram feature of each frame of video image. the

Preferably, the extraction of the key feature vector in the spatial feature vector of the video feature sequence comprises the following steps:

(21) the first spatial feature vector of the video feature sequence is defaulted as the key feature vector;

(22) Calculate the Mahalanobis distance between each spatial feature vector and the previous key feature vector;

(23) Extract the spatial feature vector corresponding to the Mahalanobis distance greater than the system preset threshold as the key feature vector. the

Preferably, the described key feature vector of all video files in the video storehouse is set up the distributed storage index storehouse of all video files, comprises the following steps:

(31) establish the subspace projection histogram of the key feature vector in the video feature sequence and record the frequency of occurrence of each key feature vector in the corresponding video;

(32) set up the inverted index file of all video files of video storehouse;

(33) Establish a distributed index database of all video files in the video library. the

Preferably, the subspace projection histogram of the key feature vector in the described establishment of the video feature sequence is specifically:

The key feature vectors in the video feature sequence are projected into the gray subspace, texture subspace and color subspace, and the subspace projection histogram of each key feature vector is obtained. the

Furthermore, the frequency of recording each key feature vector in the corresponding video is specifically:

A feature value representing the frequency of appearance of the key feature vector in the video is recorded in the subspace projection histogram corresponding to each key feature vector. the

Further, the described inverted index file of all video files of building video storehouse, comprises the following steps:

(321) the corresponding key feature vector set of each video file in the statistical video storehouse and constitute the statistical key feature vector storehouse of this video storehouse;

(322) Establishing a document set corresponding to each key feature vector in the statistical key feature vector library having the key feature vector;

(323) sort the documents of the key feature vector set according to the number of key feature vectors contained in them from many to few;

(324) Establish an inverted index file of all video files in the video library according to each subspace. the

Further still, the described distributed index database of all video files of setting up video storehouse, comprises the following steps:

(331) Utilize the local sensitive hash algorithm based on p-stable to map the key feature vectors of each subspace to one-dimensional space;

(332) Based on the Hadoop distributed file system architecture, name_node is used to maintain the hash table and data_node is used to store index data as a distributed index database for all video files. the

More preferably, the video similarity retrieval is performed in the distributed storage index library according to the video index file of the video to be retrieved, specifically:

(51) Computing the intersection of the video subspace projection histogram to be retrieved and each video subspace projection histogram in the video library as the similarity of each video in the video to be retrieved and the video library;

(52) According to the time-space structure consistency of the video to be retrieved and the key feature vectors of each video in the video library, video files that do not meet the time-space structure consistency requirements are eliminated. the

Furthermore, the video retrieval result of the output similarity greater than the system preset value includes the following steps:

(52) Extract each subspace projection histogram of the key feature vector of the video to be retrieved and map each key feature vector into a hash value in each subspace;

(53) select the video file whose similarity in the distributed index database meets the system preset requirements through the inverted index file as output;

(54) Calculate the spatio-temporal structure consistency of the video to be retrieved and the key feature vectors of each video in the video database and output a video file whose similarity with the video to be retrieved is greater than the system preset value. the

Adopting the method for realizing massive video fast retrieval in this invention has the following beneficial effects:

The invention mainly aims at the integrity of video index information and the selection of index features, and proposes a subspace method based on video fingerprints to solve the current fast and robust retrieval problem facing massive data. First of all, this patent adopts a novel key frame extraction method, replaces key frame extraction with key feature vector extraction, and directly replaces key frames with representative visual features, which is equivalent to encoding the original video in the feature space, completely expressing The video information has neither a large amount of redundancy, but also is very compact, and overcomes the current key frame extraction parameter selection problem. Secondly, each visual feature is mapped into a one-dimensional hash value, and the name_node (name node) and data_node (data node) of the appropriate HDFS (Hadoop Distributed File System, Hadoop Distributed File System) are selected according to the range of the hash value of the visual feature. node), which not only speeds up the retrieval speed, but also enables it to have the ability to process massive data concurrently, and has a wider range of applications. the

Description of drawings

FIG. 1 is a flowchart of a method for realizing fast retrieval of massive videos in the present invention. the

FIG. 2 is a flow chart of a method for realizing fast retrieval of massive videos of the present invention applied to a specific embodiment. the

Fig. 3 is a flowchart of mapping a video frame sequence to a video feature sequence in the present invention. the

Fig. 4 is a flow chart of calculating gray-scale spatial distribution characteristics in the present invention. the

Fig. 5 is a flow chart of extracting key feature vectors in the present invention. the

Detailed ways

In order to describe the technical content of the present invention more clearly, further description will be given below in conjunction with specific embodiments. the

The invention discloses a method and system for fast retrieval of massive videos, wherein the method includes: mapping a video frame sequence to a video feature sequence composed of spatial feature vectors, and extracting representative features among them as key feature vectors of the video feature sequence; Map the key feature vector through a hash function, construct a distributed index according to the hash bucket where the mapped hash value is located; calculate the number of the hash bucket where each corresponding hash value is located according to the key feature vector set of the video to be retrieved , extract the video index files corresponding to the features, obtain candidate video files by voting, calculate the similarity between the video to be retrieved and the candidate video files, and output the similarity greater than a certain threshold as the retrieval result. the

As shown in Figure 1, the method for realizing massive video fast retrieval of the present invention comprises the following steps:

(1) Extract the spatial feature vectors for each frame of video images in the video stream of the video library to obtain the video feature sequence;

In a preferred embodiment, the spatial feature vector includes the grayscale spatial distribution feature and the texture spatial distribution feature of the corresponding frame image, therefore, each frame video image in the video stream of the video library is extracted respectively The spatial feature vector obtains the video feature sequence, including the following steps:

(11) calculate the grayscale image and the edge texture image of each frame video image in the video stream of the video storehouse;

In a preferred embodiment, the grayscale image and the edge texture image can be calculated in the following manner, that is, the grayscale image and the edge texture image of each frame of the video image in the video stream of the video library are calculated, including The following steps:

(111) divide each frame video image in the video stream of the video storehouse into several sub-images of the same size and calculate the gray value and the number of texture edge points of each sub-image;

(112) calculate the grayscale value of each sub-image of each frame video image to obtain the grayscale image of this frame video image;

(113) Calculate the number of texture edge points of each sub-image of each frame of video image to obtain the edge texture image of the frame of video image. the

(12) calculate the central space feature and the boundary space feature of the grayscale image of each frame video image and obtain the grayscale space distribution feature of this frame video image formed by the center space feature and the boundary space feature;

The central spatial feature and the boundary spatial feature may be the central spatial feature and the boundary spatial feature based on local binary patterns. the

(13) Calculate the texture spatial distribution feature of the edge texture image of each frame of video image. the

Wherein, the texture spatial distribution feature may be a texture spatial distribution feature based on a local binary mode. the

In a more preferred embodiment, the spatial feature vector can further include color histogram features, so that the spatial feature vector can better represent the video features, that is, each frame video image in the video stream of the video library Extracting spatial feature vectors respectively, also includes the following steps:

(14) Calculate the color histogram feature of each frame of video image. the

(2) extract key feature vector in the spatial feature vector of described video feature sequence;

In a preferred embodiment, extracting key feature vectors includes the following steps:

(21) the first spatial feature vector of the video feature sequence is defaulted as the key feature vector;

(22) Calculate the Mahalanobis distance between each spatial feature vector and the previous key feature vector;

(23) Extract the spatial feature vector corresponding to the Mahalanobis distance greater than the system preset threshold as the key feature vector. the

(3) set up the distributed storage index library of all video files according to the key feature vector of all video files in the video library;

In a preferred embodiment, establishing a distributed storage index library includes the following steps:

(31) establish the subspace projection histogram of the key feature vector in the video feature sequence and record the frequency of occurrence of each key feature vector in the corresponding video;

Furthermore, the subspace can be a grayscale subspace and a texture subspace, and can also include a color subspace, so the subspace projection histogram of the key feature vector in the described establishment of the video feature sequence is specifically:

The key feature vectors in the video feature sequence are projected into the gray subspace, texture subspace and color subspace, and the subspace projection histogram of each key feature vector is obtained. the

Furthermore, the frequency of recording each key feature vector in the corresponding video is specifically:

A feature value representing the frequency of appearance of the key feature vector in the video is recorded in the subspace projection histogram corresponding to each key feature vector. the

(32) set up the inverted index file of all video files of video storehouse;

Further, the described inverted index file of all video files of building video storehouse, comprises the following steps:

(321) the corresponding key feature vector set of each video file in the statistical video storehouse and constitute the statistical key feature vector storehouse of this video storehouse;

(322) set up the document set corresponding to each key feature vector in the statistical key feature vector library with the key feature vector;

(323) sort the documents of the key feature vector set according to the number of key feature vectors contained in them from many to few;

(324) Establish an inverted index file of all video files in the video library according to each subspace. the

(33) Establish a distributed index database of all video files in the video library. the

Further, the described distributed index database of all video files of building video storehouse, comprises the following steps:

(331) Utilize the local sensitive hash algorithm based on p-stable to map the key feature vectors of each subspace to one-dimensional space;

(332) Based on the Hadoop distributed file system architecture, name_node is used to maintain the hash table and data_node is used to store index data as a distributed index database for all video files. the

(4) Extract the key feature vector set of the video to be retrieved and extract the video index file of the video to be retrieved; in specific implementation applications, extract the key feature vector of the video to be retrieved here as steps (1) and (2) The key feature vector extraction method in . the

(5) Perform video similarity retrieval in the distributed storage index library according to the video index file of the video to be retrieved, and output video retrieval results whose similarity is greater than the system preset value. the

In a preferred embodiment, the video similarity retrieval is performed in the distributed storage index library according to the video index file of the video to be retrieved, specifically:

(51) Computing the intersection of the video subspace projection histogram to be retrieved and each video subspace projection histogram in the video library as the similarity of each video in the video to be retrieved and the video library;

(52) According to the time-space structure consistency of the video to be retrieved and the key feature vectors of each video in the video library, video files that do not meet the time-space structure consistency requirements are eliminated. the

In a preferred embodiment, the video retrieval result whose similarity is greater than the system preset value includes the following steps:

(52) Extract each subspace projection histogram of the key feature vector of the video to be retrieved and map each key feature vector into a hash value in each subspace;

(53) select the video file whose similarity in the distributed index database meets the system preset requirements through the inverted index file as output;

(54) Calculate the spatio-temporal structure consistency of the video to be retrieved and the key feature vectors of each video in the video database and output a video file whose similarity with the video to be retrieved is greater than the system preset value. the

The method for realizing massive video fast retrieval of the present invention is further described below with a specific embodiment, as shown in Figure 2, in specific application, this method comprises the following steps:

(1) Video space feature encoding, which is to map the video frame sequence to the video feature sequence; as shown in Figure 3, it specifically includes the following sub-steps:

(11) Read a frame of video image from the video stream, divide the image into MxN sub-images of the same size, and calculate the gray value of each sub-image and the number of texture edge points;

(12) Calculate the two types of LBP (Local binary pattern, local binary pattern) spatial features v_gray of the grayscale image, as shown in Figure 4. The center feature (f) and the boundary feature (g) are shared by the center feature and the boundary feature Constitute the spatial distribution characteristics of 8-bit video frames, see (h) in Fig. 4;

(13) As above, calculate the LBP spatial distribution feature v_texture of the edge texture image. For simplicity, the number of internal edges of the image block can be counted as a measure of texture complexity, and the calculation result is an 8-bit spatial texture distribution feature as above;

(14) Combining v_gray and v_texture features to construct multiple frame features v=(v_gray, v_texture), we call a frame feature v a frame visual word (visual word);

(15) In addition to calculating the grayscale and texture LBP spatial features of graphics, other frame features can also be added, such as the color histogram v_color_his_16 of 8 or 16 bins, at this time v=(v_gray, v_texture, v_color_his_16); the frame feature composition method It can overcome the defect that a single feature subspace cannot express video frames well. the

This patent does not consider the time feature, because the time feature of the retrieved video is affected by other interference factors such as low frame rate or lack of frames, which has great uncertainty, and the spatiotemporal features of the frames constructed according to the time series are likely to be wrong. Instead, temporal order consistency is verified during similarity retrieval. the

(2) Video key feature extraction, that is, extracting representative features as the key feature vector of the video feature sequence; as shown in Figure 5, specifically including the following sub-steps:

(21) The first spatial feature vector of the video feature sequence is used as the default key feature vector;

(22) Extract the spatial feature vector v(n) of the current nth frame, if the Mahalanobis distance between the current feature v(n) and the previous key feature vector (v(m),m) is greater than the threshold thrsh, considering the noise factor , in this paper 1<=thrsh<=2, then the current frame is the key feature vector, denoted as (v(n),n). the

Two different feature vectors v1 and v2 express different video contents. Replacing the traditional key frame vector with a representative key feature vector key vector not only saves the step of key frame extraction, but also expresses video content more directly and accurately with original features. The completeness of video index information and the selection of index features are solved. the

We call the key feature vector (key vector) a visual word, and the collection of visual words is called a visual vocabulary. The histogram of the feature vector set of a single video file is called a feature histogram (vector histogram or visual word histogram). In order to make the key vector have rich expression ability and abstract generalization ability, the key vector is composed of different but independent sub vectors to form the gray-LBP vector of the spatial gray distribution feature, the Texture-LBP vector and the color vector of the spatial texture distribution feature, which can be simply expressed It is key vector={Gray-LBP, Texture-LBP, Color}. The multiplicative description space composed of different abstract feature concept spaces realizes the rich expressive ability and abstract generalization ability of key vector. the

The difference between this patent and other key frame extraction is that the present invention directly extracts key features from video streams, rather than traditional key frame extraction. the

Traditional key frame extraction is to use the key frame extraction algorithm to extract key frames, and then use the extracted key frames to extract retrieval features. The method used to extract key frames is not exactly the same as the retrieval features calculated after extracting key frames, and sometimes the difference is very large , which will lead to inaccurate description; this is one of the reasons why the accuracy of traditional retrieval features is not high enough. the

(3) The mapping of the video frame sequence to the video visual word histogram specifically includes the following sub-steps:

(31) Since visual words may have very high dimensions, for example, the dimensions (8, 8, 16) of (f_gray, f_texture, f_color_his_16) have a total of 32 dimensions, and its memory requirement is nearly 1GB. We re-project the 32-dimensional space respectively Go to the f_gray8-bit subspace, f_texture8-bit subspace, and f_color_his_16-bit subspace, respectively count their histograms in the subspaces, the memory requirements are significantly reduced, less than 70MB, and the histogram size of a single video file hardly exceeds 10MB;

(32) The value of the bin (a subspace feature, such as an 8-bit LBP feature) of the subspace projection histogram represents the frequency of the feature appearing in the video. In order to maintain the time distribution of the feature in the same bin, the following method is used Record bin content:

bin: (The frequency of occurrence of this feature is the sum of n1+n2+...+nk, frame number T1, number of consecutive occurrences n1, T2, n2,..., Tk, nk). the

(4) Establish the video file inverted index file, which specifically includes the following sub-steps:

(41) Statistically count the set of visual words corresponding to each video in the video library to form the statistical visual word library VwSet of the video library. according to

Vw_i (the i-th visual word in the visual lexicon) establishes a document collection {vf1, vf2, vf3,...,vfni} with the visual word, ni is the size of the document collection;

(42) The documents of the visual word document collection are sorted from large to small according to the number of visual words contained;

(43) Since high-dimensional visual words are projected into low-dimensional feature subspaces, an inverted index file is established according to each subspace. the

(5) Establish a distributed storage index library, specifically including the following steps:

(51) Use the p-stable-based Local Sensitive Hash Algorithm (LSH) to map the subspace feature f_v, (such as f_colo_his_16) to a one-dimensional space [0-Range);

(52) Adopt the HDFS file system architecture of hadoop, maintain the LSH table with name_node, and save index data with data_node. the

(6) Video similarity calculation, specifically including the following steps:

The histogram of the retrieved video Vq subspace is {Bin_q_1, Bin_q_2,...,Bin_q_M}, M is the size of the feature subspace, the histogram of the video library Vi is {Bin_i_1, Bin_i_2,...,Bin_i_M}, in Bin_id_n, id is the unique number of the video , n is the serial number of the histogram bin,

Bin_id_n is the number of occurrences of this feature;

(61) Video similarity is the intersection of histograms,

$\mathrm{<span\; class="notranslate">\; sim</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; wxya</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Vi</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; Bin</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Bin</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>\underset{\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; Bin</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Bin</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>$

(62) If the similarity is greater than the threshold thrsh_sim, compare the time series relationship of visual words. Histogram time series information has been recorded in step (32), and its algorithm is as follows:

Represent the video in the order in which the visual words of the retrieved video appear in time, such as {(Vq_vw1,Bin_k1),(Vq_vw2,Bin_k2),…, 

(Vq_vwl, Bin_kl))}, where vw1 is the first visual word that occurs in time, vw2 occurs subsequently, and Bin_k1 represents that the sequence number of the histogram Bin where the visual word is located is k1, the total number of kl histogram bins;

(63) If the visual word Vq_vw(x) in the retrieved video appears earlier than Vq_vw(y), x<y, then the matching video histogram corresponds to all sequences of the same visual word contained in the Bin with the sequence number Bin_kx Among the serial numbers, at least one is smaller than one of the serial numbers corresponding to Bin_ky; we believe that the sequence of appearance of the retrieval visual words should be consistent with the sequence of appearance of the same visual words in similar videos, that is, the corresponding spatio-temporal structures are consistent, and through time The order is desirable to remove a large number of suspected similar videos. the

(7) Retrieve the video, specifically in the following ways:

Extract the histogram of visual words from the retrieved video, map the visual word features into hash values in each subspace, determine the name_node and data_nodes where the access hash bucket is located, and select the most similar top 20% as output through inverted index video files. Then calculate the consistency of the space-time structure, and output all retrieved video files with a similarity greater than 0.7 according to the similarity. the

Adopting the method for realizing massive video fast retrieval in this invention has the following beneficial effects:

The invention mainly aims at the integrity of video index information and the selection of index features, and proposes a subspace method based on video fingerprints to solve the current fast and robust retrieval problem facing massive data. First of all, this patent adopts a novel key frame extraction method, replaces key frame extraction with key feature vector extraction, and directly replaces key frames with representative visual features, which is equivalent to encoding the original video in the feature space and completely expressing The video information has neither a large amount of redundancy, but also is very compact, and overcomes the current key frame extraction parameter selection problem. Secondly, each visual feature is mapped into a one-dimensional hash value, and the name_node (name node) and data_node (data node) of the appropriate HDFS (Hadoop Distributed File System, Hadoop Distributed File System) are selected according to the range of the hash value of the visual feature. node), which not only speeds up the retrieval speed, but also enables it to have the ability to process massive data concurrently, and has a wider range of applications. the

In this specification, the invention has been described with reference to specific embodiments thereof. However, it is obvious that various modifications and changes can be made without departing from the spirit and scope of the invention. Accordingly, the specification and drawings are to be regarded as illustrative rather than restrictive. the

Claims (13)

1. A method for realizing massive video fast retrieval, is characterized in that, described method comprises the following steps: (1) Each frame of video images in the video stream of the video library extracts the spatial feature vector to obtain the video feature sequence; (2) extract key feature vector in the spatial feature vector of described video feature sequence; (3) set up the distributed storage index library of all video files according to the key feature vector of all video files in the video library; (4) extract the key feature vector set of the video to be retrieved and extract the video index file of the video to be retrieved; (5) Perform video similarity retrieval in the distributed storage index library according to the video index file of the video to be retrieved, and output video retrieval results whose similarity is greater than the system preset value. 2. the method for realizing massive video fast retrieval according to claim 1, is characterized in that, described spatial characteristic vector comprises the grayscale spatial distribution characteristic and the texture spatial distribution characteristic of corresponding frame image, and described pair video storehouse Each frame video image in the video stream extracts the spatial feature vector respectively, comprises the following steps: (11) calculate the grayscale image and the edge texture image of each frame video image in the video stream of the video storehouse; (12) calculate the center space feature and boundary space feature of the grayscale image of each frame video image and obtain the grayscale space distribution feature of this frame video image that is made up of described center space feature and boundary space feature; (13) Calculate the texture spatial distribution feature of the edge texture image of each frame of video image. 3. the method for realizing massive video fast retrieval according to claim 2, is characterized in that, described calculation obtains the grayscale image and the edge texture image of each frame video image in the video stream of video storehouse, comprises the following steps: (111) divide each frame video image in the video flow of video bank into several sub-images of the same size and calculate the gray value and texture edge point number of each sub-image; (112) calculate the grayscale value of each sub-image of each frame video image to obtain the grayscale image of this frame video image; (113) Calculate the number of texture edge points of each sub-image of each frame of video image to obtain the edge texture image of the frame of video image. 4. the method for realizing mass video fast retrieval according to claim 2, is characterized in that, the central space feature and the boundary space feature of the grayscale image of each frame video image of the described calculation, specifically: Calculate the central space feature and the boundary space feature of the local binary pattern of the grayscale image of each frame video image; The texture space distribution feature of the edge texture image of the described calculation each frame video image is specifically: The texture spatial distribution feature of the local binary pattern of the edge texture image of each frame video image is calculated. 5. the method for realizing massive video fast retrieval according to claim 2, is characterized in that, described spatial feature vector also comprises color histogram feature, and each frame video image extracts respectively in the video flow of described video storehouse The spatial feature vector also includes the following steps: (14) Calculate the color histogram feature of each frame of video image. 6. the method for realizing massive video fast retrieval according to claim 1, is characterized in that, described extracting key feature vector in the spatial feature vector of described video feature sequence, comprises the following steps: (21) the first spatial feature vector of the video feature sequence is defaulted as the key feature vector; (22) Calculate the Mahalanobis distance between each spatial eigenvector and the previous key eigenvector; (23) Extract the spatial feature vector corresponding to the Mahalanobis distance greater than the system preset threshold as the key feature vector. 7. the method for realizing massive video fast retrieval according to claim 1, is characterized in that, described according to the key feature vector of all video files in the video library, the distributed storage index storehouse of all video files is set up, comprises the following steps: (31) establish the subspace projection histogram of the key feature vector in the video feature sequence and record the frequency of occurrence of each key feature vector in the corresponding video; (32) set up the inverted index file of all video files of video storehouse; (33) Establish a distributed index database of all video files in the video library. 8. the method for realizing mass video fast retrieval according to claim 7, is characterized in that, the subspace projection histogram of the key feature vector in the described establishment video feature sequence is specifically: The key feature vectors in the video feature sequence are projected into the gray subspace, texture subspace and color subspace, and the subspace projection histogram of each key feature vector is obtained. 9. the method for realizing massive video fast retrieval according to claim 8, is characterized in that, the frequency that each key feature vector of described recording occurs in the corresponding video is specifically: A feature value representing the frequency of appearance of the key feature vector in the video is recorded in the subspace projection histogram corresponding to each key feature vector. 10. the method for realizing massive video fast retrieval according to claim 8, is characterized in that, the inverted index file of all video files of described setting up video storehouse, comprises the following steps: (321) the corresponding key feature vector set of each video file in the statistical video library and constitute the statistical key feature vector library of the video library; (322) Establishing a collection of documents corresponding to each key feature vector in the statistical key feature vector library having the key feature vector; (323) sorting the documents of the key feature vector set according to the number of key feature vectors contained therein from more to less; (324) Establish an inverted index file of all video files in the video library according to each subspace. 11. the method for realizing massive video fast retrieval according to claim 10, is characterized in that, the distributed index database of all video files of described setting up video storehouse, comprises the following steps: (331) utilizing a locality-sensitive hash algorithm based on p-stable to map the key feature vectors of each subspace to a one-dimensional space; (332) Based on the Hadoop distributed file system architecture, name_node is used to maintain the hash table and data_node is used to store index data as a distributed index database for all video files. 12. The method for realizing fast retrieval of massive videos according to claim 7, wherein the video similarity retrieval is performed in the distributed storage index library according to the video index file of the video to be retrieved ,Specifically: (51) Computing the intersection of each video subspace projection histogram in the video subspace projection histogram to be retrieved and the video library as the similarity of each video in the video to be retrieved and the video library; (52) According to the time-space structure consistency of the video to be retrieved and the key feature vectors of each video in the video library, video files that do not meet the time-space structure consistency requirements are eliminated. 13. the method for realizing massive video fast retrieval according to claim 12, is characterized in that, described output similarity is greater than the video retrieval result of system default value, comprises the following steps: (52) extract each subspace projection histogram of the key feature vector of the video to be retrieved and map each key feature vector into a hash value in each subspace; (53) selecting a video file whose similarity in the distributed index database meets the system preset requirements through the inverted index file as output; (54) Calculate the spatio-temporal structure consistency of the video to be retrieved and the key feature vectors of each video in the video database and output a video file whose similarity with the video to be retrieved is greater than the system preset value.