Supercompression for Full-HD and 4k-3D (8k) Digital TV Systems

Transcription

1 Supercompression for Full-HD and 4k-3D (8k Digial TV Sysems Mario Masriani Absrac In his work, we developed he concep of supercompression, i.e., compression above he compression sandard used. In his conex, boh compression raes are muliplied. In fac, supercompression is based on super-resoluion. Tha is o say, supercompression is a daa compression echnique ha superpose spaial image compression on op of bi-per-pixel compression o achieve very high compression raios. If he compression raio is very high, hen we use a convoluive mask inside decoder ha resores he edges, eliminaing he blur. Finally, boh, he encoder and he complee decoder are implemened on General-Purpose compuaion on Graphics Processing Unis (GPGPU cards. Specifically, he menio-ned mask is coded inside exure memory of a GPGPU. Keywords General-Purpose compuaion on Graphics Process ing Unis, Image Compression, Inerpolaion, Super-resoluion. S I. INTRODUCTION UPERCOMPRESSION represens he mos revoluionary concep in image and video compression [1]. This concep is based on wo simple principles: a Downsampling/upsampling, i.e., spaial decimaion, and b deblurring, super-resoluion, or sharpening [2-6]. While he firs was performed using bilinear inerpolaion, he second we do hrough a horizonal rafer wih a convoluion mask, which is based on he Van Cier s ieraive algorihm [7, 8], and an improvemen (non-ieraive ha makes he menioned algorihm compuaionally more efficien, and which was developed by our eam [1]. Specifically, he super-compression is a combinaion of wo compressions, i.e., he spaial decimaion and he compression of he employed sandard. Therefore, he super-compression is a compression above he compression sandard used. In his conex, boh compression raes are muliplied. Supercompression is based on super-resoluion, because, i increases he compression on he basis of a reducion in size of he image (or frame, in he case of videos [1]. Tha is o say, super-compression is a lossy compression echnique ha superpose spaial image compression on op of bi-per-pixel compression o achieve very high compression raios. If he compression raio is very high, hen we use a convoluive mask inside decoder ha resores he edges, eliminaing he blur. Finally, boh, he encoder and he comple-e decoder are implemened on General-Purpose Mario Masriani is wih he Grupo de Invesigación sobre Procesamieno de Señales e Imágenes (GIPSI, Univ. Nac. de Tres de Febrero (UNTreF, 910 Florida S., Floor 6h, Room B, (C1005AAT, CABA, Argenina. phone: ; fax: ; mmasriani@unref.edu.ar. compuaion on Graphics Processing Unis (GPGPU cards. Specifically, he menioned mask is coded inside exure memory of a GPGPU [1, 9-12]. The Bilinear Inerpolaion is oulined in Secion II, where we discuss he problem of inerpolaing visually accepable images a a higher resoluion. We firs presen he inerpolaion problem and why linear inerpolaion filers are inadequae for image daa. To represen he major mahemaical approaches o image processing, we discuss and evaluae five differen image inerpolaion mehods. Superresoluion scheme for compression including linear inerpolaion are oulined in Secion III. Merics are oulined in Secion IV. Simulaions are ouline in Secion V. Finally, Secion VI provides a conclusion of he paper. II. BILINEAR INTERPOLATION Bilinear inerpolaion is by far he mos common inerpolaion mehod [1-6]. The idea is o inerpolae along one dimension using values ha were hemselves inerpolaed along he oher dimension, see Fig.1. Fig.1: Bilinear inerpolaion. If we have values a (x, y 0 and (x, y 1, hen we could linearly inerpolae along he verical line. This is no a problem, jus generae hem by inerpolaing along he horizonals. z x0 = (1 α z 00 + α z 10 α = (x x 0 /(x 1 x 0 z x1 = (1 α z 01 + α z 11 (1 z xy = (1 β z x0 + β z x1 β = (y y 0 /(y 1 y 0 Noe ha i does no maer wheher we inerpolae across and hen down or down and hen across (i.e. on x firs or y firs. Eiher way we end up wih z xy = (1 α (1 β z 00 + (1 α β z 01 + α (1 β z 10 + α β z 11 (2 This is bilinear inerpolaion. I resuls in a piecewise funcion ha is no piecewise linear of course i can be, because 1799

2 i maches he daa a four differen poins, and hree poins uniquely deermine he linear funcion. I has a piece for each cell in he grid of daa poins, bu he inerpolaion defined over ha recangle is no linear. Look a his mos recen equaion, remembering ha α is a linear funcion of x and β is a linear funcion of y. The full expression for z xy is going o conain a consan erm, an x erm, a y erm, and a xy erm. Because of he presence of his las erm is no linear. This kind of funcion is called bilinear because i is linear as a funcion of x when y is held fixed and also linear as a funcion of y when x is held fixed. The qualiy is obvious, see Fig.2. Fig.2: Image inerpolaion using bilinear mehod of inerp2 buil-in MATLAB funcion. Top: original image. Medium: close-up of eye in image. Down: inerpolaed image. III. SUPER-RESOLUTION SCHEME FOR COMPRESSION This secion is organized ino four pars, for a beer undersanding of he conceps: A. Super-resoluion vs Deblurring, B. Compression vs Super-compression, C. Deducion of he mask D. Applicaions A. Super-resoluion vs Deblurring: As we saw in Secion I, here is much confusion beween he conceps of super-resoluion and deblurring in Digial Image Processing [13, 14]. We are going o esablish here wo rigorous definiions for he purpose of eliminaing his confusion. We say ha a process is super-resoluion if i resores he sharpness of an image involving an increase in he resoluion of he same [1-6, 13, 14]. We say ha a process is deblurring if i resores he sharpness of an image no involving an increase in he resoluion of he same. This process is applied when he image sharpness suffers an aberraion called blur [13, 14], which comes from a high relaive speed of he objec in focus in relaion o he camera, fas opening and closing he shuer, ec. We consider imporan o menion ha boh processes can involve each oher as par of he process of improving he sharpness of he image. In fac, we can undersand he superresoluion as a process of increasing he resoluion followed by a resoraion of he edges by a deblurring process. On he oher hand, previously esablished definiions are fundamenal o undersanding wha follows. B. Compression vs Super-compression: We define compression as he process reduces he average number of bi-per-pixel (bpp of an image. In Fig. 3, we represen he se of bi-planes in which decomposes a gray or color image. As seen in Fig. 3, he compression process does no aler he image size [13, 14]. Insead, we define supercompression as he process reduces he average number of bi-per-pixel (bpp of an image afer downsizing. The size reducion process is performed by down-sampling, which akes shrinkage in rows and columns, wihou obligaion o respec he aspec raio (16:9. In fac, for ISDB-Tb (Inegraed Services Digial Broadcasing Brazilian Digial TV Sysem we use 5:1 as compression rae over he original com-pression of he sysem, which uses H.264 as video compress-ion sandard [15]. When we say, we increase he sandard compression 5 imes, his means ha we move from a resolu-ion of 1920x1080 (Full-High Definiion: Full- HD o anoher 5 imes lower of 720x576 (Sandard Definiion: SD. The sandard video compression H.264 is no affeced by he supercompression. As discussed in Sub- Secion D, supercom-pression requires minimal equipmen a he ransmier and he reverse procedure o supercompression in he receiver (se-op-box [16]. 1800

3 wih bilinear inerpolaion, while he deblurring is done by a bidimensional convoluive mask of NxN pixels, which makes a rafer over he upsampled (blurred image. Fig.3: Compression. However, he unavailabiliy of he laer, he sysem is compaible, since he receiver will send he SD signal o he Liquid Display Crysal (LCD TV, which naurally made upsampling obviously changing he aspec raio, as when a Full-HD LCD TV receive a SD signal. In Fig. 4, we represen he se of bi-planes in which decomposes a gray or color image. As discussed in Sub-Secion D, our supercompression procedure consiss in wo pars spread in ransmier and receiver. In ransmier we have hree seps: 1. Video slicing: frame-by-frame 2. Downsampling 3. Video reassembling and in receiver inside se-op-box we have four seps: 1. Receiver of sreaming/h H Upsampling 4. Deblurring In our case, he downsampling and upsampling is done Fig.4: Supercompression. The parameers of his squared mask (where N is odd are criicals, herefore, he such parameers mus be calculaed and adjused wih oal accu-racy. In he nex secion, we will proceed o deduc he mask and se he opimal relaionship beween is parameers. Laer we will proceed o adjus hem via rial and error. C. Deducion of he mask: Based on he las secion, he single frame is recovered afer suffering a pair of processes: downsampling and upsampling, see lef side of Fig.5. In his figure: X means original single frame. Y means recovered (blurred single frame. M b means square mask of NxN pixels (where N is odd. This mask is known as a blurred mask, smoohing ope raor or Poin Spread Funcion (PSF [2]. 1801

4 Sub-index means -ieraion. means downsampling. means upsampling. Fig.5: Downsampling/upsampling as a blurred mask. In hese processes ( and, he single frame is affeced by a space/ime invarian blur. On he basis of his, we need an esimaor o recover he single frame of he processes affecing i. Then, for an image affeced by a downsampling/upsampling as Fig.5, we deduce ha he bes esimaor is he Van Cier s recursive algorihm [7, 8]. The se of equaions reflecing he above model can be divided ino wo sages: he model and he esimaor [1]. Based on Fig.5, we have: Model: X = Y M +1 X (3 = / X (4 d u Where means bidimensional convoluion, and M d/u represens a convoluive and unknown mask which summarizes he combined acion of downsampling and upsampling ogeher. Esimaor: X = + λ ε = Y Y Y M X + 1 X (5 ε (6 = / (7 d u Where 0<λ<2 is a consan parameer o adjus. Therefore, X 0 = Y On he oher hand, he compuaional implemenaion of he above se of equaions involves he use of four nesed for s plus a sric conrol of he sabiliy of he Eq.5 (wih a predicor form from resricing he possible values of λ, i.e., only i is possible o use 0<λ <2. Therefore, i is much more efficien o implemen such filering hrough a simple bidimensional mask convoluion, eliminaing he predicor form of Eq.5, which allows much more efficien implemenaions using - for example - a convoluion hrough he Fas Fourier Transform (FFT [13, 14]. In consequence, we need deduce such mask. If we replace (8 Eq.7 in Eq.6, we have, ε Y M X (9 = d / u Now, we replace Eq. 9 inside Eq.5, obaining, X = X + ( Y M X + 1 λ d / u (10 Reagrouping erms of Eq.10, and remembering a model of low noise and linear space and ime invarian blur, we have, X +1 = M s Y (11 Where M s is a mask as shown in Fig.6, and he following relaionships o consider are very imporan, 2 ( N 1 α + β = 1, (for deblurring (12 2 ( N 1 α + β = 0, (for edge deecion (13 Thus, a new and simplified model of deblurring appears on he scene, see Fig.7, where α < 0 and β > 1. We need o esablish precisely boh parameers, hen, here are wo possible ways forward: 1. Choose N (ineger, posiive, odd and small, and β > 1 (and arbirarily less han 2, hen α is derived from Eq Sar wih arbirary values of α and β (abou cerain recommendaions, e.g., -0.1 < α < 0 and 1 < β 2 and generaing a random populaion of he pair [α, β], and deducing N from Eq.12. Fig.6: Deblurring mask M s The acion of his mask can be seen in Fig.8. Firs, we performed an up-scaling of he image, and second, we apply he mask of he Fig.6 on he middle image, hus, obaining a much higher qualiy final image [1]. 1802

5 he upscaling, wih an obvious change in he original aspec raio, i.e., Fig.7: New and simplified model of deblurring. D. Applicaions: Fig.8: Our echnology wih super-resoluion We presen hree main applicaions of video compression in real ime for Digial TV, according o sandard ISDB-Tb [17]. In he firs, we move from a resoluion of 1920x1080 Full- HD o anoher 5 imes lower of 720x576 SD. As we have said before, he sandard video compression H.264 is no affeced by he supercompression. The Fig.9 shows a diagram of he encoder wih hree modules embedded ino GPGPU cards [1, 9-12]. In fac, we work wih hree GPUs in he encoder. For sarers, he camera delivers picures wih a resoluion of 1920x1080 pixels HD-SDI, SDI which means Serial Digial Inerface [1]. The firs GPU performs a separaion of he frames of he video, frame by frame. This procedure is called video slicing [1]. This allows us o individually access each frame o apply downsampling. The above sampling is conduced in he second GPU, for which, we have seleced NVIDIA Tesla 2050 [18] for sric cusomer requiremens (see Fig.10, however, can be carried ou he same downsampling wih a much less powerful, and herefore much less expensive NVIDIA GPU, wihou any problem. By downsampling, we pass from a resoluion of 1920x1080 pixels o anoher 5 imes less, i.e. 720x576. Tha is o say, from Full-HD o SD. Therefore, we achieve lower bi rae (and herefore he bandwidh used 5 imes. This seemingly arbirary compression raio and reducion of Full-HD forma o SD is required by he Argenine governmen, so ha if a user does no have our decoder, o enjoy he SD broadcas. In his case is he same TV who makes Fig.9: Encoder. from 16:9 o 4:3. A very imporan aspec o consider is ha his procedure does no require any kind of color ransform, i.e., i works direcly on he RGB (reed, green and blue componens of each frame. This eliminaes he wo possible conversions and hus he compuaional cos hey enail. We used Texure Memory of GPGPU o a compuaional efficien implemenaion of he differen modules of encoder and decoder, allowing us o reach TV imes. Fig.10 shows in deail he employed echnology for he real implemenaion of Fig.9, which consiss in wo Quadro GPUs [18] he firs for video slicing. frame-by-frame, and he second 1803

6 Fig.10: Encoder implemenaion wih GPGPUs. for video reassembling, respecively. The downsampling is im-plemened on a Tesla 250 [18]. However, currenly, we have found a way o perform his experimen using only one Quadro GPU. Moreover, in Fig.10: TX means ransmier On he oher hand, Fig.11 shows a diagram of he decoder implemened inside a se-op-box (STB. So ha, if he STB has he superdecompression and depending on he resoluion of he LCD TV, we obain resoluions of High Definiion (HD 720x1280 or Full-HD 1080x1920. However, if he STB hasn he superdecompression, he sysem mus be compaible, here-fore we obain only SD 576x720. The Fig.12 shows he Super-Resoluion Module (SRM used inside STB of Fig.11, which includes upsampling and deblurring, hus resoring he original resoluion. Fig.11: Decoder. Fig.13 represens he real implemenaion of Fig.11, in which, we can see, he se-op-box used in his work, developed by Dixar Inc. [16]. This STB works equally wih Terres-rial Digial TV, IPTV, WebTV, 3DTV and Digial Cinema. Besides, his STB has camera and moion sensors, which can be used as ineracive gaming plaform. Acually, we are working on an inegraed circui (chip [19] o replace he curren GPGPU inside he STB, minimizing he power consumpion and he size of his [16]. Finally, he second applicaion of his echnology presened here is shows in Fig.14, where we use a mobile phone wih High-Definiion Mulimedia Inerface (HDMI video ou as a recepor. 1804

7 Fig.12: Super-resoluion Module (SRM. Fig.14: Mobile phone as HD or Full-HD recepor. As shows in Fig.14, we ake he HDMI video ou, and we inroduce i in he STB. Depending on he resoluion of he LCD TV we obain HD o Full-HD resoluions. The original resoluion of he mobile phone employed is Low Definiion (LD 320x240 One-Seg (one of 13 segmens ha form he ISDB-T norm, see Fig.15. In his case, he addiional compression rae of STB on H.264 is 27:1 [16]. Fig.13: Se-op-box of Dixar Inc. Fig.15: Deail of 13 segmens inside ISDB-T channel. 1805

8 Finally, he hird and las applicaion of his echnology presened here is shows in Fig.16 (a he end of his paper, which consiues he modern Digial TV Sysem 4k-3D designed for Argenine governmen, which has he same informaion rae 8k monochannel. All encoders currenly in use (wihou excepion use inra and inerframe compression simulaneously, even modern European coder known as HEVC/H.265 (High Efficiency Video Coding, and which will begin esing in The inerframe compression is composed of hree pars: 1. Scene deecion 2. Moion deecion 3. Region of Ineres (ROI deecion These hree modules are responsible for he delay known as laency, which for European sysem of digial TV knowed as DVB (Digial Video Broadcasing is 5.5 seconds, while for he Brazilian sysem of digial TV knowed as ISDB-Tb sysem is 4.5 seconds. If an encoder such as H.264 is used for 3D-4k, hus, he laency would be beween 25 and 35 seconds. This is unaccepable. A his poin, we define laency as he delay beween he digial and analog ransmission. On he oher hand, H.264 was originally designed for video ransmission of low and medium resoluion. In fac, for rans-mission for 2k resoluion and up, i has shorcomings, morpho-logical defecs, and chromaic aberraions. Tha is o say, i does no fi he 4K-3D, as well as oher codecs. This is he reason ha since 2k resoluions can only use inraframe compression, especially he JPEG2000 codec [21, 22]. Since i can no serve inerframe compression, hen, he compression raes obained are very low, wih he spending disproporionae bandwidh of he channel for ransmissions of his ype. Moreover, given ha he Argenine governmen wishes o reuse he digial TV plaform insalled of Full-HD, hen he only viable soluion ha mees all boundary condiions is one based on supercompression [1]. As shown in Fig.16, we have wo images of 3840x2160 pixeles, one for righ eye and one for lef eye, i.e. a oal resoluion of 3840 x 2160 x 2 (i.e., sereo. We firs performed he downsampling, obaining wo images of 1920x1080 each, which are encoded in H.264 and sen o he ransmier. Once he receiver, upsampling and deblurring is applied o boh images, hereby resoring he original resoluion. Tha is o say, we lower he bi rae of he wo images combined wih a quarer of is value, bu, however, boh images of 1920x1080 combined occupy abou 60% of he bi rae ha would occupy he original image ransmied by a sysem of single-channel Full-HD. This urns ou o be a produc arising from he characerisics of he mehod iself [1], which creaes a seamless smoohing. IV. METRICS A. Daa Compression Raio (CR Daa compression raio, also known as compression power, is a compuer-science erm used o quanify he reducion in daa-represenaion size produced by a daa compression algorihm. The daa compression raio is analogous o he physical compression raio used o measure physical compression of subsances, and is defined in he same way, as he raio beween he uncompressed size and he compressed size [13, 14]: Uncompressed Size CR = (14 Compressed Size Thus a represenaion ha compresses a 10MB file o 2MB has a compression raio of 10/2 = 5, ofen noaed as an explici raio, 5:1 (read "five o one", or as an implici raio, 5X. Noe ha his formulaion applies equally for compression, where he uncompressed size is ha of he original; and for decompression, where he uncompressed size is ha of he reproducion. B. Bi-per-pixel (bpp The "bis per pixel" refers o he sum of he bis in all hree color channels and represens he sum colors available a each pixel before compression ( bpp. However, as a compression bc meric, he bis-per-pixel refers o he average of he bis in all hree color channels, afer of compression process ( bpp. ac Compressed Size bpp bpp = bpp = bc (15 ac Uncompressed Size bc CR Besides, bpp is also defined as Number of coded bis bpp = (16 ac Number of pixels C. Mean Absolue Error (MAE The mean absolue error is a quaniy used o measure how close forecass or predicions are o he evenual oucomes. The mean absolue error (MAE is given by 1 NR 1NC 1 MAE = NRxNC nr = 0 nc = 0 X ( nr, nc X ( nr, nc (17 which for wo NR NC (rows-by-columns monochrome images X and X, where he second one of he images is considered a decompressed approximaion of he oher of he firs one. D. Mean Squared Error (MSE The mean square error or MSE in Image Compression is one of many ways o quanify he difference beween an original 1806

9 Fig.16: Deail of he resoluions involved in he proposed 4k-3D TV Sysem. 1807

10 image and he rue value of he quaniy being decompressed image, which for wo NR NC (rows-by-columns monochrome images X and X, where he second one of he images is considered a decompressed approximaion of he oher is defined as: 1 NR 1NC 1 MSE = NRxNC nr = 0 nc = 0 2 X ( nr, nc X ( nr, nc (18 E. Peak Signal-To-Noise Raio (PSNR The phrase peak signal-o-noise raio, ofen abbreviaed PSNR, is an engineering erm for he raio beween he maximum possible power of a signal and he power of corruping noise ha affecs he fideliy of is represenaion. Because many signals have a very wide dynamic range, PSNR is usually expressed in erms of he logarihmic decibel scale. The PSNR is mos commonly used as a measure of qualiy of reconsrucion in image compression, ec [13]. I is mos easily defined via he mean squared error (MSE, so, he PSNR is defined as [14]: MAX 2 MAX PSNR = 10 log 10 ( X = 20 log 10 ( X (19 MSE MSE Here, MAX X is he maximum pixel value of he image. When he pixels are represened using 8 bis per sample, his is 256. More generally, when samples are represened using linear pulse code modulaion (PCM wih B bis per sample (bps, maximum possible value of MAX X is 2 B -1. For color images wih hree red-green-blue (RGB values per pixel, he definiion of PSNR is he same excep he MSE is he sum over all squared value differences divided by image size and by hree [13, 14]. Typical values for he PSNR in lossy image and video compression are beween 30 and 50 db, where higher is beer. V. SIMULATIONS The simulaions are organized in four experimens, separaed in wo groups: sill images (for obvious reasons, however, idenical resuls were achieved in video, HDTV and Digial Cinema by color and gray. All experimens include calculaions of MAE, MSE, PSNR, bpp and CR. All hese experimens involve he comparison beween he use of JPEG vs SC (JPEG+SR, and JPEG2000 vs SC (JPEG2000+SR for sill color and gray images, in boh cases over a BMP file (which doesn have compression, o raw daa mode, where he used acronym means: BMP: BiMap file forma [20] JPEG: Join Picure Group [20] JPEG2000: JPEG wih waveles [21, 22] SC: Super-compression SR: Super-resoluion A. Group 1: Main characerisics of employed image: File = angelina.bmp Color = yes Size = 1920-by-1080 pixels Original bpp = 24 Experimen 1: JPEG vs SC (JPEG+SR JPEG: See Table I, column JPEG, and Fig.17 (2 nd from op. 1. From BMP (24 bpp, 1920x To JPEG ( bpp, 1920x From JPEG ( bpp, 1920x To BMP(24 bpp, 1920x1080 SC (JPEG+SR: See Table I, column SC (JPEG+SR, and Fig.17 (3 rd from op. 1. BMP (24 bpp, 1920x Downsampling (24 bpp, 720x JPEG ( bpp, 720x JPEG ( bpp, 720x Upsampling ( bpp, 1920x Deblurring ( bpp, 1920x BMP (24 bpp, 1920x1080 Experimen 2: JPEG2000 vs SC (JPEG2000+SR JPEG2000: See Table II, column JPEG2000, and Fig.17 (4 h from op. 1. From BMP (24 bpp, 1920x To JPEG2000 ( bpp, 1920x From JPEG2000 ( bpp, 1920x To BMP (24 bpp, 1920x1080 SC (JPEG2000+SR: See Table II, column SC (JPEG2000+ SR, and Fig.17 (down. 1. BMP (24 bpp, 1920x Downsampling (24 bpp, 720x JPEG2000 ( bpp, 720x JPEG2000 ( bpp, 720x Upsampling ( bpp, 1920x Deblurring ( bpp, 1920x BMP (24 bpp, 1920x1080 The following ables show he merics vs he Algorihms for boh cases, i.e., JPEG and JPEG2000 vs Supercompression. 1808

11 TABLE I ANGELINA (COLOR, 24 BPP, 1920X1080: JPEG VS SC (JPEG+SR Merics JPEG SC (JPEG+SR MAE MSE PSNR bpp CR TABLE II ANGELINA (COLOR, 24 BPP, 1920X1080: JPEG2000 VS SC (JPEG2000+SR Merics JPEG2000 SC (JPEG2000+SR MAE MSE PSNR bpp CR B. Group 2: Main characerisics of employed image: File = lena.bmp Color = gray Size = 512-by-512 pixels Original bpp = 8 Experimen 3: JPEG vs SC (JPEG+SR JPEG: See Table III, column JPEG, and Fig.18 (2 nd from op. 1. From BMP (8 bpp, 512x To JPEG ( bpp, 512x From JPEG ( bpp, 512x To BMP(24 bpp, 512x512 SC (JPEG+SR: See Table III, column SC (JPEG+SR, and Fig.18 (3 rd from op. 1. BMP (8 bpp, 512x Downsampling (8 bpp, 256x JPEG ( bpp, 256x JPEG ( bpp, 256x Upsampling ( bpp, 512x Deblurring ( bpp, 512x BMP (8 bpp, 512x512 Experimen 4: JPEG2000 vs SC (JPEG2000+SR JPEG2000: See Table IV, column JPEG2000, and Fig.18 (4 h from op. 1. From BMP (8 bpp, 512x To JPEG2000 ( bpp, 512x From JPEG2000 ( bpp, 512x To BMP (8 bpp, 512x512 Fig.17: Firs (op original image, second (coded and decoded wih JPEG, hird (coded and decoded wih JPEG+Supercompression, fourh (coded and decoded wih JPEG2000, fifh (down, coded and decoded wih JPEG2000+Supercompression. 1809

12 SC (JPEG2000+SR: See Table IV, column SC (JPEG2000+ SR, and Fig.18 (down. 1. BMP (8 bpp, 512x Downsampling (8 bpp, 256x JPEG2000 ( bpp, 256x JPEG2000 ( bpp, 256x Upsampling ( bpp, 512x Deblurring ( bpp, 512x BMP (8 bpp, 512x512 The following ables show he merics vs he Algorihms for boh cases, i.e., JPEG and JPEG2000 vs Supercompression. TABLE III LENA (GRAY, 8 BPP, 512X512: JPEG VS SC (JPEG+SR Merics JPEG SC (JPEG+SR MAE MSE PSNR bpp CR TABLE IV LENA (GRAY, 8 BPP, 512X512: JPEG2000 VS SC (JPEG2000+SR Merics JPEG2000 SC (JPEG2000+SR MAE MSE PSNR bpp CR Finally, all echniques were previously implemened in MATLAB R2010b (Mahworks, Naick, MA [23] on a Noebook wih Inel Core(TM i5 CPU M 2.27 GHz and 6 GB RAM on Microsof Windows 7 Home Premium 64 bis, and hen in NeSream of Dixar Inc. [18] on NVIDIA [18] wo Quadro Tesla 2050 GPUs for encoder, and NVIDIA GTX285 GPU inside STB developed by Dixar Inc. [16] for decoder, as shown in Fig.16. VI. CONCLUSION A. Group 1: Experimen 1: JPEG vs SC (JPEG+SR In his experimen SC (JPEG+SR has MAE, MSE and PSNR wih pracically he same order of magniude han JPEG alone, however, bpp is five imes lower, a he same ime, CR is five imes higher, see Table I. As shown in Fig.17, he second (coded and decoded wih JPEG and he hird (coded and decoded wih JPEG+Supercompression from he op, have he same lookand-feel and image qualiy han he op, i.e., original image of Angelina. Experimen 2: JPEG2000 vs SC (JPEG2000+SR We make similar consideraions for his experimen, regar Fig.18: Firs (op original image, second (coded and decoded wih JPEG, hird (coded and decoded wih JPEG+Supercompression, fourh (coded and decoded wih JPEG2000, fifh (down, coded and decoded wih JPEG2000+Supercompression. 1810

13 ding o he las experimen, see Table II and Fig.17 (fourh coded and decoded wih JPEG2000 alone, and fifh coded and decoded wih JPEG2000+Supercompression, however, here is a big difference beween JPEG and JPEG-2000 o compress his ype of image (compare bpp and CR of Table I and II. B. Group 2: Experimen 3: JPEG vs SC (JPEG+SR In his experimen SC (JPEG+SR has MAE, MSE and PSNR wih pracically he same order of magniude han JPEG alone, however, bpp is five imes lower, a he same ime, CR is five imes higher, see Table III, idem Experimen 1. As shown in Fig.18, he second (coded and decoded wih JPEG and he hird (coded and decoded wih JPEG+Supercompression from he op, have he same lookand-feel and image qualiy han he op, i.e., original image of Lena. Experimen 4: JPEG2000 vs SC (JPEG2000+SR Idenical consideraions han Experimen 2 are necessary, see Table IV and Fig.18, wih he same conclusions abou he difference beween JPEG and JPEG-2000 o compress his ype of image (compare bpp and CR of Table III and IV. C. For boh groups: We used Texure Memory inside STB [16] GPGPU o a compuaional efficien implemenaion of he bidimensional convoluive mask of deblurring module, allowing us o reach TV imes, i.e., a frame every 40 milliseconds. ACKNOWLEDGMENT M. Masriani hanks Prof. Marin Kaufmann, vice chancellor of Universidad Nacional de Tres de Febrero, for his remendous help and suppor. REFERENCES [1] M. Masriani, Single Frame Supercompression of Sill Images, Video, High Definiion TV and Digial Cinema, Inernaional Journal of Informaion and Mahemaical Sciences, vol. 6:3, pp , [2] A. Gilman, D. G. Bailey, S. R. Marsland, Inerpolaion Models for Image Super-resoluion, in Proc. 4h IEEE Inernaional Symposium on Elecronic Design, Tes & Applicaions, DELTA 2008, Hong Kong, 2008, pp [3] D. Glassner, S. Bagon, M. Irani. Super-Resoluion from a Single Image. Available: hp:// e_image_sr.pdf [4] A. Lukin, A. S. Krylov, A. Nasonov. Image Inerpolaion by Super- Resoluion. Available: hp://graphicon.ru/oldgr/en/publicaions/ex/lukinkrylovnasonov.pdf [5] Y. Huang, Wavele-based image inerpolaion using mulilayer perceprons, Neural Compu. & Applic., vol.14, pp.1-10, [6] N. Mueller, Y. Lu, and M. N. Do. Image inerpolaion using muliscale geomeric represenaions. Avalilable: hp://lcav.epfl.ch/~lu/papers/inerp_conourle.pdf [7] S.H.M. Allon, M.G. Deberrand, and B.T.H.M. Sleujes, "Fas Deblurring Algorihms", Available: hp:// [8] A. Bennia and S.M. Riad, Filering Capabiliies and Convergence of he Van-Cier Deconvoluion Technique, IEEE, Trans. Insrum. Meas., Vol. 41, no. 2, pp , Apr [9] M. Kraus, M. Eissele, and M. Srenger. GPU-Based Edge-Direced Image Inerpolaion. Available: hp://cieseerx.is.psu.edu/viewdoc/summary?doi= [10] -. NVIDIA CUDA: Bes Pracices Guide, version 3.0, 2/4/2010. Available: hp://developer.download.nvidia.com/compue/cuda/3_0/oolki/docs/n VIDIA_CUDA_BesPracicesGuide.pdf [11] V. Podlozhnyuk. Image Convoluion wih CUDA, June Available: hp://developer.download.nvidia.com/compue/cuda/1_1/websie/projec s/convoluionseparable/doc/convoluionseparable.pdf [12] V. Simek, and R. Rakesh, GPU Acceleraion of 2D-DWT Image Compression in MATLAB wih CUDA, in Proc. Second UKSIM European Symposium on Compuer Modeling and Simulaions, Liverpool, UK, 2008, pp [13] R.C. Gonzalez, R.E. Woods, Digial Image Processing, 2nd Ediion, Prenice- Hall, Jan. 2002, pp [14] A.K. Jain, Fundamenals of Digial Image Processing, Englewood Cliffs, New Jersey, [15] I. E. Richardson, H.264 and MPEG-4 Video Compression: Video Coding for Nex Generaion Mulimedia, Ed. Wiley, N.Y., [16] hp:// [17] hp:// [18] NVIDIA (NVIDIA Corporaion, Sana Clara, CA. [19] hp:// [20] J. Miano, Compressed Image File Formas: JPEG, PNG, GIF, XBM, BMP; Ed. Addison-Wesley, N.Y., [21] T. Acharya, and P-S Tsai, JPEG2000 Sandard for Image Compression: Conceps, Algorihms and VLSI Archiecures, Ed. Wiley, N.Y., [22] A. Bilgin, and M. W. Marcellin, JPEG2000 for Digial Cinema in Proceedings of 2006 Inernaional Symposium on Circuis and Sysems (ISCAS, (invied paper, May [23] MATLAB R2010b (Mahworks, Naick, MA. Mario Masriani was born in Buenos Aires, Argenina on February 1, He received he B.Eng. degree in 1989 and he Ph.D. degree in 2006, boh in elecrical engineering. Besides, he received he second Ph.D. degree in Compuer Science in All degrees from Buenos Aires Universiy. He is Professor of Digial Signal and Image Processing of he Engineering College, a Buenos Aires Universiy (UBA. Professor Masriani is he Coordinaor of Technological Innovaion (CIT of ANSES, and he Compuer Engineering Deparmen of he Naional Universiy of Tres de Febrero, a Buenos Aires, Argenina. He published 51 papers. He is a currenly reviewer of IEEE Transacions on Neural Neworks, Signal Processing Leers, Transacions on Image Processing, Transacions on Signal Processing, Communicaions Leers, Transacions on Geoscience and Remoe Sensing, Transacions on Medical Imaging, Transacions on Biomedical Engineering, Transacions on Fuzzy Sysems, Transacions on Mulimedia; Springer-Verlag Journal of Digial Imaging, SPIE Opical Engineering Journal; and Taylor & Francis Inernaional Journal of Remoe Sensing. He (M 05 became a member (M of WASET in His areas of ineres include Digial Signal Processing, Digial Image Processing, Compression and Super-resoluion. 1811