【視頻相似性檢測(cè)算法探究文獻(xiàn)綜述5200字】

視頻相似性檢測(cè)算法研究國(guó)內(nèi)外文獻(xiàn)綜述目錄TOC\o"1-2"\h\u2946視頻相似性檢測(cè)算法研究國(guó)內(nèi)外文獻(xiàn)綜述 167051視頻分類(lèi)算法研究現(xiàn)狀 1218372視頻關(guān)鍵幀選取研究現(xiàn)狀 2278203視頻相似性判定及檢測(cè)算法研究現(xiàn)狀 34584(1)視頻相似性判定 320444(2)視頻指紋提取 332250(3)視頻相似性檢測(cè) 410466（4）現(xiàn)存問(wèn)題及解決 61視頻分類(lèi)算法研究現(xiàn)狀視頻分類(lèi)將待檢測(cè)視頻分入與其類(lèi)別相對(duì)應(yīng)的小類(lèi)目中，加快視頻相似性檢測(cè)的速度，是計(jì)算機(jī)視覺(jué)領(lǐng)域和數(shù)字媒體應(yīng)用中的一個(gè)熱點(diǎn)研究課題。短視頻是一類(lèi)內(nèi)容龐雜的視頻數(shù)據(jù)，如何在海量短視頻數(shù)據(jù)中尋找到有效信息一直是用戶(hù)關(guān)注的問(wèn)題，由此產(chǎn)生了視頻索引、視頻檢索、視頻相似性檢測(cè)等相關(guān)應(yīng)用。視頻分類(lèi)是指給定一個(gè)視頻片段，對(duì)其中所展示的內(nèi)容進(jìn)行分類(lèi)，通過(guò)視頻分類(lèi)進(jìn)行視頻的預(yù)處理是快速獲取有效信息的一種常用手段。視頻分類(lèi)技術(shù)將同類(lèi)型視頻分為小類(lèi)目，對(duì)于創(chuàng)作者而言快速檢索到相同短視頻有助于進(jìn)行版權(quán)保護(hù)，對(duì)于視頻平臺(tái)而言較為精確的分類(lèi)有助于進(jìn)行優(yōu)化推薦算法。目前的視頻分類(lèi)的主要分類(lèi)研究方向是以視頻中人物做出的動(dòng)作為研究切入點(diǎn)進(jìn)行分類(lèi)ADDINEN.CITE<EndNote><Cite><Author>Soomro</Author><Year>2012</Year><RecNum>129</RecNum><DisplayText><styleface="superscript">[3-5]</style></DisplayText><record><rec-number>129</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618404154">129</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Soomro,K.</author><author>Zamir,A.R.</author><author>Shah,M.%JComputerScience</author></authors></contributors><titles><title>UCF101:ADatasetof101HumanActionsClassesFromVideosinTheWild</title></titles><dates><year>2012</year></dates><urls></urls></record></Cite><Cite><Author>Kuehne</Author><Year>2011</Year><RecNum>130</RecNum><record><rec-number>130</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618404299">130</key></foreign-keys><ref-typename="ConferenceProceedings">10</ref-type><contributors><authors><author>Kuehne,H.</author><author>Jhuang,H.</author><author>Garrote,E.</author><author>Poggio,T.</author><author>Serre,T.</author></authors></contributors><titles><title>HMDB:ALargeVideoDatabaseforHumanMotionRecognition</title><secondary-title>IEEEInternationalConferenceonComputerVision</secondary-title></titles><dates><year>2011</year></dates><urls></urls></record></Cite><Cite><Author>Heilbron</Author><RecNum>131</RecNum><record><rec-number>131</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618404466">131</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Heilbron,F.C.</author><author>Escorcia,V.</author><author>Ghanem,B.</author><author>Niebles,J.C.%JIEEE</author></authors></contributors><titles><title>ActivityNet:Alarge-scalevideobenchmarkforhumanactivityunderstanding</title></titles><dates></dates><urls></urls></record></Cite></EndNote>[\o"Soomro,2012#129"3-5]。早期針對(duì)視頻分類(lèi)主要采用的方法是將視頻以幀的形式進(jìn)行存儲(chǔ)，通過(guò)對(duì)單幀的圖像場(chǎng)景或人物動(dòng)作進(jìn)行分類(lèi)，從而達(dá)到對(duì)視頻進(jìn)行分類(lèi)的目的。即識(shí)別動(dòng)作靠場(chǎng)景，識(shí)別場(chǎng)景靠動(dòng)作，在UCF101數(shù)據(jù)集的準(zhǔn)確率最高只有68.7%。近年來(lái)，隨著深度學(xué)習(xí)的發(fā)展，2014年AndrejKarpathy等人ADDINEN.CITE<EndNote><Cite><Author>Karpathy</Author><Year>2014</Year><RecNum>122</RecNum><DisplayText><styleface="superscript">[6]</style></DisplayText><record><rec-number>122</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618402196">122</key></foreign-keys><ref-typename="ConferenceProceedings">10</ref-type><contributors><authors><author>Karpathy,A.</author><author>Toderici,G.</author><author>Shetty,S.</author><author>Leung,T.</author><author>Li,F.F.</author></authors></contributors><titles><title>Large-ScaleVideoClassificationwithConvolutionalNeuralNetworks</title><secondary-title>ComputerVision&PatternRecognition</secondary-title></titles><dates><year>2014</year></dates><urls></urls></record></Cite></EndNote>[\o"Karpathy,2014#122"6]將時(shí)序語(yǔ)義融合的方法分為四種。2015年TRAN等人ADDINEN.CITE<EndNote><Cite><Author>Tran</Author><Year>2014</Year><RecNum>121</RecNum><DisplayText><styleface="superscript">[7]</style></DisplayText><record><rec-number>121</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618400459">121</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Tran,D.</author><author>Bourdev,L.</author><author>Fergus,R.</author><author>Torresani,L.</author><author>Paluri,M.</author></authors></contributors><titles><title>LearningSpatiotemporalFeatureswith3DConvolutionalNetworks</title></titles><dates><year>2014</year></dates><urls></urls></record></Cite></EndNote>[\o"Tran,2014#121"7]提出C3D網(wǎng)絡(luò)，利用三維卷積提取連續(xù)幀序列的時(shí)空域特征進(jìn)行視頻分類(lèi)，在UCF101數(shù)據(jù)集上準(zhǔn)確率最高為85.2%，F(xiàn)PS為313.9。在優(yōu)化過(guò)程中，2017年CARREIRA等人ADDINEN.CITE<EndNote><Cite><Author>Carreira</Author><Year>2017</Year><RecNum>123</RecNum><DisplayText><styleface="superscript">[8]</style></DisplayText><record><rec-number>123</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618402477">123</key></foreign-keys><ref-typename="ConferenceProceedings">10</ref-type><contributors><authors><author>Carreira,J.</author><author>Zisserman,A.</author></authors></contributors><titles><title>QuoVadis,ActionRecognition?ANewModelandtheKineticsDataset</title><secondary-title>2017IEEEConferenceonComputerVisionandPatternRecognition(CVPR)</secondary-title></titles><dates><year>2017</year></dates><urls></urls></record></Cite></EndNote>[\o"Carreira,2017#123"8]提出I3D網(wǎng)絡(luò)，通過(guò)在三維卷積網(wǎng)絡(luò)基礎(chǔ)上增加網(wǎng)絡(luò)寬度的方式提高網(wǎng)絡(luò)分類(lèi)性能，在UCF101數(shù)據(jù)集上僅取RGB情況下其準(zhǔn)確率為84.5%，通過(guò)在kinetics數(shù)據(jù)集預(yù)訓(xùn)練參數(shù)進(jìn)行優(yōu)化最終結(jié)果達(dá)到98%，然而文章使用64個(gè)GPU并行訓(xùn)練，其網(wǎng)絡(luò)復(fù)雜及較大計(jì)算量較難在現(xiàn)實(shí)場(chǎng)景下應(yīng)用。針對(duì)傳統(tǒng)的三維卷積缺乏空間及通道依賴(lài)性問(wèn)題，2020年王輝濤ADDINEN.CITE<EndNote><Cite><Author>王輝濤</Author><Year>2020</Year><RecNum>124</RecNum><DisplayText><styleface="superscript">[9]</style></DisplayText><record><rec-number>124</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618403172">124</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>王輝濤</author><author>胡燕</author></authors></contributors><auth-address>武漢理工大學(xué)計(jì)算機(jī)學(xué)院;</auth-address><titles><title>基于全局時(shí)空感受野的高效視頻分類(lèi)方法%J小型微型計(jì)算機(jī)系統(tǒng)</title></titles><pages>1768-1775</pages><volume>41</volume><number>08</number><keywords><keyword>視頻分類(lèi)</keyword><keyword>卷積神經(jīng)網(wǎng)絡(luò)</keyword><keyword>通道和空間注意力</keyword><keyword>全局時(shí)空感受野</keyword><keyword>三維卷積核分解</keyword></keywords><dates><year>2020</year></dates><isbn>1000-1220</isbn><call-num>21-1106/TP</call-num><urls></urls><remote-database-provider>Cnki</remote-database-provider></record></Cite></EndNote>[\o"王輝濤,2020#124"9]提出了首先將傳統(tǒng)的三維卷積核分解成空域卷積核和時(shí)域卷積核，然后在二維網(wǎng)絡(luò)引入通道和空間的方法，在UCF101數(shù)據(jù)集準(zhǔn)確率最高為96.6%，雖然在一定程度上通過(guò)放棄準(zhǔn)確度來(lái)提升速度，然而由于將3維卷積核拆分和多網(wǎng)絡(luò)拼接，難以進(jìn)行快速實(shí)時(shí)分類(lèi)。在改進(jìn)3D卷積網(wǎng)絡(luò)方面：2017年楊曙光ADDINEN.CITE<EndNote><Cite><Author>現(xiàn)代計(jì)算機(jī)</Author><Year>2017</Year><RecNum>125</RecNum><DisplayText><styleface="superscript">[10]</style></DisplayText><record><rec-number>125</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618403460">125</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>楊曙光%J現(xiàn)代計(jì)算機(jī)</author></authors></contributors><titles><title>一種改進(jìn)的深度學(xué)習(xí)視頻分類(lèi)方法</title></titles><pages>68-71</pages><number>08</number><dates><year>2017</year></dates><urls></urls></record></Cite><Cite><Author>現(xiàn)代計(jì)算機(jī)</Author><Year>2017</Year><RecNum>125</RecNum><record><rec-number>125</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618403460">125</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>楊曙光%J現(xiàn)代計(jì)算機(jī)</author></authors></contributors><titles><title>一種改進(jìn)的深度學(xué)習(xí)視頻分類(lèi)方法</title></titles><pages>68-71</pages><number>08</number><dates><year>2017</year></dates><urls></urls></record></Cite></EndNote>[\o"現(xiàn)代計(jì)算機(jī),2017#125"10]為解決精度下降問(wèn)題，將3D卷積方法提取的短視頻特征放入LSTM進(jìn)行序列識(shí)別，在UCF101數(shù)據(jù)集上分別進(jìn)行了5類(lèi)、30類(lèi)、101類(lèi)的視頻分類(lèi)試驗(yàn)，準(zhǔn)確率分別為80%、63%和43%。相較于單一圖片特征，2021年張麗娟、井佩光ADDINEN.CITE<EndNote><Cite><Author>張麗娟</Author><RecNum>126</RecNum><DisplayText><styleface="superscript">[11]</style></DisplayText><record><rec-number>126</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618403760">126</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>張麗娟</author><author>崔天舒</author><author>井佩光</author><author>蘇育挺</author></authors></contributors><auth-address>天津大學(xué)電氣自動(dòng)化與信息工程學(xué)院;</auth-address><titles><title>基于深度多模態(tài)特征融合的短視頻分類(lèi)%J北京航空航天大學(xué)學(xué)報(bào)</title></titles><pages>1-9</pages><keywords><keyword>短視頻</keyword><keyword>多模態(tài)學(xué)習(xí)</keyword><keyword>深度網(wǎng)絡(luò)</keyword><keyword>分類(lèi)</keyword><keyword>特征空間</keyword></keywords><dates></dates><isbn>1001-5965</isbn><call-num>11-2625/V</call-num><urls></urls><remote-database-provider>Cnki</remote-database-provider></record></Cite></EndNote>[\o"張麗娟,#126"11]等人基于音頻模態(tài)采用相似性損失函數(shù)提取特征進(jìn)行視頻分類(lèi)。2021年陳意ADDINEN.CITE<EndNote><Cite><Author>陳意</Author><Year>2021</Year><RecNum>127</RecNum><DisplayText><styleface="superscript">[12]</style></DisplayText><record><rec-number>127</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618403871">127</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>陳意</author><author>黃山</author></authors></contributors><auth-address>四川大學(xué)電氣工程學(xué)院;</auth-address><titles><title>基于改進(jìn)NeXtVLAD的視頻分類(lèi)%J計(jì)算機(jī)工程與設(shè)計(jì)</title></titles><pages>749-754</pages><volume>42</volume><number>03</number><keywords><keyword>深度學(xué)習(xí)</keyword><keyword>視頻分類(lèi)</keyword><keyword>局部聚合描述子向量</keyword><keyword>特征融合</keyword><keyword>卷積神經(jīng)網(wǎng)絡(luò)</keyword></keywords><dates><year>2021</year></dates><isbn>1000-7024</isbn><call-num>11-1775/TP</call-num><urls></urls><remote-database-provider>Cnki</remote-database-provider></record></Cite></EndNote>[\o"陳意,2021#127"12]等人基于改進(jìn)NeXtVLAD在VideoNet-100數(shù)據(jù)集中分類(lèi)準(zhǔn)確率最高為85.4%。在細(xì)分類(lèi)領(lǐng)域，2020年李釗光ADDINEN.CITE<EndNote><Cite><Author>電子測(cè)量技術(shù)</Author><Year>2020</Year><RecNum>128</RecNum><DisplayText><styleface="superscript">[13]</style></DisplayText><record><rec-number>128</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618403939">128</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>李釗光%J電子測(cè)量技術(shù)</author></authors></contributors><titles><title>基于深度學(xué)習(xí)和遷移學(xué)習(xí)的體育視頻分類(lèi)研究</title></titles><pages>27-31</pages><volume>v.43;No.350</volume><number>18</number><dates><year>2020</year></dates><urls></urls></record></Cite></EndNote>[\o"電子測(cè)量技術(shù),2020#128"13]針對(duì)體育視頻提出了基于對(duì)VGG-16模型遷移學(xué)習(xí)大方法，對(duì)其自行采集的10和15類(lèi)體育視頻數(shù)據(jù)集進(jìn)行分類(lèi)，準(zhǔn)確率分別達(dá)到94%和92%，然而僅對(duì)體育視頻分類(lèi)并不能滿(mǎn)足需求。2視頻關(guān)鍵幀選取研究現(xiàn)狀視頻關(guān)鍵幀是將視頻變成視頻幀的形式，并通過(guò)選取若干視頻幀作為整個(gè)視頻信息代表。在一定程度上減少了視頻的冗余度，用較小的內(nèi)存空間和較少的信息特征表示較復(fù)雜的信息內(nèi)容。將較為成熟的圖像研究方法遷移至視頻研究中，視頻關(guān)鍵幀選取是目前對(duì)視頻進(jìn)行深入研究的關(guān)鍵處理步驟，目前視頻關(guān)鍵幀選取的方法可分為：(1)鏡頭檢測(cè)法。一種常用的方法是選取每個(gè)鏡頭中的第一幀選為關(guān)鍵幀ADDINEN.CITE<EndNote><Cite><Author>Cernekova</Author><Year>2005</Year><RecNum>135</RecNum><DisplayText><styleface="superscript">[14]</style></DisplayText><record><rec-number>135</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618485942">135</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Cernekova,Z.</author><author>Pitas,I.</author><author>Nikou,C.%JIEEETransactionsonCircuits</author><author>SystemsforVideoTechnology</author></authors></contributors><titles><title>Informationtheory-basedshotcut/fadedetectionandvideosummarization</title></titles><pages>82-91</pages><volume>16</volume><number>1</number><dates><year>2005</year></dates><urls></urls></record></Cite></EndNote>[\o"Cernekova,2005#135"14]，劉政凱等人ADDINEN.CITE<EndNote><Cite><Author>劉政凱</Author><Year>2002</Year><RecNum>136</RecNum><DisplayText><styleface="superscript">[15]</style></DisplayText><record><rec-number>136</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618486714">136</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>劉政凱</author><author>湯曉鷗%J計(jì)算機(jī)工程與應(yīng)用</author></authors></contributors><titles><title>視頻檢索中鏡頭分割方法綜述</title></titles><pages>84-87</pages><number>23</number><dates><year>2002</year></dates><urls></urls></record></Cite></EndNote>[\o"劉政凱,2002#136"15]先將視頻分成若干個(gè)相鄰視頻幀特征圖無(wú)明顯變化的鏡頭，再?gòu)拿總€(gè)鏡頭中提取關(guān)鍵幀。2019年梁建勝ADDINEN.CITE<EndNote><Cite><Author>梁建勝</Author><Year>2019</Year><RecNum>153</RecNum><DisplayText><styleface="superscript">[16]</style></DisplayText><record><rec-number>153</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618923677">153</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>梁建勝</author><author>溫賀平%J控制工程</author></authors></contributors><titles><title>基于深度學(xué)習(xí)的視頻關(guān)鍵幀提取與視頻檢索</title></titles><pages>965-970</pages><volume>026</volume><number>005</number><dates><year>2019</year></dates><urls></urls></record></Cite></EndNote>[\o"梁建勝,2019#153"16]設(shè)計(jì)一種自適應(yīng)的關(guān)鍵幀選擇算法，在小波變換后，度量小波變換的距離并將其保存為向量，識(shí)別并提取視頻中每個(gè)鏡頭的摘要信息，選取包含最多顯著特征的幀作為該鏡頭的關(guān)鍵幀。KucuktuncOADDINEN.CITE<EndNote><Cite><Author>Kuecuektunc</Author><Year>2010</Year><RecNum>155</RecNum><DisplayText><styleface="superscript">[17]</style></DisplayText><record><rec-number>155</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618924274">155</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Kuecuektunc,Onur</author><author>Gueduekbay,Ugur</author><author>Ulusoy,Oezguer%JComputerVision</author><author>ImageUnderstanding</author></authors></contributors><titles><title>Fuzzycolorhistogram-basedvideosegmentation</title></titles><pages>125-134</pages><volume>114</volume><number>1</number><dates><year>2010</year></dates><urls></urls></record></Cite></EndNote>[\o"Kuecuektunc,2010#155"17]提出一種基于模糊顏色直方圖的鏡頭分割方法，除了檢測(cè)突變、漸變鏡頭，還可以對(duì)掉幀、加字幕、加圖標(biāo)等較難識(shí)別視頻幀進(jìn)行檢測(cè)。鏡頭檢測(cè)法具有計(jì)算簡(jiǎn)單且易于實(shí)現(xiàn)的特點(diǎn)，然而提取的關(guān)鍵幀一般為鏡頭突變或漸變處的視頻幀，既無(wú)法處理關(guān)鍵幀冗余問(wèn)題，又無(wú)法充分描述視頻內(nèi)容。(2)特征提取法。在對(duì)視頻幀圖像提取如顏色、亮度、紋理等底層特征后，計(jì)算相鄰兩幀圖像特征之間的差值，通過(guò)比較特征差值是否大于預(yù)設(shè)閾值來(lái)確定是否為關(guān)鍵幀，ZhangADDINEN.CITE<EndNote><Cite><Author>Hong</Author><Year>1997</Year><RecNum>137</RecNum><DisplayText><styleface="superscript">[18]</style></DisplayText><record><rec-number>137</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618487742">137</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Hong</author><author>Jiang</author><author>Zhang</author><author>and</author><author>Jianhua</author><author>Wu</author><author>and</author><author>Di</author><author>Zhong</author><author>and%JPatternRecognition</author></authors></contributors><titles><title>Anintegratedsystemforcontent-basedvideoretrievalandbrowsing</title></titles><dates><year>1997</year></dates><urls></urls></record></Cite></EndNote>[\o"Hong,1997#137"18]和GunselADDINEN.CITE<EndNote><Cite><Author>Gunsel</Author><Year>2002</Year><RecNum>138</RecNum><DisplayText><styleface="superscript">[19]</style></DisplayText><record><rec-number>138</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618487815">138</key></foreign-keys><ref-typename="ConferenceProceedings">10</ref-type><contributors><authors><author>Gunsel,B.</author><author>Tekalp,A.M.</author></authors></contributors><titles><title>Content-basedvideoabstraction</title><secondary-title>InternationalConferenceonImageProcessing</secondary-title></titles><dates><year>2002</year></dates><urls></urls></record></Cite></EndNote>[\o"Gunsel,2002#138"19]等人提出了基于顏色直方圖的方法。因?yàn)橐獙?duì)視頻中的每幅圖像進(jìn)行一種或多種特征的提取，所以計(jì)算量比較大。(3)運(yùn)動(dòng)分析法。首先將視頻分成若干個(gè)鏡頭，然后通過(guò)光流法計(jì)算每個(gè)鏡頭的運(yùn)動(dòng)量，當(dāng)運(yùn)動(dòng)量取局部最小值時(shí)，選擇該鏡頭中對(duì)應(yīng)的幀作為最終的關(guān)鍵幀。FanLADDINEN.CITE<EndNote><Cite><Author>Fan</Author><Year>2013</Year><RecNum>154</RecNum><DisplayText><styleface="superscript">[20]</style></DisplayText><record><rec-number>154</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618923852">154</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Fan,L.</author><author>Wu,Q.</author><author>Ruan,C.</author><author>Zhuo,Z.</author><author>Wang,X.%JIEEProceedings-VisionImage</author><author>SignalProcessing</author></authors></contributors><titles><title>Afeatureextractionalgorithmbasedon2Dcomplexityofgaborwaveletstransformforfacialexpressionrecognition</title></titles><dates><year>2013</year></dates><urls></urls></record></Cite></EndNote>[\o"Fan,2013#154"20]采用運(yùn)動(dòng)向量特征作為視頻關(guān)鍵幀特征，由于特征提取時(shí)間長(zhǎng)，故不適用于大規(guī)模的視頻相似性檢測(cè)。(4)聚類(lèi)選取法。結(jié)合了圖像全局或局部特征，利用特定的相似度或相異度將處理對(duì)象進(jìn)行分組，具有相似特征劃分到同一個(gè)簇中，再?gòu)母鞔刂羞x取關(guān)鍵幀。Wu等人ADDINEN.CITE<EndNote><Cite><Author>Wu</Author><Year>2017</Year><RecNum>140</RecNum><DisplayText><styleface="superscript">[21]</style></DisplayText><record><rec-number>140</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618489014">140</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>JiaxinWu</author><author>Sheng-huaZhong</author><author>JianminJiang</author><author>YunyunYang</author></authors></contributors><auth-address>ShenzhenUniversity;;ShenzhenUniversity;;ShenzhenUniversity;;HarbinInstituteofTechnologyShenzhenGraduateSchool</auth-address><titles><title>Anovelclusteringmethodforstaticvideosummarization%JMultimediaToolsandApplications</title></titles><volume>76</volume><number>7</number><keywords><keyword>Staticvideosummarization</keyword><keyword>Clusteringmethod</keyword><keyword>Videorepresentation</keyword></keywords><dates><year>2017</year></dates><isbn>1380-7501</isbn><urls></urls><remote-database-provider>Cnki</remote-database-provider></record></Cite></EndNote>[\o"Wu,2017#140"21]先利用特征向量的奇異值分解來(lái)進(jìn)行預(yù)采樣，減少給定視頻的冗余信息，再用詞袋模型來(lái)表示候選幀的視覺(jué)內(nèi)容，最后利用基于視頻表示的高密度峰值聚類(lèi)算法選取關(guān)鍵幀。吳先宇A(yù)DDINEN.CITE<EndNote><Cite><Author>吳先宇</Author><Year>2018</Year><RecNum>152</RecNum><DisplayText><styleface="superscript">[22]</style></DisplayText><record><rec-number>152</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618923341">152</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>吳先宇</author><author>王之杰</author><author>石蘊(yùn)玉</author><author>吳寅騰</author><author>張一鳴</author><author>程凱亮%J福建電腦</author></authors></contributors><titles><title>廣告視頻關(guān)鍵幀提取軟件的設(shè)計(jì)與開(kāi)發(fā)</title></titles><pages>130-131+177</pages><volume>v.34</volume><number>04</number><dates><year>2018</year></dates><urls></urls></record></Cite></EndNote>[\o"吳先宇,2018#152"22]針對(duì)廣告視頻通過(guò)舍棄過(guò)亮過(guò)暗的視頻幀，結(jié)合均值哈希算法、聚類(lèi)查重，將一段廣告視頻濃縮為十幾或幾十張圖像。3視頻相似性判定及檢測(cè)算法研究現(xiàn)狀(1)視頻相似性判定視頻相似性檢測(cè)包括重復(fù)視頻檢測(cè)、近似重復(fù)視頻ADDINEN.CITE<EndNote><Cite><Author>Wu</Author><Year>2007</Year><RecNum>158</RecNum><DisplayText><styleface="superscript">[23]</style></DisplayText><record><rec-number>158</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618968978">158</key></foreign-keys><ref-typename="ConferenceProceedings">10</ref-type><contributors><authors><author>Wu,X.</author><author>Hauptmann,A.G.</author><author>Ngo,C.W.</author><author>Wu,P.</author></authors></contributors><titles><title>Practicaleliminationofnear-duplicatesfromwebvideosearch</title><secondary-title>InternationalconferenceonMultimedia</secondary-title></titles><dates><year>2007</year></dates><urls></urls></record></Cite></EndNote>[\o"Wu,2007#158"23]檢測(cè)和拷貝視頻檢測(cè)。重復(fù)視頻即為幾乎一模一樣的視頻，范圍較窄；近似重復(fù)視頻，要求語(yǔ)義一致、畫(huà)面近似，視頻來(lái)源一般不同；而拷貝視頻，要求語(yǔ)義一致、畫(huà)面近似且視頻來(lái)源相同ADDINEN.CITE<EndNote><Cite><Author>顧佳偉</Author><Year>2017</Year><RecNum>161</RecNum><DisplayText><styleface="superscript">[24]</style></DisplayText><record><rec-number>161</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618971857">161</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>顧佳偉</author><author>趙瑞瑋</author><author>姜育剛%J計(jì)算機(jī)研究與發(fā)展</author></authors></contributors><titles><title>視頻拷貝檢測(cè)方法綜述</title></titles><pages>1238-1250</pages><volume>54</volume><number>006</number><dates><year>2017</year></dates><urls></urls></record></Cite></EndNote>[\o"顧佳偉,2017#161"24]。例如，學(xué)生用各自的手機(jī)分別記錄同一時(shí)間段內(nèi)教師授課狀態(tài)及授課內(nèi)容，這2個(gè)視頻視為近似重復(fù)視頻而不是拷貝視頻；如果一名同學(xué)對(duì)另一名同學(xué)拍攝的視頻進(jìn)行后期加工，如加入字幕或貼紙?jiān)?，則新視頻才被視為拷貝視頻。早期認(rèn)為拷貝檢測(cè)與近似重復(fù)檢測(cè)有明顯的差異，后來(lái)，Basharat等人ADDINEN.CITE<EndNote><Cite><Author>Basharat</Author><Year>2008</Year><RecNum>159</RecNum><DisplayText><styleface="superscript">[25]</style></DisplayText><record><rec-number>159</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618969197">159</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Basharat,A.</author><author>Yun,Z.</author><author>Shah,M.%JComputerVision</author><author>ImageUnderstanding</author></authors></contributors><titles><title>Contentbasedvideomatchingusingspatiotemporalvolumes</title></titles><pages>360-377</pages><volume>110</volume><number>3</number><dates><year>2008</year></dates><urls></urls></record></Cite></EndNote>[\o"Basharat,2008#159"25]建議放寬定義，本文視頻相似性檢測(cè)包括如上視頻檢測(cè)，包括在文件格式、編碼參數(shù)、光度變化(顏色、光線變化)、編輯操作(標(biāo)題、logo和邊框插入)、不同長(zhǎng)度和某些修改(幀添加/刪除)方面的不同，用戶(hù)會(huì)清楚地識(shí)別出這些視頻“本質(zhì)上是相同的”。針對(duì)視頻相似性檢測(cè)所使用的檢測(cè)方法，不同研究者選擇的方法存在共通性ADDINEN.CITE<EndNote><Cite><Author>Liu</Author><Year>2013</Year><RecNum>160</RecNum><DisplayText><styleface="superscript">[26]</style></DisplayText><record><rec-number>160</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618971185">160</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Liu,J.</author><author>Huang,Z.I.</author><author>Cai,H.</author><author>Shen,H.T.</author><author>Chong,W.N.</author><author>Wang,W.%JIEEEMultimedia</author></authors></contributors><titles><title>Near-DuplicateVideoRetrieval:CurrentResearchandFutureTrends</title></titles><pages>1-1</pages><volume>45</volume><number>4</number><dates><year>2013</year></dates><urls></urls></record></Cite></EndNote>[\o"Liu,2013#160"26]。(2)視頻指紋提取視頻指紋是指基于視頻內(nèi)容形成的簽名，可以專(zhuān)門(mén)用來(lái)表示某一個(gè)視頻ADDINEN.CITE<EndNote><Cite><Author>Kim</Author><Year>2014</Year><RecNum>133</RecNum><DisplayText><styleface="superscript">[27]</style></DisplayText><record><rec-number>133</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618484899">133</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>SeminKim</author><author>SeungHoLee</author><author>YongManRo</author></authors></contributors><auth-address>ImageandVideoSystemsLab,Dept.ofElectricalEngineering,KoreaAdvancedInstituteofScienceandTechnology(KAIST),Yuseong-Gu,Daejeon305-701,RepublicofKorea</auth-address><titles><title>Rotationandflippingrobustregionbinarypatternsforvideocopydetection%JJournalofVisualCommunicationandImageRepresentation</title></titles><volume>25</volume><number>2</number><keywords><keyword>Regionbinarypattern</keyword><keyword>Rotationrobustpattern</keyword><keyword>Flippingrobustpattern</keyword><keyword>Videocopydetection</keyword><keyword>Videofingerprint</keyword><keyword>Localbinarypattern</keyword><keyword>ImageDescriptor</keyword><keyword>VideoDescriptor</keyword></keywords><dates><year>2014</year></dates><isbn>1047-3203</isbn><urls></urls><remote-database-provider>Cnki</remote-database-provider></record></Cite></EndNote>[\o"Kim,2014#133"27]。要在視頻數(shù)據(jù)庫(kù)中查找某個(gè)視頻的相似視頻，可以在相應(yīng)的指紋數(shù)據(jù)庫(kù)中搜索其指紋并進(jìn)行匹配。兩個(gè)指紋的緊密性代表了相應(yīng)視頻之間的相似性，兩個(gè)感知上不同的視頻應(yīng)該有不同的指紋ADDINEN.CITE<EndNote><Cite><Author>Zahedi</Author><Year>2015</Year><RecNum>134</RecNum><DisplayText><styleface="superscript">[28]</style></DisplayText><record><rec-number>134</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618485046">134</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Zahedi,M.</author><author>Ghadi,O.R.%JSignalImage</author><author>VideoProcessing</author></authors></contributors><titles><title>CombiningGaborfilterandFFTforfingerprintenhancementbasedonaregionaladaptionmethodandautomaticsegmentation</title></titles><pages>267-275</pages><volume>9</volume><number>2</number><dates><year>2015</year></dates><urls></urls></record></Cite></EndNote>[\o"Zahedi,2015#134"28]。視頻指紋提取可以分為三種：基于空域提取視頻指紋、基于時(shí)域提取視頻指紋和基于時(shí)空結(jié)合生成視頻指紋。1)基于空域提取視頻指紋。以視頻的圖像特征為基礎(chǔ)提取視頻指紋，如形狀、顏色、紋理等特征作為視頻的唯一身份證明。E.MADDINEN.CITE<EndNote><Cite><Author>Maani</Author><Year>2008</Year><RecNum>141</RecNum><DisplayText><styleface="superscript">[29]</style></DisplayText><record><rec-number>141</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618493082">141</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Maani,E.</author><author>Tsaftaris,S.A.</author><author>Katsaggelos,A.K.%JIEEE</author></authors></contributors><titles><title>Localfeatureextractionforvideocopydetectioninadatabase</title></titles><dates><year>2008</year></dates><urls></urls></record></Cite></EndNote>[\o"Maani,2008#141"29]基于幀運(yùn)動(dòng)強(qiáng)度差異提取關(guān)鍵幀，以一個(gè)矢量來(lái)描述K.MADDINEN.CITE<EndNote><Cite><Author>Mikolajczyk</Author><Year>2005</Year><RecNum>142</RecNum><DisplayText><styleface="superscript">[30]</style></DisplayText><record><rec-number>142</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618493164">142</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Mikolajczyk,K.</author><author>Schmid,C.%JIEEETransactionsonPatternAnalysis</author><author>MachineIntelligence</author></authors></contributors><titles><title>APerformanceEvaluationofLocalDescriptors</title></titles><pages>1615-1630</pages><volume>27</volume><number>10</number><dates><year>2005</year></dates><urls></urls></record></Cite></EndNote>[\o"Mikolajczyk,2005#142"30]提取的興趣點(diǎn)，使用DP技術(shù)降低計(jì)算復(fù)雜度，對(duì)低比特率壓縮魯棒性較好。A.SarADDINEN.CITE<EndNote><Cite><Author>Sarkar</Author><Year>2008</Year><RecNum>143</RecNum><DisplayText><styleface="superscript">[31]</style></DisplayText><record><rec-number>143</rec-number><foreign-keys><keyapp="EN"db-id="59px5dpxe0e0dpe2fs6ps5xgetrpwxvf2p0w"timestamp="1618493382">143</key></foreign-keys><ref-typename="ConferenceProceedings">10</ref-type><contributors><authors><author>Sarkar,A.</author><author>Ghosh,P.</author><author>Moxley,E.</author><author>Manjunath,B.S.</author></authors></contributors><titles><title>VideoFingerprinting:FeaturesforDuplicateandSimilarVideoDetectionandQuery-basedVideoRetrieval</title><secondary-title>MultimediaContentAccess:AlgorithmsandSystemsII</secondary-title></titles><dates><year>2008</year></dates><urls></urls></record></Cite></EndNote>[\o"Sarkar,2008#143"31]提取CFMT(CompactFourier-MellinTransform)、SIFT、亮度色度直方圖和有序直方圖等特征，指出CFMT用于復(fù)制視頻檢測(cè)時(shí)效果比SIFT好，但相似視頻檢索時(shí)SIFT比CFMT效果更佳。2）基于時(shí)域提取視頻指紋。以運(yùn)動(dòng)特征為基礎(chǔ)提取視頻指紋，如亮度變化當(dāng)作能辨別視頻身份的證明。LiChenADDINEN.CITE<EndNote><Cite><Author>Chen</Author><Year>2008</Year>