【摘要】：各種數(shù)字圖像、視頻編輯軟件以及相應(yīng)教程的普及,導(dǎo)致圖像和視頻可能成為作惡者的工具,人們長久以來對影像媒體的信任也發(fā)生了動搖。數(shù)字圖像和視頻篡改檢測的相關(guān)理論和技術(shù)就是在這種背景下應(yīng)運(yùn)而生。數(shù)字圖像和視頻篡改檢測屬于媒體內(nèi)容取證范疇,目的是鑒別影像媒體內(nèi)容的真?zhèn)巍?shù)字圖像和視頻從拍攝到成像的過程中,場景中感知對象反射的光線需要經(jīng)過光學(xué)鏡頭的折射,并經(jīng)歷光學(xué)濾波、光電轉(zhuǎn)換、色彩插值、后處理等操作,這個(gè)過程中的每一個(gè)步驟可以看做是對光線的一次變換,本文將一組變換的有序組合定義為一個(gè)變換鏈,并將數(shù)字圖像和視頻建模為經(jīng)歷了某個(gè)變換鏈的若干感知對象的結(jié)構(gòu)化組合,進(jìn)而對圖像和視頻的篡改手段進(jìn)行了形式化的分析,并將篡改行為在媒體中留下的痕跡歸結(jié)為兩種情況:一是媒體中出現(xiàn)異常相似的感知對象,二是某些感知對象經(jīng)歷了與其他感知對象不同的變換鏈。在此基礎(chǔ)上,本文將圖像及視頻的篡改檢測建模為一個(gè)“描述——發(fā)現(xiàn)”的過程:對于待檢測的媒體,首先找到某種特征,用以描述媒體中的感知對象或感知對象所經(jīng)歷的變換鏈中的某個(gè)環(huán)節(jié),進(jìn)而通過匹配或校驗(yàn)的方式,去發(fā)現(xiàn)媒體中異常相似的感知對象,或去發(fā)現(xiàn)某些感知對象經(jīng)歷了與其他感知對象不一致的變換鏈。本文圍繞篡改檢測模型中的特征構(gòu)造、匹配和校驗(yàn)方法等關(guān)鍵技術(shù)展開研究,解決了數(shù)字圖像及視頻篡改檢測技術(shù)中存在的若干問題。本文的主要研究工作和創(chuàng)新點(diǎn)包括以下幾個(gè)方面:第一,提出了一種基于有序序列聚類的特征匹配方法。特征匹配是檢測媒體中異常相似對象的關(guān)鍵技術(shù)之一,在基于特征點(diǎn)的圖像區(qū)域拷貝檢測方法中,當(dāng)特征空間中同時(shí)存在多個(gè)高度相似的特征時(shí),現(xiàn)有的特征匹配方法會漏檢大量實(shí)際上匹配的特征對。針對該問題,本文提出了基于有序序列聚類的特征匹配方法,并基于貝葉斯分類器實(shí)現(xiàn)了聚類過程中參數(shù)的自適應(yīng)選擇,顯著地提高了合格匹配特征收集的完備性。第二,針對基于特征點(diǎn)的圖像區(qū)域拷貝檢測方法對平滑區(qū)域拷貝行為檢測能力弱的問題,提出了層次化特征點(diǎn)檢測結(jié)合特征融合的區(qū)域拷貝檢測方法。該方法在不顯著增加特征點(diǎn)總數(shù)的情況下,保證不同區(qū)域內(nèi)特征點(diǎn)的覆蓋率。對于平滑區(qū)域的特征點(diǎn),本文構(gòu)造了基于局部梯度和色彩的融合特征,提高了局部特征在平滑區(qū)域內(nèi)的區(qū)分性。第三,提出了基于DCT系數(shù)分析的壓縮歷史不一致檢測所應(yīng)滿足的邊界條件,并設(shè)計(jì)了相應(yīng)的參數(shù)求解方法。圖像中各感知對象的壓縮歷史不一致即某些感知對象在壓縮編碼階段所經(jīng)歷的變換與其他感知對象不同,這意味著圖像有局部區(qū)域遭到了篡改。在基于DCT系數(shù)分析的JPEG壓縮歷史不一致檢測方法中,通常以篡改成分和非篡改成分的DCT系數(shù)分布作為特征來反映各DCT塊所經(jīng)歷的壓縮歷史。而篡改成分和非篡改成分的DCT系數(shù)分布的估計(jì)則依賴于對二者的混合模型的參數(shù)求解�，F(xiàn)有的方法普遍以“盲”的方式進(jìn)行參數(shù)估計(jì),因而對參數(shù)的估值往往不夠準(zhǔn)確。考慮到DCT系數(shù)所應(yīng)遵循的實(shí)際約束,本文向混合模型對應(yīng)的似然函數(shù)中補(bǔ)充了必要的邊界條件,并結(jié)合似然函數(shù)的平滑特性設(shè)計(jì)了粗粒度搜索結(jié)合梯度上升的參數(shù)估計(jì)方法,實(shí)現(xiàn)了更為準(zhǔn)確的篡改檢測和定位。第四,針對降質(zhì)視頻,提出了基于位置敏感哈希和幀配準(zhǔn)的匹配方法。在降質(zhì)視頻中,受各種降質(zhì)因素的影響,視頻幀局部結(jié)構(gòu)細(xì)微變化的累積將導(dǎo)致幀特征產(chǎn)生實(shí)質(zhì)性的變化。因此傳統(tǒng)方法中普遍采用的“特征提取——閾值化”的匹配方法很難實(shí)現(xiàn)穩(wěn)定的雷同幀檢測。本文基于位置敏感哈希實(shí)現(xiàn)視頻幀序列的初步匹配,并基于配準(zhǔn)技術(shù)完成雷同幀的校驗(yàn)。為了實(shí)現(xiàn)對視頻降質(zhì)的魯棒性,本文將視頻幀中各區(qū)域的穩(wěn)定性信息編碼到配準(zhǔn)能量函數(shù)中,并基于概率推理的方式近似求解全局最優(yōu)匹配問題,實(shí)現(xiàn)了更為魯棒的降質(zhì)視頻幀拷貝檢測。第五,提出了針對高碼率視頻的快速幀拷貝檢測方法。在高碼率視頻中,當(dāng)具有相同內(nèi)容的幀之間并不存在顯著的波動時(shí),為了在保證檢測能力的前提下降低幀拷貝檢測的時(shí)間開銷,本文提出了二維視頻幀的三維骨架特征并設(shè)計(jì)了相應(yīng)的匹配方法。本文首先基于骨架的拓?fù)湫畔?shí)現(xiàn)數(shù)據(jù)篩選,進(jìn)而基于幾何信息進(jìn)行細(xì)粒度的雷同幀判別,實(shí)現(xiàn)了高碼率視頻中的快速幀拷貝檢測。第六,提出了基于碼流異常突變的視頻刪/插幀檢測方法�；诖a流分析的刪/插幀檢測方法通常以碼流中存在異常的周期效應(yīng)作為碼流異常的特征。然而,異常的周期效應(yīng)并非總能夠可靠地檢測到。此外,現(xiàn)有的方法不能有效地定位篡改操作發(fā)生的位置。本文把碼流中各P幀對應(yīng)的預(yù)測殘差均值和幀內(nèi)預(yù)測宏塊數(shù)量的同時(shí)突變作為檢測刪/插幀操作的依據(jù),設(shè)計(jì)了用于度量預(yù)測殘差均值和幀內(nèi)預(yù)測宏塊數(shù)量的變化強(qiáng)度的指標(biāo),進(jìn)而基于這兩種指標(biāo)構(gòu)造了融合特征,以及用于捕捉碼流信息突變的校驗(yàn)方法。在對視頻編碼參數(shù)不做任何約束的情況下,本文的方法能夠有效地檢測和定位視頻幀插入/刪除操作。
[Abstract]:The popularity of various digital images, video editing software, and corresponding tutorials has led to the potential for images and videos to become the tools of the perpetrator, and the trust of the image media has been shaken for a long time. The related theories and techniques of digital image and video tamper detection are in this background. The digital image and the video tampering detection belong to the category of the media content forensic, and the purpose of the invention is to identify the authenticity of the image media content. in the process of shooting the digital image and the video from the shooting to the image, the light reflected by the sensing object in the scene needs to pass through the refraction of the optical lens and undergo the operations of optical filtering, photoelectric conversion, color interpolation, post-processing and the like, each step in this process can be seen as a transformation of light, which is defined herein as a transform chain and the digital image and the video are modeled as a structured combination of several perceptual objects that have undergone a transformation chain, In this paper, a formal analysis of the tampering of the image and the video is carried out, and the trace of the alteration behavior in the media is summed up as two cases: one is a perception object in the media which is similar to the other, and the other is that some of the perceptive objects experience different transformation chains than the other perceptual objects. On the basis of this, the process of modeling the tampering detection of the image and the video is a "Description _ Discovery" process: for the media to be detected, a certain feature is first found to describe a link in the transformation chain experienced by the perceived object or the perceived object in the media, Furthermore, by matching or checking, the perceptually similar perceived objects in the media are found, or some sense objects are found to experience a transformation chain that is not consistent with other perceived objects. In this paper, the key technologies such as feature structure, matching and checking method in the tamper detection model are studied, and some problems in the detection technology of digital image and video tampering are solved. The main research work and innovation points of this paper include the following aspects: first, a feature matching method based on ordered sequence clustering is proposed. The feature matching is one of the key technologies for detecting the abnormal similar objects in the media. In the method of image area copy detection based on the feature point, when a plurality of highly similar features exist in the feature space, the existing feature matching method can miss a large number of actually matching feature pairs. In order to solve this problem, the feature matching method based on the ordered sequence clustering is proposed, and the self-adaptive selection of the parameters in the clustering process is realized based on the Bayesian classifier, and the completeness of the collection of the qualified matching features is remarkably improved. Secondly, aiming at the problem that the image area copy detection method based on the feature point is weak in the detection capability of the smooth region copy behavior, a method for detecting the region copy of the hierarchical feature point detection and combining feature fusion is proposed. The method ensures the coverage rate of the characteristic points in different regions without significantly increasing the total number of the feature points. For the feature point of the smooth region, this paper constructs the fusion feature based on the local gradient and the color, and improves the distinguishing between the local feature and the smooth region. Thirdly, the boundary conditions to be satisfied by the non-uniform detection of the compression history based on the DCT coefficient analysis are put forward, and the corresponding method for solving the parameters is designed. The compression history of each of the perceptual objects in the image is not consistent, that is, the transformation experienced by some of the perceptual objects during the compression encoding stage is different from the other perceptual objects, which means that the image has a local area being tampered with. In the JPEG compression history inconsistency detection method based on the DCT coefficient analysis, the compression history experienced by each DCT block is generally reflected by the DCT coefficient distribution of the tamper component and the non-tamper component as a feature. The estimation of the DCT coefficient distribution of the tampered component and the non-tamper component is dependent on the parameter solving of the mixed model of the two. The existing methods are generally used in "blind" to estimate parameters, and therefore the valuation of the parameters is often not accurate. Considering the actual constraints to be followed by the DCT coefficients, the necessary boundary conditions are added to the likelihood function corresponding to the mixed model, and the method of parameter estimation of the coarse-size search combined with the gradient increase is designed in combination with the smoothness of the likelihood function. And the more accurate detection and positioning of the tampering is realized. Fourth, aiming at the quality reduction video, a matching method based on location sensitive hash and frame registration is proposed. In the quality reduction video, the accumulation of fine changes in the local structure of the video frame will lead to a substantial change in the frame characteristics due to the influence of various quality-reducing factors. Therefore, the conventional "feature extraction _ thresholding" matching method in the traditional method is difficult to realize the stable thunder and the frame detection. In this paper, the preliminary matching of the video frame sequence is realized based on the location-sensitive hash, and the verification of the same frame is completed based on the registration technique. In order to achieve the robustness of the video quality reduction, the stability information of each region in the video frame is encoded into the registration energy function, and the global optimal matching problem is solved based on the way of probability reasoning, and a more robust quality-reducing video frame copy detection is realized. Fifth, a fast frame copy detection method for high code rate video is proposed. In the high code rate video, when there is no significant fluctuation between the frames with the same content, in order to reduce the time cost of the frame copy detection under the premise of ensuring the detection capability, the three-dimensional skeleton feature of the two-dimensional video frame is proposed and the corresponding matching method is designed. In this paper, based on the topology information of the skeleton, the data selection is realized, and the fine-grained Lei-frame discrimination is carried out based on the geometric information, and the fast frame copy detection in the high-code-rate video is realized. And sixth, a method for detecting the video deletion/ insertion frame based on the abnormal mutation of the code stream is provided. The method of deleting/ inserting the frame based on the code stream analysis usually takes the periodic effect of the abnormality in the code stream as the characteristic of the code stream abnormality. However, the periodic effect of the abnormality is not always able to be detected reliably. In addition, the existing method cannot effectively locate the position of the tampering operation. in that invention, the mean value of the prediction residual corresponding to the P-frame in the code stream and the simultaneous mutation of the number of intra-prediction macro-block are taken as the basis for detecting the deletion/ interpolation frame operation, and the index for measuring the variation intensity of the mean value of the prediction residual and the number of the intra-prediction macro-blocks is designed, And then the fusion characteristic is constructed based on the two indexes, and the verification method for capturing the code stream information mutation is realized. In the case of no constraint to the video coding parameters, the method of the present invention can effectively detect and position the video frame insertion/ deletion operation.
【學(xué)位授予單位】：哈爾濱工業(yè)大學(xué)
【學(xué)位級別】：博士
【學(xué)位授予年份】：2016
【分類號】：TP391.41

【相似文獻(xiàn)】

相關(guān)期刊論文 前10條

1 駱偉祺;黃繼武;丘國平;;魯棒的區(qū)域復(fù)制圖像篡改檢測技術(shù)[J];計(jì)算機(jī)學(xué)報(bào);2007年11期

2 張雯;李學(xué)明;;改進(jìn)的基于顏色濾波陣列特性的篡改檢測[J];計(jì)算機(jī)工程與應(yīng)用;2009年06期

3 王鑫;魯志波;;數(shù)字圖像拷貝移動篡改檢測方法的分析[J];計(jì)算機(jī)工程與設(shè)計(jì);2009年12期

4 周治平;張小祥;;基于質(zhì)量評價(jià)量和方差分析的圖像篡改檢測[J];計(jì)算機(jī)工程;2011年09期

5 馬巧梅;鄧啟森;;基于交叉劃分的灰度圖像篡改檢測與恢復(fù)算法[J];微電子學(xué)與計(jì)算機(jī);2014年09期

6 張震;蘇白娜;喻宙;;一種新的圖像拷貝篡改檢測方法[J];鄭州大學(xué)學(xué)報(bào)(工學(xué)版);2012年06期

7 黃添強(qiáng);陳智文;蘇立超;鄭之;袁秀娟;;利用內(nèi)容連續(xù)性的數(shù)字視頻篡改檢測[J];南京大學(xué)學(xué)報(bào)(自然科學(xué)版);2011年05期

8 黃添強(qiáng);曾文賦;;基于壓縮感知的視頻異源篡改檢測[J];計(jì)算機(jī)工程與應(yīng)用;2011年30期

9 周龍龍;;圖像空域分析下的篡改檢測算法[J];電視技術(shù);2012年23期

10 李曉梅;;基于殘留噪聲相關(guān)性的視頻篡改檢測算法[J];信息技術(shù);2013年10期

相關(guān)會議論文 前2條

1 孔輝;梁洪亮;辛陽;楊義先;陳林順;;一種高性能的網(wǎng)頁篡改檢測與恢復(fù)機(jī)制[A];2010年全國通信安全學(xué)術(shù)會議論文集[C];2010年

2 韓杰思;朱中梁;沈建京;彭韶峰;;針對拼接圖像篡改的檢測技術(shù)[A];第十四屆全國圖象圖形學(xué)學(xué)術(shù)會議論文集[C];2008年

相關(guān)博士學(xué)位論文 前3條

1 楊本娟;圖像篡改檢測中的投影核方法[D];貴州大學(xué);2015年

2 于立洋;數(shù)字圖像及視頻篡改檢測關(guān)鍵技術(shù)研究[D];哈爾濱工業(yè)大學(xué);2016年

3 丁琦;數(shù)字音頻篡改檢測與隱寫分析技術(shù)研究[D];解放軍信息工程大學(xué);2011年

相關(guān)碩士學(xué)位論文 前10條

1 張璐波;基于運(yùn)動目標(biāo)的視頻幀內(nèi)篡改檢測算法研究與實(shí)現(xiàn)[D];上海交通大學(xué);2015年

2 劉躍強(qiáng);用于提高篡改檢測精度的半脆弱性數(shù)字水印技術(shù)研究[D];電子科技大學(xué);2014年

3 林志新;基于內(nèi)容的視頻篡改檢測系統(tǒng)設(shè)計(jì)與實(shí)現(xiàn)[D];電子科技大學(xué);2014年

4 孫艷茹;數(shù)字圖像區(qū)域復(fù)制粘貼篡改檢測算法分析與研究[D];黑龍江大學(xué);2015年

5 鄧峰;網(wǎng)頁防篡改中篡改檢測技術(shù)的研究[D];西安工業(yè)大學(xué);2013年

6 吳俞醒;基于連續(xù)性特征的視頻幀間篡改檢測算法的研究與實(shí)現(xiàn)[D];上海交通大學(xué);2015年

7 李君;基于客戶端的網(wǎng)頁篡改檢測研究[D];浙江工業(yè)大學(xué);2015年

8 王春井;基于網(wǎng)絡(luò)爬蟲與HASH的網(wǎng)站篡改檢測系統(tǒng)的設(shè)計(jì)與實(shí)現(xiàn)[D];內(nèi)蒙古大學(xué);2016年

9 曾文賦;視頻數(shù)據(jù)特征降維及其在篡改檢測中的應(yīng)用[D];福建師范大學(xué);2011年

10 蘇立超;壓縮感知在數(shù)字視頻篡改檢測中的應(yīng)用研究[D];福建師范大學(xué);2014年

，

本文編號：2487512