Robust Block Selection for Watermarking Video Streams

Transcription

1 Proceedngs of the World Congress on Engneerng 8 Vol I WCE 8, July - 4, 8, London, U.K. Robust Block Selecton for Watermarkng Vdeo Streams Tamás Polyák, Gábor Fehér Abstract In ths paper a robust vdeo atermarkng method s presented, hch embeds data n the avelet doman usng edge detecton. The algorthm uses the lumnance values around the edges here changes are less notceable for the human vsual system. Watermark s embedded n an addtve ay usng spread spectrum nose. The complexty of the method s lo. The algorthms used for avelet transformaton and edge detecton are fast and smple. Results sho that the atermark sgnal s mperceptble, and the method s robust aganst lo btrate lossy compressons, lke H.64 and XvD. Index Terms Edge detecton, spread spectrum, vdeo stream, avelet transform, atermarkng I. INTRODUCTION IGITAL vdeo streamng s a more and more popular Dservce among content provders. These streams can be dsplayed on varous types of devces: computer, PDA etc. The content can be easly dstrbuted through the nternet, but dgtal contents carry a bg securty rsk,.e. the copyng and reproducton of the content s very easy and effortless. Even users thout specal knoledge can easly share the donloaded content th other people. To avod llegtmate access, dgtal content s often encrypted, and travels n an encrypted form to the consumer. Although encrypton secures the content on the ay to the consumer, durng playback t must be decrypted, and ths s the pont here t exposed to llegal copyng. In these cases atermarks can be used to gve some extra protecton to the content, snce atermarkng can embed addtonal nformaton n the dgtal content. Watermarks are embedded n the dgtal content n an mperceptble ay. Watermarks can carry some nformaton about the content: oner of the content, metadata, QoS parameters etc. Ths protecton does not elmnate the need for encrypton, but s a supplemental technque to secure the content, or store addtonal data. These atermarks are called robust atermarks because they must survve transformatons of the underlyng content e.g. lossy compresson. Dgtal atermarkng algorthms are generally dvded nto Manuscrpt receved March, 8. T. Polyák s th the Department of Telecommuncatons and Meda Informatcs, Budapest Unversty of Technology and Economcs, Magyar tudósok krt.., 7 Budapest, Hungary (e-mal: tpolyak@tmt.bme.hu). G. Fehér s th the Department of Telecommuncatons and Meda Informatcs, Budapest Unversty of Technology and Economcs, Magyar tudósok krt.., 7 Budapest, Hungary (e-mal: feher@tmt.bme.hu). to groups: algorthms that hde data n the spatal doman. It means that nformaton s embedded by modfyng the pxel values drectly []-[3], and algorthms that use a transform doman for data hdng. Dscrete cosne transform (DCT) and dscrete avelet transform (DWT) are often used at transform doman atermarkng [4]-[7]. These atermarkng technques modfy coeffcents of the gven transform doman. Wavelet transform s commonly used because t has many advantages over DCT transform. It s closer to the human vsual system (HVS), nstead of processng 8 x 8 pxel blocks t processes the hole frame. It dvdes the frame nto four parts th dfferent frequences (LL, LH, HL and HH) and drectons. Some atermarkng technques embed data n object borders, and other perceptually mportant areas. It has several advantages: the HVS s less senstve to changes made n these components, and modern lossy compresson algorthms leave them relatvely untouched to mantan hgh vdeo qualty. These propertes make these regons deal for atermarkng. The detecton of sutable regons can be realzed n the uncompressed doman. Pröfrock, Schlaueg and Müller use the Normed Centre of Gravty (NCG) algorthm to fnd the blocks that contan object borders or other sgnfcant changes n the vdeo [8]. Ellnas presents an algorthm that embeds data n mages usng four level DWT and edge detecton [7]. Ths algorthm also modfes the surroundngs of the edges; t s accomplshed usng a morphologcal dlataton th a structurng element of 9x9. The proposed algorthm s desgned for atermarkng lo resoluton (CIF and QCIF) vdeo streams. Durng the desgn of a vdeo atermarkng algorthm complexty has to be taken nto count. A trade-off has to be made beteen complexty, qualty loss, and robustness. The algorthm uses the avelet doman for data hdng. It modfes the avelet coeffcents that belong to object borders. Vsble artfacts ll not appear on the source vdeo. The sutable coeffcents are selected by an edge detecton algorthm. Watermark s nserted n an addtve ay that a spread spectrum pseudorandom nose s added to the lumnance pane of the mddle frequency components. II. THE WATERMARK EMBEDDING PROCESS Fg. shos the process of atermark embeddng. Frst the nput vdeo sgnal (X,j ) s transformed usng a fast to level DWT transform, the Haar avelet [9]. ISBN: WCE 8

2 Proceedngs of the World Congress on Engneerng 8 Vol I WCE 8, July - 4, 8, London, U.K. N*D R Te frequences. X,j DWT Xu,v Sobel edge detector X u,v Insert atermark bts Yu,v IDWT Fg.. Block dagram of the atermark embeddng process The transformaton produces a lo frequency, lo resoluton approxmaton subband, and 6 detal subbands. Watermark s nserted nto the mddle frequency components (HL and LH) of each level. Fg.. shos the vdeo frame after the avelet decomposton Fg.. Vdeo frame after avelet transformaton The transformed vdeo then gets through an edge detectng flter, here edge detecton s appled on the mddle frequency components. Edge detecton s performed usng the Sobel edge detector []. The Sobel operator performs a -D spatal gradent measurement on an mage. It s used to fnd the approxmate absolute gradent magntude at each pont n an nput grayscale mage. The Sobel edge detector uses a par of 3x3 convoluton masks, one estmatng the gradent n the x- drecton (G x ) and the other estmatng the gradent n the y- drecton (G y ). A convoluton mask s usually much smaller than the actual mage. As a result, the mask s sld over the mage, manpulatng a square of pxels at a tme. The to kernels are shon n (). G x, G y, The magntude of the gradent (G) s then calculated usng () G + G x G. () y To mprove performance an approxmate magntude s used (3) G G x + G y. (3) () Fg.3. Vdeo frame th Sobel edge detecton n the mddle frequency DWT coeffcents Durng atermark nserton the edges th value greater than a gven threshold (T e ), and the pxels around them n a gven radus (R) are selected. The value of the radus s dfferent at the dfferent levels of the transformed mage. Radus of to pxels s used at the frst level, and radus of one pxel s used at the second level consderng that the hgher level coeffcents contan loer frequency components, hch affect the qualty of the vdeo. Data s embedded nto the four selected mddle frequency areas by addng a pseudo random spread spectrum nose. Frst, data s extended th a chp rate cr to mprove robustness. d Y D, here x cr,..., (x + ) cr (4) x Inserton s done n an addtve ay usng (5). ' u,v X u,v + α N d (5) here Y u,v are the modfed avelet coeffcents, X u,v are the selected avelet coeffcents for atermarkng, α s the strength of the atermark (the value hch the lumnance values are modfed by), N s the pseudorandom Gaussan nose consstng of - and, d s the data to embed. The data also conssts of - and accordng to and. The pseudorandom nose N s calculated from a seed, and from the u and v coordnates of the selected coeffcents. Ths seed s used as the key for the atermarkng method. The same data are embedded nto consequent frames to mprove robustness. III. THE WATERMARK DETECTION PROCESS The process of the atermark detecton s also made n the avelet doman. After the to level DWT the mddle frequency blocks get through an edge detecton. Fg. 4. shos the block scheme of the atermark detectng process. Fg. 3. shos the frame after edge detecton n the mddle ISBN: WCE 8

3 Proceedngs of the World Congress on Engneerng 8 Vol I WCE 8, July - 4, 8, London, U.K. Y,j DWT Yu, v Sobel edge detector Y u,v N R Te Detect atermark Fg.4. Block dagram of the atermark detecton process dj Decson The embedded data are extracted by correlatng the pseudorandom nose N th the area contanng the detected edges and ther surroundngs. The detecton s blnd. Assumng that the value of Y u,v coeffcents s almost equal, d j Yu v N, (6), K u, v can be used, here K s the number of sutable coeffcents and d j s the extracted value. The embedded bt s calculated the follong ay: δ, δ, f d < T f d > T, (7) here T s a threshold used to decrease the number of false results. Its value s: α T. (8) Because the embedded nformaton changes only after 9 frames, smple votng s appled on the detected bts. D t s more complex than the PSNR, the results are more close to the HVS. The smlarty s calculated usng 3 factors: lumnance comparson, contrast comparson and structural comparson. Its value s beteen and, here means the perfect qualty. V. EXPERIMENTAL RESULTS The atermarkng algorthm as tested on four vdeo streams: Moble, Bus and Floer, hch are hghly detaled streams, and Foreman, hch contans more smooth regons. The sze of all vdeo streams s 35x88 pxels. A. Qualty Results Fg. 5. shos the orgnal and the atermarked verson of the same vdeo frame. For better vsblty t also contans cropped, x pxels blocks of the to frames, hch contan sgnfcant edges (stuated at the top of the bus on the rght sde of the frame). It can be seen that the nose caused by atermark sgnal s mperceptble. IV. VIDEO QUALITY MEASUREMENT A. PSNR PSNR (Peak Sgnal-To-Nose Rato) s a dely used method for measurng objectve qualty produced by mage and vdeo manpulatng algorthms []. PSNR s expressed n terms of the logarthmc decbel scale. The typcal values of the PSNR are beteen 3 and 5 db, here hgher means better qualty. It s calculated usng (9) PSNR log 55, (9) MSE here MSE s the value of the Mean Square Error beteen the orgnal and the modfed frame. It s calculated usng () MSE, () MN M N [ ] x(, j) y(, j) j here M s the heght, N s the dth of the frame, x and y are the orgnal and the atermarked mages. In some cases hoever PSNR does not correlate ell th the HVS, although t s a dely used metrc at measurng the qualty loss caused by vdeo manpulatng and compressng algorthms. B. SSIM SSIM (Structured Smlarty) s another dely used method for measurng the smlarty beteen to mages []. Whle Fg.5. Orgnal and atermarked vdeo frame and blocks of x pxels from the orgnal and atermarked vdeo frame Fg. 6. shos the PSNR and SSIM error maps beteen the orgnal and the atermarked frame. The error maps are converted to be negatve for better vsblty. It can be seen that atermark s nserted around the edges. The PSNR error map shos that numerous pxels have been altered (PSNR4,3 db), but the SSIM error map shos a better qualty (SSIM,995). ISBN: WCE 8

4 Proceedngs of the World Congress on Engneerng 8 Vol I WCE 8, July - 4, 8, London, U.K. Fg.6. PSNR and SSIM error maps beteen orgnal and atermarked frame (converted to negatve for better vsblty) Table I presents the objectve qualty values of the tested vdeos. TABLE I AVERAGE PSNR AND SSIM VALUES OF WATERMARKED TEST SEQUENCES Vdeos Moble Foreman Bus Floer PSNR[dB] 35,9 45,7 37,37 35,7 SSIM,984,9969,995,995 It can be seen that the vdeo qualty depends on the content. If the content contans more edges or textured components, then more nose s added to t. Although t mproves robustness of the atermark, the objectve qualty of the vdeo becomes orse. B. Robustness of the algorthm Robustness of the algorthm has been tested aganst lossy compresson. To codecs ere used: the H.64/AVC th Baselne profle, and the XvD codec. These codecs are dely used for lo btrate vdeo codng. Frst, test data of 8 bts ere embedded n the vdeo streams; 8 bts ere embedded at every frame usng a atermarkng strength (α) of 5. Then the atermarked vdeos have been compressed usng the to codecs th dfferent btrates. Fnally, the accordance as measured beteen the orgnal and the extracted data produced by the detector. Accordance of % means that every embedded bt can be correctly dentfed, hle 5% means that atermark sgnal can not been retreved, the embedded and the extracted data are uncorrelated. Fg.7. Blocks of 8x8 pxels of the uncompressed atermarked frame, and after H.64/AVC compresson th btrates 8 kbps, 7 kbps, 384 kbps Table II presents the objectve qualty values of the dfferent btrates used at vdeo compresson. TABLE II AVERAGE PSNR AND SSIM VALUES OF H.64 COMPRESSION OF VIDEO BUS Btrates 8 kbps 7 kbps 384 kbps PSNR[dB] 33,8 3,7 8,96 SSIM,9364,948,8573 Fg. 8. shos the test results of the H.64/AVC lossy compresson. ) Results after H.64/AVC compresson Fg. 7. presents blocks of 8x8 pxels cropped from a selected frame of the orgnal, and from the compressed streams. It can be seen that the loer the btrate of the compresson s, the less detals the vdeo frame has. Fg. 8. Robustness results aganst H.64/AVC lossy compresson Correct bts means the number of bts that can be detected correctly. Erroneous bts s the number of false bts. The embedded bt as and the detected as and vce versa. ISBN: WCE 8

5 Proceedngs of the World Congress on Engneerng 8 Vol I WCE 8, July - 4, 8, London, U.K. Unrecognzed bts s the number of bts that the detectng algorthm can not decde hether the embedded bt as or. Results sho that the atermark can be extracted even from vdeos compressed usng lo btrate H.64 compresson. ) Results after Xvd compresson Test results sho that XvD codec produce orse mage qualty at the same btrate than H.64/AVC codec. Fg. 9. presents blocks of 8x8 pxels cropped from a selected frame of the orgnal, and the compressed streams. Fg.. Robustness results aganst XvD lossy compresson VI. CONCLUSION In ths paper a novel vdeo atermarkng algorthm s presented. Data embeddng s made n the avelet doman. It uses the Sobel edge detector for fndng the approprate areas for atermarkng these areas are the sgnfcant edges of the frames and ther surroundngs. The HVS s less senstve to modfcatons on mddle and hgh frequences. Compresson algorthms make only mnor changes to these areas. The atermark tself s a pseudo random nose, hch s calculated from the nput data and a seed value. Watermark s detected by correlatng the pxel values of the selected area th a pseudo random nose. Test results sho that the embedded atermark s mperceptble to the HVS, and the method s resstant to the H.64/AVC and the XvD lossy compresson algorthms. Fg.9. Blocks of 8x8 pxels from the uncompressed atermarked frame, and from the frames compressed th XvD th btrates: 8 kbps, 7 kbps and 384 kbps TABLE III AVERAGE PSNR AND SSIM VALUES OF XVID COMPRESSION OF VIDEO BUS Btrates 8 kbps 7 kbps 384 kbps PSNR[dB] 9,55 8,77 6,86 SSIM,895,879,847 Fg.. shos the results of the tests. Correct bts, Erroneous bts and Unrecognzed bts mean the same. Test results sho that the atermark also survves XvD compresson. REFERENCES [] R. Schyndel, A. Trkel and C. Osborne, A dgtal atermark, IEEE Proc. Int. Conf. Image Processng, vol., pp. 86 9, 994. [] W. Bender, D. Gruhl, N. Mormoto, and A. Lu, Technques for data hdng, IBM Systems Journal, vol. 35, no. 3-4, pp , 996. [3] R. B. Wolfgang and E. J. Delp, A atermark for dgtal mages, n IEEE Proc. Int. Conf. Image Processng, vol. 3, pp. 9, 996. [4] I. J. Cox, J. Klan, T. Leghton, and T. Shamoon, Secure spread spectrum atermarkng for multmeda, IEEE Trans. Image Processng, vol. 6, no., pp , 997. [5] M. D. Sanson, B. Zhu, and A. H. Tefk, Transparent robust mage atermarkng, n IEEE Proc. Int. Conf. Image Processng, vol. 3, pp. 4, 996. [6] M. Barn, F. Bartoln,a and A. Pva, Improved avelet-based atermarkng through pxel-se maskng, IEEE Trans. Image Processng, vol., no. 5, pp ,. [7] J. N. Ellnas, A robust avelet-based atermarkng algorthm usng edge detecton, Internatonal Journal of Computer Scence vol., no. 3, pp. 6, 7. [8] D. Pröfrock, M. Schlaueg, E. Müller, A ne uncompressed-doman vdeo atermarkng approach robust to H.64/AVC compresson, Proc. of the IASTED Int. Conf. on Sgnal Processng, Pattern Recognton, and Applcatons, pp. 99 4, 6. [9] P. Schrder, W. Seldens, M. Cohen, T. DeRose, D. Salesn, Wavelets n Computer Graphcs, SIGGRAPH 96 Course Notes. [] M. Nxon, A. S. Aguado, Feature extracton & mage processng, Academc Press, 7. [] S. Wnkler, Dgtal Vdeo Qualty: Vson Models and Metrcs, Wley, 5. [] Z. Wang, A. C. Bovk, H. R. Shekh and E. P. Smoncell, Image qualty assessment: From error vsblty to structural smlarty, IEEE Transactons on Image Processng, vol. 3, no. 4, pp. 6-6, 4. ISBN: WCE 8