WO2016086731A1 - 多级并行关键帧云提取方法及系统 - Google Patents

多级并行关键帧云提取方法及系统 Download PDF

Info

Publication number
WO2016086731A1
WO2016086731A1 PCT/CN2015/092349 CN2015092349W WO2016086731A1 WO 2016086731 A1 WO2016086731 A1 WO 2016086731A1 CN 2015092349 W CN2015092349 W CN 2015092349W WO 2016086731 A1 WO2016086731 A1 WO 2016086731A1
Authority
WO
WIPO (PCT)
Prior art keywords
key frame
video
parallel
lens
extraction
Prior art date
Application number
PCT/CN2015/092349
Other languages
English (en)
French (fr)
Inventor
朱定局
汤庸
蒋运承
Original Assignee
华南师范大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华南师范大学 filed Critical 华南师范大学
Publication of WO2016086731A1 publication Critical patent/WO2016086731A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
    • G06V20/47—Detecting features for summarising video content
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/49—Segmenting video sequences, i.e. computational techniques such as parsing or cutting the sequence, low-level clustering or determining units such as shots or scenes

Definitions

  • the present invention relates to a key frame cloud extraction technology, and more particularly to a multi-level parallel key frame cloud extraction method and a multi-level parallel key frame cloud extraction system.
  • a video keyframe can represent one of the most important, representative images or images in a shot.
  • the extraction of key frames can greatly reduce the processing capacity of video data, which has attracted wide attention of researchers.
  • Cloud computing can be seamlessly extended to large-scale clusters, and can tolerate error codes of some nodes. Even a large number of nodes fail to affect the correct operation of the program, so cloud computing has better scalability and stability. .
  • the main methods of key frame extraction include: lens boundary based method, key frame extraction based on motion analysis, key frame extraction based on image information, key frame extraction based on clustering, and the like.
  • Most of the research focuses on the extraction accuracy of video keyframes, but the extraction of video keyframes involves many digital operations. The number of video frames is large, the serial video keyframe extraction takes a long time, and the single-machine processing capability is limited. It is necessary to study a multi-level parallel key frame cloud extraction method and cloud extraction system.
  • an object of the present invention is to provide a multi-level parallel key frame cloud extraction system, wherein the multi-level parallel key frame cloud extraction system includes: a video input module configured to extract a key a video input interface of the frame; the video lens segmentation module is configured to perform lens segmentation on the input video for key frame extraction parallel processing; the multi-level parallel key frame cloud processing module is configured to pass the multi-level parallel strategy for the video lens By extracting the frame feature vector and the parallel clustering operation in parallel, each lens key frame group is generated; the key frame output module is configured to combine the key frame groups generated by the respective lenses, and finally output the key frame group of the entire video.
  • Another object of the present invention is to provide a multi-level parallel key frame cloud extraction method, wherein the multi-level parallel key frame cloud extraction method includes: accepting a video that needs to extract a key frame; and performing lens segmentation on the video; Multi-level parallel key frame cloud extraction operation is performed on the lens; the extracted lens key frame groups are combined to output a final video key frame group.
  • the multi-level parallel key frame cloud extraction operation includes parallel extraction of video frame feature vectors and parallel clustering operations according to video frame feature vectors.
  • the lens key frame group combination includes a final cluster to generate a video final key frame.
  • the video feature vector clustering operation may be k-means clustering, fuzzy C-means clustering or other image feature clustering algorithm suitable for parallel processing.
  • the multi-level parallel key frame cloud extraction method and the cloud extraction system of the invention can greatly improve the key frame extraction efficiency, and have good scalability and stability.
  • FIG. 1 is a schematic diagram of a multi-level parallel key frame cloud extraction system in accordance with an embodiment of the present invention.
  • FIG. 2 is a flow diagram of a multi-level parallel keyframe cloud extraction scheduling method in accordance with an embodiment of the present invention.
  • FIG. 3 is a diagram of an example operation of a multi-level parallel keyframe cloud extraction in accordance with an embodiment of the present invention.
  • FIG. 1 is a schematic diagram of a multi-level parallel key frame cloud extraction system in accordance with an embodiment of the present invention.
  • a multi-level parallel key frame cloud extraction system includes a video input module 10 configured to extract a video input interface of a key frame, and a video shot segmentation module 20 configured to input an input video.
  • the lens segmentation is performed for key frame extraction parallel processing;
  • the multi-level parallel key frame cloud processing module 30 is configured to generate a frame key by parallelly extracting frame feature vectors and parallel clustering operations by using a multi-level parallel strategy for the video lens.
  • the frame group; the key frame output module 40 is configured to combine the key frame groups generated by the respective lenses, and finally output the key frame group of the entire video.
  • the multi-level parallel key frame cloud extraction operation includes parallel extraction of video frame feature vectors and parallel clustering operations according to video frame feature vectors.
  • the video feature vector clustering operation may be k-means clustering, fuzzy C-means clustering or other image feature clustering algorithm suitable for parallel processing.
  • the present invention further provides a multi-level parallel key frame cloud extraction method.
  • FIG. 2 is a flowchart of a multi-level parallel key frame cloud extraction method according to an embodiment of the present invention.
  • a multi-level parallel key frame cloud extraction method includes: S1, accepting a video that needs to perform key frame extraction; S2, performing video shot segmentation on the video; S3, and extracting video in multiple stages of the lens. Key frame; S4, combined output of the lens key frame group.
  • the video frame feature vector may be extracted in parallel, and then the video feature vector clustering operation may be performed in parallel, and the clustering may be k-means clustering, fuzzy C-means clustering or other suitable parallel processing.
  • Image feature clustering algorithm The present invention will be described below by taking parallel extraction video frame vector processing and performing k-means clustering to extract key frames in parallel.
  • the K-means algorithm is a typical distance-based clustering algorithm, using distance as the The evaluation index of similarity, that is, the closer the distance between two objects is, the greater the similarity.
  • fuzzy c-means algorithm FCMA
  • FCMA fuzzy c-means algorithm
  • S3 is constructed as follows:
  • the first-level lens map operation the operation is to distribute the split lens to the first-level cloud computing platform node.
  • the node described here has a dual role, specifically: the Datanode and the TaskTasker role in the first-level cloud platform.
  • the Namenode and Jobtracker roles In the second-level cloud computing platform node, the Namenode and Jobtracker roles.
  • the map function constructs the ⁇ key, value> of the input data record as ⁇ lens ID, video lens position>, and the function operation is: copying the lens to the first-level cloud computing platform node, so as to perform the second-level map operating.
  • S32 Perform a second-level map operation preprocessing on the lens distributed to the node. Split the shot into a set of video frame images.
  • the lens 1 includes the number of video frames N 1
  • the lens 2 includes the number of video frames N 2
  • the lens K includes the number of video frames N k .
  • N 1 , N 2 , ... N k refers to the number of video frames actually included in each of the lenses N 1 , N 2 , ..., N k (N 1 , N 2 , ... N k ⁇ 1, N 1 , N 2 , ... N k ⁇ Z).
  • S33 Perform a second level map operation on each lens in parallel.
  • the second level map operation is to distribute each frame to the next-level cloud computing platform Tasktracker node, and extract the video frame feature vector in parallel.
  • the map function constructs the ⁇ key, value> of the input data record as ⁇ frame ID, frame image position>, and the function operates to extract the feature vector of the video frame, and the output result ⁇ key, value> is in the form of ⁇ frame ID, frame feature vector>.
  • S34 Perform a third-level map operation on the result obtained by the S33 operation.
  • the operation of S34 is to perform a cloud clustering operation in parallel on the ⁇ frame ID, frame feature vector> record group generated by S33. Specifically, it can be described as: calculating the distance from each data frame to the initial frame cluster center, and re-marking the new cluster category to which it belongs, and the input is the ⁇ frame ID, frame feature vector> all records and on the S33 generated.
  • Each map function reads the cluster center description file, and the map function calculates its nearest class center for each input point entered, and marks the new category.
  • the ⁇ key, value> of the Map function input data record is ⁇ frame ID, frame feature vector>; the output intermediate result ⁇ key, value> is of the form ⁇ frame belonging category, frame feature vector>.
  • the number of initial cluster centers of the above lens 1, lens 2, ..., lens k is described as m 1 , m 2 , ..., m k (m 1 , m 2 , ..., m k ⁇ 1, m 1 , m 2 , ..., m k ⁇ Z)
  • the value is determined according to the total number of data frames of each lens and a certain rule.
  • m 1 , m 2 , ..., m k also represent the number of key frames that each lens will produce.
  • Said initial cluster center is configured to: according to 1, 2, ..., k m values of m m, respectively, from the lens 1, the lens 2, ..., k randomly selected lens m 1, m 2, ..., m k th sign
  • the feature vector is used as the initial clustering center of the lens 1, the lens 2, ..., the lens k.
  • the distance calculation of the above data frame to the cluster center can be described as Euclidean distance, Mahalanobis distance, and the like.
  • the map function constructed by S34 can be described as:
  • the above map stage performs a shuffle operation to complete the grouping of the intermediate calculation results.
  • S35 Update the cluster center according to the output of S34 for use in the next round of map-reduce.
  • the input data ⁇ key, value> pair is in the form of ⁇ cluster class ID, ⁇ record attribute vector set ⁇ >; all records with the same key (ie, records of the same cluster center category ID) will be sent to one Reduce task.
  • the S35 operation is described as: summing the number of points with the same key and the sum of the respective recorded components, and finding the mean of each component to obtain a new cluster center.
  • the Skey operation output result ⁇ key, value> is in the form of ⁇ cluster class ID, mean vector>.
  • the S35 operation process can be described as:
  • step S36 Iterative process.
  • the step S36 can be described as: judging whether the cluster has converged on the output result of S34. Specifically, it can be described as: comparing the distance between the cluster center obtained by the previous round of map-reduce and the current map-reduce cluster center. If the distance is less than a given threshold, the algorithm ends. Otherwise, replace the cluster center of this round with the cluster center of the previous round and start a new round of map-reduce operation.
  • S4 Combined output of the lens key frame group.
  • the output video key frames are combined for the final lens video key group of the S3 step.
  • the multi-level parallel key frame cloud extraction method and the cloud extraction system have the characteristics of multi-level parallel operation, which can greatly improve the key frame extraction efficiency.
  • the system is based on the cloud computing platform, so it has better scalability and stability.

Abstract

本发明公开一种多级并行关键帧云提取方法及系统。所述关键帧云提取系统包括：视频输入模块，被构造为提取关键帧的视频输入接口；视频镜头分割模块，被构造为对输入的视频进行镜头分割，以便进行关键帧提取并行处理；多级并行关键帧云处理模块，被构造为对视频镜头通过多级并行策略，通过并行提取帧特征向量、并行聚类操作，生成各镜头关键帧组；关键帧输出模块，被构造为对各镜头所生成的关键帧组进行组合，最后输出整个视频的关键帧组。本发明的多级并行关键帧云提取方法及系统对处理大规模关键帧提取时具有较高的效率，同时，具有很好的可扩展性及稳定性，能够满足大规模关键帧提取的需求。

Description

多级并行关键帧云提取方法及系统 技术领域
本发明涉及关键帧云提取技术，更具体地讲，涉及一种多级并行关键帧云提取方法及多级并行关键帧云提取系统。
背景技术
视频关键帧指能代表镜头中的最重要的、有代表性的一幅或多幅图像。关键帧的提取能大大减少视频数据的处理量，受到研究者的广泛关注。
云计算能够无缝扩展到大规模的集群，且能够容忍部分节点的错误码，甚至很大部分节点发生失效也不会影响程序的正确运行，因此云计算具有较好的可扩展性及稳定性。
现有的关键帧提取的主要方法包括：基于镜头边界的方法、基于运动分析提取关键帧、基于图像信息提取关键帧、基于聚类提取关键帧等。大部分研究主要集中在视频关键帧的提取准确度方面，但视频关键帧的提取涉及许多数字运算，具视频帧数量多，串行视频关键帧抽取会耗时长，且单机处理能力有限，故有必要研究一种多级并行关键帧云提取方法和云提取系统。
发明内容
为了解决上述现有技术存在的问题，本发明的目的在于提供一种多级并行关键帧云提取系统，其中，所述多级并行关键帧云提取系统包括：视频输入模块，被构造为提取关键帧的视频输入接口；视频镜头分割模块，被构造为对输入的视频进行镜头分割，以便进行关键帧提取并行处理；多级并行关键帧云处理模块，被构造为对视频镜头通过多级并行策略，通过并行提取帧特征向量、并行聚类操作，生成各镜头关键帧组；关键帧输出模块，被构造为对各镜头所生成的关键帧组进行组合，最后输出整个视频的关键帧组。
本发明的另一目的还在于提供一种多级并行关键帧云提取方法，其中，所述多级并行关键帧云提取方法包括：接受需提取关键帧的视频；对所述视频进行镜头分割；对所述镜头进行多级并行关键帧云提取操作；对所提取的镜头关键帧组进行组合，输出最终视频关键帧组。
进一步地，所述多级并行关键帧云提取操作包含并行提取视频帧特征向量、根据视频帧特征向量并行聚类操作。
进一步地，所述镜头关键帧组组合，包含最终聚类产生视频最终关键帧。
进一步地，所述视频特征向量聚类操作可为k-means聚类、模糊C均值聚类或其他适合并行处理的图像特征聚类算法。
本发明的多级并行关键帧云提取方法及云提取系统能大大提高关键帧提取效率，同时具有很好的可扩展性及稳定性。
附图说明
图1是根据本发明的实施例的多级并行关键帧云提取系统示意图。
图2是根据本发明的实施例的多级并行关键帧云提取调度方法的流程图。
图3是根据本发明的实施例的多级并行关键帧云提取某实例操作图。
具体实施方式
现在对本发明的实施例进行详细的描述，其示例表示在附图中，其中，相同的标号始终表示相同部件。下面通过参照附图对实施例进行描述以解释本发明。在附图中，为了清晰起见，可以夸大层和区域的厚度。在下面的描述中，为了避免公知结构和/或功能的不必要的详细描述所导致的本发明构思的混淆，可省略公知结构和/或功能的不必要的详细描述。
图1是根据本发明的实施例的多级并行关键帧云提取系统示意图。
参照图1，根据本发明的实施例的多级并行关键帧云提取系统包括：视频输入模块10，被构造为提取关键帧的视频输入接口；视频镜头分割模块20，被构造为对输入的视频进行镜头分割，以便进行关键帧提取并行处理；多级并行关键帧云处理模块30，被构造为对视频镜头通过多级并行策略，通过并行提取帧特征向量、并行聚类操作，生成各镜头关键帧组；关键帧输出模块40，被构造为对各镜头所生成的关键帧组进行组合，最后输出整个视频的关键帧组。
此外，多级并行关键帧云提取操作包含并行提取视频帧特征向量、根据视频帧特征向量并行聚类操作。
所述视频特征向量聚类操作可为k-means聚类、模糊C均值聚类或其他适合并行处理的图像特征聚类算法。
相对应地，本发明还提供了一种多级并行关键帧云提取方法，具体请参照图2，其是根据本发明的实施例的多级并行关键帧云提取方法的流程图。
参照图2，根据本发明的实施例的多级并行关键帧云提取方法包括：S1、接受需进行关键帧抽取的视频；S2、对视频进行视频镜头分割；S3、对镜头多级并行提取视频关键帧；S4、对镜头关键帧组进行组合输出。
在本实施例中，视频镜头分割后，可并行提取视频帧特征向量，之后可并行进行视频特征向量聚类操作，聚类可为k-means聚类、模糊C均值聚类或其他适合并行处理的图像特征聚类算法。下面将以并行提取视频帧向量处理及并行进行k-means聚类提取关键帧为例来对本发明进行说明。其中，K-means算法是很典型的基于距离的聚类算法，采用距离作为 相似性的评价指标，即认为两个对象的距离越近，其相似度就越大。而，模糊c-均值聚类算法fuzzy c-means algorithm(FCMA)作为无监督机器学习的主要技术之一，是用模糊理论对重要数据分析和建模的方法，建立了样本类属的不确定性描述，能比较客观地反映现实世界。
具体而言，参照图3，假如经过S1、S2步骤后，视频已分割成k个镜头(k≥1，k∈Z)。S3被构造为如下步骤：
S31：一级镜头map操作，此操作实现将分割的镜头分布到一级云计算平台节点中，此处所述的节点为双重角色，具体为：在第一级云平台中为Datanode及TaskTasker角色，在第二级云计算平台节点中为Namenode及Jobtracker角色。S301操作中，map函数构造输入数据记录的<key,value>为<镜头ID，视频镜头位置>，函数操作为：将镜头拷贝至第一级云计算平台节点中，以便进行第二级的map操作。
S32：对分发到节点上的镜头作第二级map操作预处理。将镜头分割成一组视频帧图像。镜头1包含视频帧数N1、镜头2包含视频帧数N2、镜头K包含视频帧数Nk。N1、N2、…Nk指各镜头N1、N2、…、Nk实际所包含的视频帧数(N1、N2、…Nk≥1,N1、N2、…Nk∈Z)。
S33：对各镜头并行进行第二级map操作。所述第二级map操作，在于将各帧分发到下一级云计算平台Tasktracker节点中，并行提取视频帧特征向量。所述S33操作中，map函数构造输入数据记录的<key,value>为<帧ID，帧图像位置>，函数操作为对视频帧进行特征向量提取，输出结果<key,value>对的形式为<帧ID，帧特征向量>。
S34：对S33操作所得到的结果进行第三级map操作。所述S34操作，在于对S33生成的<帧ID，帧特征向量>记录组并行进行云聚类操作。具体可描述为：完成每个数据帧到初始帧聚类中心的距离计算，并重新标记其属于的新聚类类别，其输入为S33所生成的<帧ID，帧特征向量>所有记录和上一轮迭代(或初始聚类)的聚类中心。每个map函数都读入聚类中心描述文件，map函数对输入的每个记录点计算其最近的类中心，并做新类别的标记。Map函数输入数据记录的<key,value>为<帧ID，帧特征向量>；输出中间结果<key,value>的形式为<帧所属类别，帧特征向量>。
上述镜头1、镜头2、…、镜头k的初始聚类中心个数被描述为m1、m2、…、mk(m1、m2、…、mk≥1，m1、m2、…、mk∈Z)其值根据各镜头数据帧总数及一定的规则确定，在一定程度上，m1、m2、…、mk也代表各镜头将产生的关键帧数。
上述初始聚类中心被构造为：根据m1、m2、…、mk的值，分别从镜头1、镜头2、…、镜头k中随机抽取m1、m2、…、mk个征特征向量作为镜头1、镜头2、…、镜头k的初始聚类中心。
上述数据帧到聚类中心的距离计算，可以描述为欧式距离、马氏距离等。
S34被构造的map函数可以描述为：
Figure PCTCN2015092349-appb-000001
上述map阶段会进行shuffle操作，完成中间计算结果的分组排序。
S35：根据S34的输出，更新聚类中心，供下一轮map-reduce使用。所述S35操作，输入数据<key,value>对的形式为<聚类类别ID，{记录属性向量集}>；所有key相同的记录(即相同聚类中心类别ID的记录)将送给一个reduce任务。S35操作被描述为：累加key相同的点的个数和各记录分量的和，求各分量的均值，得到新的聚类中心。S35操作输出结果<key,value>对的形式为<聚类类别ID，均值向量>。S35操作过程可描述为：
Figure PCTCN2015092349-appb-000002
S36：迭代过程。S36步骤可描述为：对S34的输出结果，判断该聚类是否已收敛。具体可描述为：比较上一轮map-reduce得到的聚类中心与本轮map-reduce聚类中心的距离。若距离小于给定阀值，则算法结束。否之，则将本轮的聚类中心替换上一轮的聚类中心，并启动新一轮的map-reduce操作。
S4：对镜头关键帧组进行组合输出。对S3步聚的最终各镜头视频关键帧组进行组合输出视频关键帧。
综上所述，根据本发明的实施例的多级并行关键帧云提取方法及云提取系统，具有多级并行操作的特点，能大大提高关键帧提取效率。同时，系统基于云计算平台，故具有较好的可扩展性和稳定性。
尽管已经参照其示例性实施例具体显示和描述了本发明，但是本领域的技术人员应该理解，在不脱离权利要求所限定的本发明的精神和范围的情况下，可以对其进行形式和细节上的各种改变。

Claims (6)

  1. 一种多级并行关键帧云提取系统，其特征在于，其包括：
    视频输入模块，被构造为提取关键帧的视频输入接口；
    视频镜头分割模块，被构造为对输入的视频进行镜头分割，以便进行关键帧提取并行处理；多级并行关键帧云处理模块，被构造为对视频镜头通过多级并行策略，通过并行提取帧特征向量、并行聚类操作，生成镜头关键帧组；
    关键帧输出模块，被构造为对镜头所生成的关键帧组进行组合，最后输出整个视频的关键帧组。
  2. 根据权利要求1所述的多级并行关键帧云提取系统，其特征在于，多级并行关键帧云处理模块进行并行提取视频帧特征向量、根据视频帧特征向量并行聚类操作。
  3. 根据权利要求2所述的多级并行关键帧云提取系统，其特征在于，所述多级并行关键帧云处理模块中的视频特征向量聚类操作可为k-means聚类、模糊C均值聚类或其他适合并行处理的图像特征聚类算法。
  4. 一种多级并行关键帧云提取方法，其特征在于，所述多级并行关键帧云提取方法包括如下步骤：
    S1.接受需进行关键帧抽取的视频；
    S2.对视频进行视频镜头分割；
    S3.对镜头多级并行提取视频关键帧；
    S4.对镜头关键帧组进行组合输出。
  5. 根据权利要求4所述的多级并行关键帧云提取方法，其特征在于，所述步骤s3包含并行提取视频帧特征向量操作，并根据视频帧特征向量并行聚类操作。
  6. 根据权利要求5所述的多级并行关键帧云提取方法，其特征在于，所述视频特征向量聚类操作为k-means聚类或模糊C均值聚类或其他适合并行处理的图像特征聚类算法中的一种或多种。
PCT/CN2015/092349 2014-12-05 2015-10-21 多级并行关键帧云提取方法及系统 WO2016086731A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201410731007.0A CN104463864B (zh) 2014-12-05 2014-12-05 多级并行关键帧云提取方法及系统
CN201410731007.0 2014-12-05

Publications (1)

Publication Number Publication Date
WO2016086731A1 true WO2016086731A1 (zh) 2016-06-09

Family

ID=52909846

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/092349 WO2016086731A1 (zh) 2014-12-05 2015-10-21 多级并行关键帧云提取方法及系统

Country Status (2)

Country Link
CN (1) CN104463864B (zh)
WO (1) WO2016086731A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104463864B (zh) * 2014-12-05 2018-08-14 华南师范大学 多级并行关键帧云提取方法及系统
CN108921773A (zh) * 2018-07-04 2018-11-30 百度在线网络技术（北京）有限公司 人体跟踪处理方法、装置、设备及系统
CN110889857A (zh) * 2019-11-15 2020-03-17 北京邮电大学 一种移动Web实时视频帧分割方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040170321A1 (en) * 1999-11-24 2004-09-02 Nec Corporation Method and system for segmentation, classification, and summarization of video images
CN101296373A (zh) * 2007-04-27 2008-10-29 新奥特硅谷视频技术有限责任公司 一种基于素材交换格式的多媒体数据处理系统及方法
CN102693299A (zh) * 2012-05-17 2012-09-26 西安交通大学 一种并行视频拷贝检测系统和方法
CN104463864A (zh) * 2014-12-05 2015-03-25 华南师范大学 多级并行关键帧云提取方法及系统

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100477809C (zh) * 2005-07-15 2009-04-08 复旦大学 一种测量音视频内容变化的方法
RU2011146075A (ru) * 2009-04-14 2013-05-20 Конинклейке Филипс Электроникс Н.В. Извлечение ключевых кадров для анализа видеоконтента
CN103064935B (zh) * 2012-12-24 2016-05-18 深圳先进技术研究院 一种多媒体数据并行处理系统及方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040170321A1 (en) * 1999-11-24 2004-09-02 Nec Corporation Method and system for segmentation, classification, and summarization of video images
CN101296373A (zh) * 2007-04-27 2008-10-29 新奥特硅谷视频技术有限责任公司 一种基于素材交换格式的多媒体数据处理系统及方法
CN102693299A (zh) * 2012-05-17 2012-09-26 西安交通大学 一种并行视频拷贝检测系统和方法
CN104463864A (zh) * 2014-12-05 2015-03-25 华南师范大学 多级并行关键帧云提取方法及系统

Also Published As

Publication number Publication date
CN104463864B (zh) 2018-08-14
CN104463864A (zh) 2015-03-25

Similar Documents

Publication Publication Date Title
Parkhi et al. Deep face recognition
WO2021109464A1 (zh) 一种面向大规模用户的个性化教学资源推荐方法
Xiao et al. Action recognition based on hierarchical dynamic Bayesian network
Zhang et al. Panorama: a data system for unbounded vocabulary querying over video
WO2022166380A1 (zh) 一种基于meanshift优化的数据处理方法和装置
Wen et al. CF-SIS: Semantic-instance segmentation of 3D point clouds by context fusion with self-attention
WO2022160772A1 (zh) 一种基于视角引导多重对抗注意力的行人重识别方法
WO2019080908A1 (zh) 实现图像识别的图像处理方法及装置、电子设备
Weyand et al. Visual landmark recognition from internet photo collections: A large-scale evaluation
Zhou et al. Attention-based neural architecture search for person re-identification
WO2022088390A1 (zh) 图像的增量聚类方法、装置、电子设备、存储介质及程序产品
CN110442618B (zh) 融合专家信息关联关系的卷积神经网络评审专家推荐方法
Petkos et al. Graph-based multimodal clustering for social event detection in large collections of images
CN103617609A (zh) 基于图论的k-means非线性流形聚类与代表点选取方法
WO2023155508A1 (zh) 一种基于图卷积神经网络和知识库的论文相关性分析方法
WO2016086731A1 (zh) 多级并行关键帧云提取方法及系统
Ouyang et al. Collaborative image relevance learning for visual re-ranking
Li et al. Adaptive betweenness clustering for semi-supervised domain adaptation
Patel Accelerated PSO swarm search feature selection with SVM for data stream mining big data [J]
Liao et al. Depthwise grouped convolution for object detection
Dhoot et al. Efficient Dimensionality Reduction for Big Data Using Clustering Technique
Wang et al. Coda: Counting objects via scale-aware adversarial density adaption
CN111160077A (zh) 一种大规模人脸动态聚类方法
JP2016014990A (ja) 動画像検索方法、動画像検索装置及びそのプログラム
JP2017021606A (ja) 動画像検索方法、動画像検索装置及びそのプログラム

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15866256

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 15866256

Country of ref document: EP

Kind code of ref document: A1