Spati-Tempral Edge-Based Weighted Fuzzy Filtering fr Prviding Interlaced Vide n a Prgressive Display Gwanggil Jen, Rafael Falcón, Jhyun Lee, and Jechang Jeng Abstract In this paper, we prpse a new weighted fuzzy filter, which selectively uses spatial and tempral infrmatin. The accuracy f the edge directin detectin, mtin detectin, and interplatin is crucial key factrs t btain excellent visual quality in deinterlaced images. The adpted fuzzy cncepts are utilized t design a weight-evaluating technique. The weights were cnsidered t be multiplied by the candidate deinterlaced pixels. Experimental results demnstrate that the prpsed deinterlacing methd perfrms better than previus techniques. I. INTRODUCTION URRENT wrldwide TV systems, such as NTSC, PAL, C SECAM, and the future HDTV, emply interlace scanning and a field frequency f either 60 Hz r 50 Hz, and thus have similar flicker disturbances [1]. Als, as high-reslutin flat panels such as plasma display, cmputers, and prjectin TV, becme mre widely used, imprving the vide quality f the displays is becming an increasingly imprtant issue. Meanwhile, mst bradcasting TV systems emply interlaced vide making a tradeff between bandwidth and vide quality []. Due t the adptin f interlaced scan, current TV systems suffer frm well-knwn visual artifacts such as interline flicker, line crawling, and field aliasing. Therefre, the interlaced-t-prgressive scan frmat cnversin becmes an imprtant issue. Previus wrks n deinterlacing techniques can be rughly categrized as mtin cmpensated [3],[4] r nn-mtin cmpensated methds (Bb [5], ELA [6], NEDI [7], Weave [8], VTMF [9], STELA [10], and EDT [11]). In general, mtin-cmpensated deinterlacing methds prvide the best perfrmance amng varius deinterlacing techniques. If mtin estimatin is accurate and reliable, the mtin-cmpensated deinterlacing methds can prvide imprved results. Hwever, mtin cmpensated methds require much higher level hardware cmplexity than ther methds [4]. Therefre, the nn-mtin cmpensated deinterlacing methds are chsen mre ften because f errr prpagatin and cmputatinal cmplexity, especially in real-time applicatins. Nn-mtin cmpensated algrithms Manuscript received December 1, 007; revised April 9, 008. This wrk was spnsred by ETRI SC Industry Prmtin Center, Human Resurce Develpment Prject fr IT SC Architect. Gwanggil Jen, Jhyun Lee, and Jechang Jeng are with the Department f Electrnics and Cmputer Engineering, Hanyang University, 17 Haengdang-dng, Sengdng-gu, Seul, Krea (e-mail: {windcap315,saint81,jjeng}@ece.hanyang.ac.kr). Rafael Falcón is with the Cmputer Science Department, Universidad Central de Las Villas, Carretera Camajuani km 5 ½ Santa Clara, Cuba (e-mail: rfalcn@uclv.edu.cu). include spatial dmain methds, tempral dmain methds, and spati-tempral schemes. The spatial dmain methds are knwn as intrafield deinterlacing which have been widely used in many applicatins since they are mre cst-efficient amng the deinterlacing techniques due t less memry requirements and hence less cmplexity. Hwever, these spatial dmain interplatin methds may intrduce flicker artifacts in mtin and edge areas. Tempral dmain methd perfrms interplatin by expliting the pixel values frm several cnsecutive field sequences and they have high implementatin cst due t multiple-field strage requirements. Hwever, they are nt expected t have gd perfrmance in cases with unreliable mtin infrmatin, and they require higher hardware cmplexity and are threatened by latent errr prpagatins. Spati-tempral dmain methds explit the crrelatin in bth spatial and tempral dmains. In this paper, we fcus n mtin adaptive deinterlacing methd which is included within spati-tempral dmain deinterlacing methds. Als, a methd fr cnstructing edge directin detectin and measuring weight techniques thrugh fuzzy inference will be presented. In literature, fuzzy lgic has successful applicatins in prcess cntrl where binary decisins are nt able t prduce gd results [1]-[19]. Until nw, sme examples f paperwrk f fuzzy cntrllers in lw-level image prcessing are a fuzzy rule-based mtin detectr [13],[14], fuzzy edge detectr [15], fuzzy rate cntrl fr MPEG vide [16], rugh sets-based deinterlacing [17], and fuzzy edge-dependent mtin adaptive deinterlacing [18]. In truth, a methd fr detecting edge directin frm an interlaced image has been cnstructed based n the gradient f the image [19]. We suppse, hwever, the perfrmance can be imprved in ther issues, such as weight measuring, by applying fuzzy inference. In this wrk, fuzzy technique is cmbined in deinterlacing methd and fund t be effective in reducing the methd s sensitivity t edge directin detectin errrs, imprving visual quality. The rest f this paper is rganized as fllws. The issues f the cnventinal deinterlacing methds are explained in Sectin II. In Sectin III, we present the prpsed algrithms. Simulatin results and analyses are prvided in Sectin IV. Finally, cncluding remarks abut this wrk and the future wrk are summarized in Sectin V. II. ISSUE OF THE CONVENTIONAL DEINTERLACING METHODS In this sectin, we briefly describe and investigate sme deep-seated prblems f intra field, inter field, and intra-inter field deinterlacing methd. Fig. 1 shws a 3D lcalized 58 978-1-444-1819-0/08/$5.00 c 008 IEEE
windw t interplate the pixel value x M (i,j,k), where M represents the utilized methd. The variable i refers t the clumn number, j t the line number, and k t the filed number. Here, u, d, l, r, p, and n represent up, dwn, left, right, previus, and next, respectively. The ntatin f the measurement LD, is the luminance difference in the edge directin represented by {=(30, 45, 90, 135, 150 )} and {=(SD, TD)} dmain, where SD and TD represent spatial dmain and tempral dmain, respectively. LD = urr dll, LD = ur dl, LD SD,30 SD,45 SD,90 = u d, LD = ul dr, LD = ull drr SD,135 SD,150 (1) (a) (b) Fig. 1. Spati-tempral windw fr the directin-based deinterlacing: (a) spatial directin filter; (b) tempral directin filter. LD = prr nll, LD = pr nl, LD = p n, TD,30 TD,45 TD,90 LD = pl nr, LD = pll nrr TD,135 TD,150 () A. Issue f the Intra Field Methd Amng spatial dmain methds, edge-based line average (ELA) algrithm is widely used since they are capable f perfrming linear interplatin alng the directin with smallest pixel difference and achieve relatively gd perfrmances with smaller cmputatinal lad [6]. A 3-by- D lcalized windw is used t cmpute directinal crrelatins and deinterlace the current pixel. Therefre, parameter SD is assigned t, the edge directin degree is calculated amng three values (45, 90, 135 ). LD, is then used t calculate the directin f the highest spatial crrelatin. The ELA perated pixel x ELA (i,j,k) is then calculated. ELA (,, ) ( ) x i j k = ur + dl, if min LD = LD ( u + d ), if min LD = LD ( ul + dr), therwise ( { 30,90,135 } ) SD ( { 30,90,135 } ) SD, SD,45, SD,90 The ELA methd gives gd perfrmance in mst cases. Hwever, they may als prduce annying artifacts when inaccurate edge infrmatin is used. Fig. shws the results perated by the ELA methd t interplate the missing pixel value. Prblems arise because the maximum crrelatin des nt always indicate the directin f an edge; althugh edge directin line average algrithm perfrms well in the area with clear edges. As can be seen in Fig., ELA des nt handle crrectly with cmplex structures and the edges are degraded severely. B. Issue f the Inter Field Methd Systems based n purely tempral filters causes a variety f artifacts when there is mving bjects in the image. The Weave (inter-field) deinterlacing methd is a simple deinterlacing methd [8], and it s utput pixel value x Weave (i,j,k) is defined as (4), Weave (3) (,, ) =, md md (4) x i j k p if j n where (i,j,k) designates the psitin, x(i,j,k) is the input field defined fr j md = n md nly. It is well-knwn that the Fig.. Examples f errrs by ELA methd: (a) riginal Krean character Chang ; (b) even field f riginal image; (c) 400% zmed riginal image; (d) 400% zmed ELA perated image; (e) determined edge directin by ELA methd. (a) (b) (c) Fig. 3. Subjective view results fr 31 st Stefan sequence [0]: (a) even filed f riginal sequence; (b) serratin artifact due t the Weave methd; (c) alias artifact due t the VTMF methd. bjective and subjective perfrmances f the Weave methd is better than that f the spatial dmain based deinterlacing methds in a static area. Figs. 3(a) and 3(b) shw the even filed f the 31 st Stefan riginal sequence and an example f the artifacts which are caused by Weave methd, respectively. The edges exhibit significant serratins and the line-crawling effect ccurs in mtin areas, which are unacceptable artifacts 008 IEEE Internatinal Cnference n Fuzzy Systems (FUZZ 008) 59
in a bradcast r prfessinal TV envirnment. C. Issue f the Intra-Inter Field Methd The vertical-tempral median filter (VTMF), which adapt t mtin r edges, is the mst ppular mtin adaptive deinterlacing methd [9]. The deinterlaced pixels are fund as the median luminance value f the upper and lwer pixels in current field and the tempral neighbr in the previus field. x ( i, j, k ) = median( u, d, p), if j md num md (5) VTMF Unfrtunately, hwever, with this filter it is als pssible t shuffle the infrmatin f different lines, causing disturbances t appear. T prve the drawbacks f VTMF, Fig. 3(c) shws a VTMF perated utput. At lw details, there is little influence n the input image. Hwever, with getting higher frequencies, the defect increases, s the fundamental details becme carse. III. PROPOSED WEIGHTED FUZZY DEINTERLACING (WFD) METHOD The values x, (i,j,k), candidate deinterlaced pixels, are the average values between tw pixels n edge directin, which are calculated as fllws. ( urr + dll ) ( ur + dl ) ( u + d ) x =, x =, x =, SD,30 SD,45 SD,90 ( ul + dr) ( ull + drr) x =, x = SD,135 SD,150 (6) ( prr + nll) ( pr + nl) ( p + n) xtd,30 =, x TD,45 =, x TD,90 =, x ( pl + nr) ( pll + nrr) =, x = TD,135 TD,150 The prpsed WFD methd is a directin-dependent interplatin technique as well, which based n a sample crrelatin and fuzzy thery fr weighting, where the detectin f edges in all pssible directins is fllwed by a weight-multiplied interplatin. In this stage, we fcus n using all infrmatin tgether nt just use nly ne edge directin. If nly the certain tempral infrmatin is cnsidered in mtin regin, sme drawbacks such as the serratin r alias artifacts are intrduced, degrade visual quality, especially n the cntur f bjects. The edge directin detectr utilized directinal crrelatins amng pixels in rder t evaluate weight in each directin t interplate a missing line. In rder t design the desired weight measuring system, we need t distinguish each edge directin whether it is wrth r nt. T this end, we assign t each f the ten directins in the 3D lcalized windw certain weights, namely with parameters w,, fr the directin (,). We cmpute LD, between tw adjacent pixels n the ten directins. Then, we figure ut the maximum bserved LD, value amng ten LD, values in the 3D lcalized windw, and we dente it as LD max. Fr each image pixel x(i,j,k) we cmpute the fuzzy index, fi,, as, (7) Fig. 4. Behavir f the S SMALL(fi,,,,). fi LD = Φ, Φ, (8) LDmax Fig. 4 shws the prpsed S-type SMALL membership functin, S SMALL (fi,,,,), where vertical axis represents a membership degree ranging frm 0 t 1, and the hrizntal axis f this functin represents all the pssible fuzzy index (fi, ) values, definitely the value range is frm 0 t 1. The S SMALL (fi,,,,) is perfrmed by (9). We assume the membership degree f S SMALL (fi,,,,) as weight w,. ( Φ ) wφ, = SSMALL fi,, ϕρτ,, = S-type SMALL membership functin 1, if fiφ, τ ( fiφ, τ ) 1, if τ < fiφ, ρ ( ϕ τ) ( fiφ, ϕ ), if ρ < fi Φ, ϕ ( ϕ τ) 0, if ϕ < fiφ, where,, and represent the upper, mid, and lwer bunds f the membership functin, S SMALL (fi,,,,), respectively. It is clearly seen that the value f the fuzzy index fi, is higher when the LD, is higher, which means the missing pixel is n the edge regin (in the case f =SD) r mtin regin (in the case f =TD). Then a membership weight w, that represent the certainty f the directin (,) t be in an edge, is calculated using the S SMALL (fi,,,,), is clse t zer. The idea behind ur study is t assign high weight t the directins that have lw luminance distance as the center. If the LD, between tw pixels n a (,) directin is lwer, then the membership degree (w, ) becmes higher. In ther wrds, we give higher weights t the directin (,) with small luminance differences. The utput f the filter can finally be illustrated, where the perated image is dented as x WFD (i,j,k), which means fuzzy weight multiplied results. xwfd ( i, j, k) = wφ, ( i, jk, ) xφ, ( i, jk, ) Φ= ( SD, TD) = ( 30,45,90,135,150 ) (10) wφ, ( i, j, k) Φ= ( SD, TD) = ( 30,45,90,135,150 ) where the result can be represented by a weighted mean with a grade and the assigned value fr membership functin, S SMALL (fi,,,,). The parameter w, (i,j,k) is determined frm the shape and width f the S SMALL (fi,,,,). (9) 60 008 IEEE Internatinal Cnference n Fuzzy Systems (FUZZ 008)
IV. SIMULATION RESULTS In rder t evaluate the perfrmance f the prpsed algrithm, we present the simulatin results in this sectin. The experiments were carried ut n the well-knwn prgressive test sequences Flwer Garden, Freman, Mbile, News, Stefan, and Table Tennis that are, respectively 35 88 pixel sized. These vide sequences are cnverted int interlaced vide by alternative sub-sampling. The dd fields in dd frames and the even fields in even frames are remved. The interlace vide is then deinterlaced by different deinterlacing algrithms and the deinterlaced vides are cmpared with the riginal vide sequences. The suitable values fr the membership functins, such as,, and values in Fig. 4, are difficult t acquire theretically and thus shuld be fund empirically. The lwer bund (=0.5), mid bund (=0.5) and upper bund (=0.75) f parameter set are selected and simulated in Sectin V. TABLE I AVERAGE PSNR AND COMPUTATIONAL CPU TIME OF EACH ALGORITHM OVER THE CORRESPONDING SIX CIF TEST SEQUENCES Sequence Flwer Freman Mbile News Stefan Table Tennis Average Spatial dmain methds Bb(dB).191 7.650 30.17 8.963 5.511 8.504 33.615 8.650 7.74 7.598 8.566 9.48 7.963 8.436 ELA(dB) 1.681 18.341 30.33 19.037 3.53 18.447 31.474 17.837 6.391 17.585 7.408.073 6.80 18.887 NEDI(dB).063 35.695 9.394 37.97 4.697 37.48 3.819 35.955 6.554 35.130 7.945 4.984 7.45 37.385 FDED(dB).068 97.47 Tempral dmain methd Weave(dB) 0.94 7.89 Spatial-Tempral dmain methds VTMF(dB).948 17.187 STELA(dB).990 7.358 EDT(dB).71 (T=0) 33.1 30.383 11.146 6.307 7.984 3.66 18.85 30.449 8.919 9.93 3.447 5.5 147.809 3.537 6.537 5.459 17.663 7.60 9.004 5.107 30.780 33.548 13.11 36.471 6.947 38.068 17.065 39.84 8.057 38.105 15.191 7.498 149.610 1.549 7.191 5.969 17.44 6.996 7.179 3.381 31.488 8.405 95.447 7.996 9.093 30.119 0.984 31.588 34.065 30.10.085 7.855 1.449 6.06 7.507 9.138 18.071 9.761 9.097 8.153 6.019 WFD(dB) 3.361 96.385 31.19 68.31 6.591 89.569 39.06 44.05 7.157 91.634 31.535 64.67 9.830 75.715 (a) (b) (c) (d) (e) (f) Fig. 5. PSNR values f the 50 frames frm (a) Flwer Garden, (b) Freman, (c) Mbile, (d) News, (e) Stefan, (f) Table Tennis vide sequences. 008 IEEE Internatinal Cnference n Fuzzy Systems (FUZZ 008) 61
(a) (b) (c) (d) (e) (f) (g) (h) (i) Fig. 6. Experimental results fr visual cmparisn: (a) frame number 178 frm Freman vide sequence; (b) detail f Freman sequence after the deinterlacing prcess f the Bb methd; (c) NEDI methd; (d) FDED methd; (e) Weave methd; (f) VTMF methd; (g) STELA methd; (h) EDT methd; (i) WFD methd. A. Objective Perfrmance Analysis The bjective quantitative measures used fr cmparisn are the peak signal-t-nise rati (PSNR) between the riginal and restred images, defined by MSE x PSNR x (, x ) ORG RES ( x ) ORG = 35 88 ( xorg ( i, j) xres ( i, j) ) (11) i= 1 j= 1, = 10lg RES 10 35 88 55 MSE x, x ( ) ORG RES (1) where x ORG is the riginal image, x RES is the restred image f size 35 88 and the 8-1 the maximum pssible intensity value with 8-bit pixel. Table I shws the PSNR and the cmputatinal CPU time level f the deinterlaced vides f different deinterlacing algrithms, including spatial dmain methds (ELA [6], NEDI [7], and FDED [15]), tempral dmain methd (Weave [8]), and spati-tempral dmain methds (VTMF [9], STELA [10], and EDT [11]). Since the rugh sets thery-based deinterlacing apprach describes the selectin rule fr the best methd amng several cnventinal methds [17], it was nt cmpared with the ther methds. In Fig. 5, PSNR result curves fr Flwer Garden, Freman, Mbile, News, and Stefan, Table Tennis vide sequences are shwn. Frm these figures, it can be seen that the perfrmance f the prpsed methds are superir with respect t the cnventinal methds. Mrever, the 6 008 IEEE Internatinal Cnference n Fuzzy Systems (FUZZ 008)
prpsed WFD methd nly require 38 % less cmputatinal CPU level as cmpared t that f FDED with a 1.975 db PSNR gain. Als, the prpsed WFD methd nly require abut.6 times cmputatinal CPU level as cmpared t that f STELA with a 0.069 db PSNR gain. In cases where the prpsed methds d nt prvide the best PSNR result, we can bserve that thse prpsed methds are always the secnd best r ne f the best methds. Hwever, we fund that smetimes bjective measures are nt suitable fr judgment f the image quality. The reasn behind this idea is that there are n direct and lgical relatinship between such bjective measures and the subjective impressin f the human bserver. Anther reasn t avid such quality measures is they generally calculate the distance between the riginal image and the enhanced image. Hwever, in practical cases, absence f nise and gd riginal images are nt prvided. Therefre, subjective tests can be used t ensure the quality f the results. The subjective cmparisn will be discussed in the fllwing sectin. B. Subjective Perfrmance Analysis Fig. 6 shws a typical deinterlaced image f the Freman sequence using varius algrithms. The spatial dmain methds d nt use tempral infrmatin and shw n mtin artifacts in the mtin regin, as shwn in Figs. 6(b-d). Hwever, they d nt wrk prperly with cmplex structures and the edges are degraded severely. Because the input vertical reslutin is halved befre the image is interplated, thus reducing the detail in the prgressive image. The results f Weave methd are prvided in Fig. 6(e). The edges exhibit significant serratins, which is an unacceptable artifact in a bradcast r prfessinal TV envirnment. Mst parts f the image are still many feathering defects n the structure f the building. In VTMF and STLEA methds, the vertical detail is gradually reduced as the tempral frequencies increase as shwn in Figs. 6(f,g). Fig. 6(h) shws the artifacts displayed due t the EDT methd. The image perated by WFD methd is presented in Fig. 6(i). The main advantage f the prpsed methd is that edge regin is nt suppressed well while fine details and edges d nt lse much sharpness. Additinally the visual results illustrate that the prpsed methd interplates riginal image much better than the ther methds. In real time, the benefit f this methd becmes nticeable since the stability f the mving hrizntal lines is assured and the traditinal flicker effect f the cnventinal methds is remved. The prpsed WFD methd prvides perceptually much better deinterlaced images with little saw tth artifacts in the hrizntal edge regin. V. CONCLUSION AND THE FUTURE WORK In this paper, a new weighted fuzzy filter which uses spatial and tempral infrmatin fr vide deinterlacing is prpsed. The prpsed deinterlacing algrithm which based n fuzzy thery is able t adapt t characteristics f sequence and give the right weights t the methds in the right amunt at the right time. Thrugh edge directin detectin, weights f each directin in each dmain are derived, s that a specific interplatin technique may be utilized t better preserve edge during deinterlacing prcess. Experimental results shw nticeable imprvement cmpared t the existing deinterlacing methds. We want t extend the fuzzy technique assisted edge preserving system. Als, we want t develp fuzzy weight evaluating system beynd ur fuzzy weight measurement in the next paper. REFERENCES [1] K. Jack, Vide Demystified - A Handbk fr the Digital Engineer, Elsevier 005. [] R. Li, B. Zeng, and L. Liu, Reliable mtin detectin/cmpensatin fr interlaced sequences and its applicatins t deinterlacing, IEEE Trans. Circuits and Systems fr Vide Technlgy, vl. 10, n. 1, pp. 3-9, Feb. 000. [3] Y. -L. Chang, S. -F. Lin, C. -Y. Chen, and L. -G. Chen, Vide de-interlacing by adaptive 4-field glbal/lcal mtin cmpensated apprach, IEEE Trans. Circuits and Systems fr Vide Technlgy, vl. 15, n. 1, pp. 1569-158, Dec. 005. [4] S. Yang, Y. -Y. Jung, Y. H. Lee, and R. -H. Park, Mtin cmpensatin assisted mtin adaptive interlaced-t-prgressive cnversin, IEEE Trans. Circuits and Systems fr Vide Technlgy, vl. 14, n. 9, pp. 1138-1148, Sep. 004. [5] E. B. Bellers and G. de Haan, Advanced de-interlacing techniques, in Prc. PrRisc/IEEE Wrkshp n Circuits, Systems and Signal Prcessing, Mierl, The Netherlands, Nv. 1996, pp. 7-17. [6] T. Dyle, Interlaced t sequential cnversin fr EDTV applicatins, in Prc. nd Int. Wrkshp Signal Prcessing f HDTV, pp. 41-430 Feb. 1990. [7] M. K. Park, M. G. Kang, K. Nam, and S. G. Oh, New edge dependent deinterlacing algrithm based n hrizntal edge pattern, IEEE Trans. Cnsumer Electrnics, vl. 49, n. 4, pp. 1508-151, Nv. 003. [8] P. L. Swan, Methd and apparatus fr prviding interlaced vide n a prgressive display, U.S. Patent 5 864 369, Jan. 6, 1999. [9] G. de Haan and E. B. Bellers, Deinterlacing An verview, Prceedings f the IEEE, vl. 86, N. 9, Sep. 1998, pp. 1839-1857. [10] H. -S. Oh, Y. Kim, Y. -Y. Jung, A. W. Mrales, and S. -J. K, Spati-tempral edge-based median filtering fr deinterlacing, IEEE Internatinal Cnference n Cnsumer Electrnics, pp. 5-53, 000. [11] M. -J. Chen, C. -H. Huang, and C. -T. Hsu, Efficient de-interlacing technique by inter-field infrmatin, IEEE Trans. Cnsumer Electrnics, vl. 50, n. 4, pp. 10-108, Nv. 004. [1] E. C. C. Tsang, D. M. Huang, D. S. Yeung, J. W. T. Lee, and X. Z. Wang, A weighted fuzzy reasning and its crrespnding neural netwrk, IEEE Internatinal Cnference n Systems, Man and Cybernetics, 00. [13] E. E. Kerre and M. Nachtegael, Fuzzy techniques in image prcessing, Springer-Verlag in the series Studies in Fuzziness and Sft Cmputing, 000. [14] S. Schulte, V. De Vitte, and E. E. Kerre, A fuzzy nise reductin methd fr clr images, IEEE Trans. Image Prcessing, vl. 16, n. 5, pp. 145-1436, May. 007. [15] F. Michaud, C. T. Le Dinh, and G. Lachiver, Fuzzy detectin f edge-directin fr vide line dubling, Trans. Circuits Systems fr Vide Technlgy, vl. 7, n. 3, pp. 539-54, June 1997. [16] D. Tsang, B. Bensau, and S. Lam, Fuzzy-Based Rate Cntrl fr Real-Time MPEG Vide, IEEE Trans. Fuzzy Systems, vl. 6, n. 4, pp. 504-516, Nv. 1998. [17] G. Jen, D. Kim, and J. Jeng, Rugh sets attributes reductin based expert system in interlaced vide sequences, IEEE Trans. Cnsumer Electrnics, vl. 5, n. 4, pp. 1348-1355, Nv. 006. [18] P. Brx, I. Baturne, S. Sánchez-Slan, J. Gutiérrez-Rís, and F. Fernández-Hernández, A fuzzy edge-dependent mtin adaptive algrithm fr de-interlacing, Fuzzy Sets and Systems, vl 158, n. 3, pp. 337-347, Feb. 007, [19] M. Nachtegael, D. Van der Weken, E. E. Kerre, and W. Philips, Sft Cmputing in Image Prcessing, Springer-Verlag in the series Studies in Fuzziness and Sft Cmputing, pp. 155-185, 007. [0] Available: http://media.xiph.rg/vide/derf/ 008 IEEE Internatinal Cnference n Fuzzy Systems (FUZZ 008) 63