IMAGE QUALITY ASSESSMENT USING THE MOTIF SCAN

Transcription

1 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 IMAGE QUALITY ASSESSMENT USING THE MOTIF SCAN Z. Cui, D.-O Kim, and R.-H. Pak Depatment of Electonic Engineeing, Sogang Univesity, Seoul, Koea {Cuizt69, hyssop, hpak}@sogang.ac.k ABSTRACT Image quality assessment (IQA) evaluates the quality of an image by computing the diffeence between the efeence and distoted images. This pape poposes an image quality metic (IQM) that uses the motif scan. Since edges often contain much infomation on an image and the human visual system is highly adapted fo extacting stuctual infomation fom a scene, we popose a new IQM compaing the similaity of block motif scans between the efeence and distoted images, whee the cente of the block is located at edge pixel. A block is divided into fou ovelapping gids with the cente pixel located at fou diffeent cones. The modified vesions of the poposed IQM using the unifom and Gaussian weights ae also poposed. Expeiments with LIVE database fo five diffeent distotion types of test images show that the poposed metic gives bette pefomance than the conventional IQMs. KEYWORDS DMOS, Image Quality Assessment, Logistic Regession, Motif Scan. INTRODUCTION Image quality assessment (IQA) evaluates how good an image is. It can be classified into two types: subjective and objective. The fome is bette than the latte because the quality of an image o video is eventually assessed accoding to the human visual peception. Howeve, such subjective quality assessments ae toublesome and expensive, thus not suitable fo pactical use in eal applications. Theefoe, the latte is desiable as the pactical IQA and thus many objective IQA methods have been developed [ 7]. The objective IQA methods ae consideed bette if they ae as close as to the subjective IQA. The goal of the objective IQA is to evaluate the image quality that is simila to the quality as people peceive. Thus, the mean squae eo (MSE) and peak signal to noise atio (PSNR) that use only the intensity diffeence fo assessing the quality cannot effectively eflect the human peception popeties and thus cannot assess the image quality coincidentally with the subjective quality like the mean opinion scoe (MOS). This eason leads to the development of the objective IQA algoithms. Conventional IQA algoithms can be classified into seveal appoaches depending on the usage of the efeence image and the kind of the infomation used fo IQA. In this pape, we classify the convention IQA algoithms into thee categoies as in []. Thee categoies ae stuctual infomation based [ ], human peception/visual attention based [5 5], and infomation theoetical appoaches [6, 7]. Motivation of stuctual infomation based IQA is that the stuctual infomation of an image changes if an image is distoted. The univesal quality index (UQI) [] was pesented as a fullefeence (FR) IQM using the stuctual infomation of an image. The stuctual similaity (SSIM) [], a modified vesion of the UQI, was also developed. Howeve, the SSIM gives poo DOI : 0.5/sipij.0.50

2 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 pefomance fo badly blued images. To educe pefomance degadation, new methods have been developed [, 5]. Edge-based SSIM [] is based on the edge infomation as the most impotant image stuctue infomation. The gadient-based SSIM has been poposed by consideing the fact that the human visual system (HVS) is sensitive to changes of edges [5]. A multi-scale SSIM (MSSSIM) was poposed [6]. An IQA method based on the edge and contast similaity between the efeence and distoted images [7] was poposed by combining the edge similaity with the contast similaity used in the SSIM. Similaly to the MSSSIM, an IQM based on multi-scale edge epesentation was poposed [8]. Also, a discete wavelet tansfom-based stuctual similaity fo IQA was poposed [9]. Shnaydeman et al. poposed an image quality metic (IQM) based on singula value decomposition (SVD) [0], which is called the MSVD. The MSVD method used the mean of the diffeences between SVD values fo assessing the image quality. An IQM using LU factoization (MLU) was poposed, whee LU factoization was used fo epesentation of the stuctual infomation of an image []. An IQM based on Hais esponse (HRQM) was poposed, in which Hais esponse was computed fom the gadient infomation matix []. In [], the joint featue-based quality metic was pesented, in which image pixels wee classified into thee stuctue types to effectively assess the visual quality. Also, featue map based IQM [] was poposed based on the SSIM. Instead of compaing images diectly, the method uses the SSIM in the featue maps (cone, edge, and symmety maps) between the efeence and distoted images. A human peception based IQM is motivated by the fact that only the image distotions that can be peceived by most people affects the subjective image quality. This means that if the distotion that cannot be peceived by the HVS occus with egad to the efeence image, people may conside that the distoted image is the same as o simila to the efeence image. Thus, some IQMs use the just-noticeable-distotion to detect the distotion that human eye can peceive. A shapness metic based on just-noticeable blus was poposed to measue blu atifacts [5] and a peceptual IQM fo blocking atifacts of joint photogaphic expets goup (JPEG) compessed images was poposed [6]. A distotion measue of the effect of fequency distotion, and a noise quality measue of the effect of additive noise wee developed [7]. DCTune is designed in the context of a discete cosine tansfom (DCT)-based model of visual quality [8]. Also, ITU ecommended fou annexes fo objective peceptual IQA [9]. A wavelet based visual signal-tonoise atio (VSNR) fo natual images was also developed based on nea-theshold and supatheshold popeties of the human vision [0]. Anothe appoach elated to the human peception is based on the phase [ ]. Phase-based IQMs wee motivated fom the fact that if an image has some stuctual distotions, stuctual distotions lead to the consistent phase change. Fo IQA, the phase conguency was used [], which is a measue of featue significance in images, a method of edge detection that is paticulaly obust against changes in illumination and contast. Zhai et al. poposed the log Gabo phase similaity [], which is a FR IQM based on measuing of similaities between phases in log Gabo tansfom domain. Also, similaity of phase spectum was used fo evaluating the image quality []. Also, visual attention based IQMs wee investigated [, 5]. Most existing IQMs do not take the human attention analysis into account. Attention to paticula objects o egions is an impotant fact of human vision and peception system in measuing peceived image and video qualities. Feng et al. pesented a saliency-based objective quality assessment metic [], fo assessing the peceptual quality of decoded video sequences affected by packet loss. Also, an appoach fo extacting visual attention egions based on a combination of a bottom-up saliency model and semantic image analysis was poposed [5].

3 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 An infomation theoetic appoach was poposed, which quantifies visual fidelity by means of an infomation fidelity citeion (IFC) deived based on natual scene statistics [6]. The visual infomation fidelity (VIF) was also pesented [7]. In this pape, we popose a new stuctual infomation based IQM, which uses the block motif scan of edges in an image. In the poposed metic, edges of the efeence image is extacted and then, pocessed by the block motif scanning. The block can be divided into fou ovelapping gids whee the edge pixel at the cente of the block is located at the diffeent cones of fou gids. The motifs can eflect the stuctual infomation of the scanned images. When an image is degaded by noise, compession, and so on, motifs of the image ae changed. Theefoe, the diffeence between fou motifs of each edge pixel between the efeence and distoted images can be used to evaluate the amount of distotion. The diffeence between the efeence and distoted images is computed by compaing whethe the motifs of fou gids in each block ae the same o not. The est of the pape is stuctued as follows. In Section, we descibe the poposed IQM. Fist, we biefly eview on the motifs which ae the base of the poposed IQM. Then, the block motif scan is descibed. Next, the poposed IQM and its modified vesions ae pesented. In Section, expeimental esults ae shown with discussions. Finally, in Section conclusions and futue woks ae given.. PROPOSED IQM USING THE MOTIF SCAN In this section, we intoduce the motif scan which is the main idea of the poposed IQM. Then, the motif scan based quality metic (MSQM) is poposed and its modified vesions ae also pesented... Review on Motif Scan Jhanwa et al. poposed an image etieval method using a local featue, which is a motif scan [8]. A motif is a kind of patten obtained fom tacing intensities of pixels in a block and is used as a descipto of that block. Motifs wee used in image etieval [9 ], since if the images ae diffeent, the motifs ae also diffeent. Also, if an image is degaded, motifs of that image also change with espect to those of the efeence image. Thus, this is a motivation of the poposed IQM, to which a modified motif scan is applied. In geneal, diffeent motifs could tavese a gid, howeve, among which in this pape we conside only six motifs that stat fom the top left cone of the gid. Also an additional motif is defined when the fou intensity values of a gid ae the same. An image is divided using gid units and each gid is eplaced by a motif. Fig. illustates seven motifs defined in a gid. The elevant motifs minimize the local intensity vaiation along the scan line. (0) () () () () (5) (6) Figue. Seven motifs used to tavese a gid p p Given a gid, then δ q ( q 6) ae defined as the sums of absolute diffeences of p p intensity values along the qth scan path:

4 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 δ = p δ = p δ = p δ = p δ = p 5 δ = p 6,,,,,. () Then, the motif of a gid is defined as 0, if min( δ = = q ) 0 m ag min( δq ), othewise () q whee 0 m 6 epesents the index of the seven motifs. In fact, motifs ae six types as defined in [9 ]. Howeve, in flat egion, a motif cannot be defined. Theefoe, we add a motif fo epesenting flat egion, esulting in seven motifs in total. In eq. (), if min( δ ) = 0, ag min( δ ) is not equal to 0 and m is equal to q that gives the minimum of δ q, q 6. In othe wods, if thee is at least one δ q that is equal to zeo, it means that the egion is flat and m becomes zeo. Othewise, m can have a value (fom to 6) among six values, which gives the minimum of δ q. Fig. shows an example of the motifs. With an 8 8 image shown in Fig. (a), it can be tansfomed into a coesponding image in Fig. (b) and also can be epesented as Fig. (c) using the index defined in Fig.. When scanned along the line as in Fig. (b), the value of the intensity vaiation is the minimum among the six motifs. q q q (a) (b) (c) Figue. An example of the motifs: (a) An 8 8 image; (b) Motif tansfomed image of (a); (c) Motifs epesented by the index.. Block Motif Scan In this pape, we use a block motif scan which was used in image etieval [, ]. In the block motif scan, each edge pixel located at the cente of the block epesents the stuctual infomation of an image. Fig. shows an example of the block motif scan. Fig. (a) epesents a block that can be divided into fou ovelapping blocks, as shown in Fig. (b). The cente pixel (i, is located at diffeent cones of the fou gids as shown in Figs. (c) (f). In a gid, the motif consides only the elationship among the fou adjacent pixels, howeve, in a block, it consides the elationship among eight neighboing pixels aound the cente pixel, whee the cente pixel is used fo constucting fou motifs with ovelapping blocks. Thus, the pixel located at the cente poduces a set of fou adjacent motifs that can

5 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 epesent the elationship among the neighboing pixels in the block and these fou motifs in a block ae used to constuct a featue vecto. (a) (b) (c) (d) (e) (f) Figue. block divided into fou ovelapping gids: (a) block with cente pixel at (i,; (b) Fou ovelapping blocks with (i, centeed in the block; (c) gid of block ; (d) gid of block ; (e) gid of block ; (f) gid of block.. Poposed MSQM Using the Block Motif Scan In geneal, if an image is distoted, the image stuctue changes. Fo example, both magnitude and diections of gadients change. Then, scan diection, which defines the motif, will also change. So, motifs can be used fo detecting the diffeence between the efeence and distoted images. It is the motivation and the fundamental concept of the poposed MSQM fo evaluating the image quality. Also, motif scan was successfully used to etieve peceptually simila images [, ], which means that motifs may be able to detect the peceived diffeence between two images. In geneal, the HVS is sensitive to the changes aound edges. Thus, it is possible to assess the image quality although only edge pixels ae seached. In pactice, many distotions might not be elated to edges. In the case of false edges, it is obvious that the image is seveely degaded, in which it is not necessay to seach all the pixels of the image fo assessing the quality. Now, conside some atifacts such as inging and noise atifacts. Ringing atifacts occu aound edges. In IQA, we use dissimilaity values at edge pixels only, howeve, we also use a block fo computing the motifs with Gaussian weights. By this pocess, inging atifacts can be measued. In the case of noise, noise is widely distibuted in the image but not in a paticula egion. Thus, as in the case of inging atifacts, noise can also be measued by using only edge pixels. Pactically, edge PSNR (EPSNR) [9] used only edge pixels fo the video quality assessment. Thus, in this pape, instead of using the whole pixels of an image, only edge pixels of an image ae used. To conside the neighboing pixels aound edges, a block containing fou ovelapping gids is used. Fig. shows the block diagam of the poposed IQM. Fist, edges of the efeence image ae extacted. Then, the block motif scan is constucted at each edge pixel in the efeence image. In the distoted image, the block motif scan is constucted at the same position as the edge position detected in the efeence image. The diffeence between the efeence and distoted images is computed by compaing whethe the motifs of fou gids in each block, defined at edge pixel, ae the same o not. 5

6 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 Figue. Block diagam of the poposed IQM Fo detecting edges, vaious edge detectos can be used and accoding to the edge detecto used, the pefomance of the poposed IQM might not be guaanteed. In this pape, the hoizontal and vetical Sobel masks ae used to obtain the edge infomation because of thei efficiency and simplicity. The edge can be easily detected by, Gx ( i, + Gy ( i, > T E( i, = 0, othewise () whee hoizontal and vetical gadients G x (i, and G y (i, ae computed by, espectively G ( i, = Y ( i +, j ) + Y ( i +, + Y ( i +, j + ) x ( Y ( i, j ) + Y ( i, + Y ( i, j + )) G ( i, = Y ( i, j + ) + Y ( i, j + ) + Y ( i +, j + ) y ( Y ( i, j ) + Y ( i, j ) + Y ( i +, j )) () (5) with T signifying the theshold and Y denoting the intensity of the efeence image. Then, the block motif scan is obtained at edge pixel of an image. In using the gid fo scanning the motifs, stuctual vaiations of the neighboing pixels at the cente pixel ae consideed by compaing the fou motifs. Of couse, symmetical 5 5 and 7 7 gids can be also consideed. Howeve, as the gid size inceases, the computational complexity inceases exponentially. Fo example, in the case that a 5 5 o 7 7 gids is used fo motif scanning, the numbe of scanning is equal to 6 o 6, espectively. In geneal, the numbe of scanning is (N ) fo an N N gid. Thus, in this pape, a gid fo scanning the motifs is selected. The featue vectos M and M d at edge pixel (i, can be defined as t t M ( i, = [ m, m, m, m ] and M d ( i, = [ md, md, md, md ], espectively, whee subscipts and d denote the efeence and distoted images, espectively, t epesents the tanspose, and components m ( u ) of the column vecto M signify the fou motifs in a block wheeas u u the components m d of the column vecto M d denote those of the distoted image. Each featue vecto includes fou motifs defined in a block centeed at edge pixel (i,. Then, the dissimilaity between motifs of the efeence and distoted images at edge pixel is defined as u = s, D( i, u= 0, if E( i, = othewise (6) 6

7 with Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 u u u 0, if m = md s = (7), othewise denoting the dissimilaity between the motif vectos M and M d. eq. (7) is defined in [], in which motifs ae used as weights epesenting whethe o not two blocks ae the same. The quality measue between the efeence image Y and distoted image Y d is defined as W H MSQM ( Y, Yd ) = D( i, (8) N i= j= whee N denotes the total numbe of edge pixels of the efeence image. W and H signify the width and height of the image, espectively, with both quantities assumed to be even numbes. If majo contents eflecting the image quality ae found below the theshold in some images, that is, those pixels ae not non-edge pixels, ou poposed MSQM might not be able to wok well. In the case whee a white image coupted with some noise, the oiginal image has no edges. Theefoe, ou poposed MSQM value is equal to zeo. It means that the distoted image is equal to the efeence one. Thus, in this pape, we assume that the degadations ae widely spead in an image. In this pape, a logistic egession [] is pefomed to descibe the elationship between the diffeential MOS (DMOS) and an IQM. The logistic egession can be witten as DMOS β = + β (, ) β5 exp( β( (, ) β)) IQM Y + Yd + IQM Y Yd (9) whee the paametes β, β, β, β, and β 5 ae obtained by logistic egession between the quality measue IQM(Y, Y d ) and the DMOS []. Fo testing the pefomance of an IQM obtained afte fitting by the logistic egession method, the fitted vesions of the IQM ae compaed with the DMOS, the subjective IQA method, in tems of the pefomance measue such as the Peason coelation coefficient (CC), oot mean squae eo (RMSE), and Speaman ank ode CC (SROCC), which wee ecommended by video quality expet goup (VQEG) [] and used in []... Modified MSQM The poposed IQM descibed in Sec.. uses the intensity values at edge pixels of the efeence and distoted images, whee a block motif scan is pocessed pixel-by-pixel ove the edge pixels. One poblem with this method is that the block motif can only eflect the local featue in a block wheeas it cannot conside moe pixels outside the block. In ode to solve this poblem, we use moe infomation of the neighboing pixels. Thus, instead of the intensity value at pixel, we use the weighted sum of pixel values computed in a pespecified window, fo example, in a 5 5 window, which can be witten as ' Y ( i, = Y ( i + k, j + l) w( k, l) (0) k= l= with k= l= w ( k, l) = () 7

8 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 whee w(k, l) denotes the nomalized weighting function. In the UQI [], the local statistics and the UQI wee computed in an 8 8 window using the unifom weight. This modified MSQM using the weighted intensity values with a unifom weighting function gives bette pefomance than the MSQM using the oiginal pixel values (with no weight) of an image. The UQI map often exhibits undesiable blocking atifacts, thus to solve the poblem, the modified vesion of this method, the SSIM, was poposed [] although thee ae some diffeences between the SSIM and UQI except fo the Gaussian weights. The SSIM uses the Gaussian weight fo computing the local statistics and the SSIM index map, which makes the pefomance of the SSIM bette than that of the UQI. Similaly, we use the modified MSQM with Gaussian weight. We get bette pefomance with the Gaussian weight than with the unifom weight, as expected. The MSQM using the unifom weighing function gives the same impotance to all the neighboing pixels aound the edge pixel in a window. Howeve, the pixels that ae fa away fom the cente pixel of the window give little effect on the cente pixel. The lage window gives bluing and thus degades the pefomance of the poposed IQM. Thus, we use a 5 5 ciculasymmetic Gaussian weighting function in eq. (0), whee w(k, l) denotes the Gaussian weighting function with standad deviation of 0.8 samples. In this pape, we have expeimented with thee diffeent weighting schemes: MSQM with no weight (MSQM N ), MSQM with unifom weight (MSQM U ), and MSQM with Gaussian weight (MSQM G ). Fig. 5 shows the efeence and thee distoted images with thei distotion maps. Fig. 5(a) is a efeence image and Figs. 5(b), 5(c), and 5(d) ae the JPEG 000 compession images of Fig. 5(a), with the bitates at.80, 0.767, and bits/pixel, espectively, and the degee of degadation inceases in ode of Figs. 5(b), 5(c), and 5(d). The DMOS of Figs. 5(b), 5(c), and 5(d) ae 6.969,.97, and , espectively, with the decease of the bitates. Figs. 5(e)- (f) epesent the distotion map between Figs. 5(a) and 5(b)-(d), espectively, with the ange between 0 and, at step of 0.5 (see eq. (6)). The ed colo denotes the value of. The moe ed colo pixels the distotion map has, the moe the image is degaded. Fig. 5(e) shows the distotion map between Figs. 5(a) and 5(b), with MSQM N =.5, MSQM U =0.9, and MSQM G =9.99. Fig. 5(f) shows the distotion map between Figs. 5(a) and 5(c), with MSQM N =6.56, MSQM U =7.056, and MSQM G =9.5. Fig. 5(g) shows the distotion map between Figs. 5(a) and 5(d) with MSQM N =9.989, MSQM U =5.978, and MSQM G =5.65. It can be obseved that the lage the degee of the degadation becomes, the moe ed pixels the distotion map has, which epesents that degadation of the distoted images becomes lage. Fo example, Fig. 5(g) has moe ed pixels than Figs. 5(e) and 5(f) in edge egions of the sail, thus we can say that Fig. 5(d) is moe degaded than Figs. 5(b) and 5(c). It can also be obseved that the poposed MSQMs can epesent the degee of degadation well when compaed with the DMOS value. 8

9 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 (a) (b) (c) (d) (e) (f) (g) Figue 5. Refeence and distoted images with thei distotion maps: (a) Refeence image; (b) JPEG 000 image (compessed at.80 bits/pixel, DMOS=6.969); (c) JPEG 000 image (compessed at bits/pixel, DMOS=.97); (d) JPEG 000 image (compessed at bits/pixel, DMOS= ); (e) Distotion map between (a) and (b) (MSQM N =.5, MSQM U =0.9, MSQM G =9.99); (f) Distotion map between (a) and (c) (MSQM N =6.56, MSQM U =7.056, MSQM G =9.5); (g) Distotion map between (a) and (d) (MSQM N =9.989, MSQM U =5.978, MSQM G =5.65). EXPERIMENTAL RESULTS AND DISCUSSIONS In this section, we evaluate the pefomance of the poposed MSQM G though pefomance compaison with conventional IQMs fo the LIVE images []. Fo all the images, the edge is detected using the theshold T=69 in eq. () and the motifs ae obtained by using the Gaussian weight of intensity of an image. We expeimentally select the theshold value 69 when the Peason CC between the MSQM and MOS values on the LIVE dataset is highest. Fig. 6 shows the gaph of the pefomance of the poposed MSQM as a function of the theshold T. The nomalized weight with a 5 5 cicula-symmetic Gaussian weighting function is used, with the standad deviation of 0.8. The conventional IQMs such as the PNSR, MSVD [0], MSSIM [], IFC [6], VIF [7], MLU [], and HRQM [] ae used fo pefomance compaison. 9

10 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 Figue 6. Gaph of the pefomance of the poposed MSQM as a function of the theshold T... Images Used in Expeiments In this pape, the LIVE data images [] ae used fo pefomance compaison of the poposed MSQM G and conventional IQMs. Thee ae total of 98 images consisting of 779 distoted and 0 efeence images. The entie images ae deived fom 9 high-esolution and high-quality colo images. The images ae vaious, such as faces, people, animals, natual scenes, and so on. Thee ae five distotion types: JPEG 000 compession (JPEG 000), JPEG compession (JPEG), white noise (WN), Gaussian blu (GBlu), and fast fading (FF). These distotions eflect a wide ange of image degadations, fom smoothing to stuctued distotion, image dependent distotion, and andom noise. The level of distotion was vaied to geneate images with a wide ange of quality, fom impeceptible levels to high levels [5]... Pefomance Compaison In ode to evaluate the pefomance of the poposed MSQMs, we compae them with the seven conventional IQMs such as the PSNR, MSVD, MSSIM, IFC, VIF, MLU, and HRQM in tems of the Peason CC, RMSE, and the Speaman ank ode CC (SROCC) which ae used as the citeia fo the pefomance evaluation []. The pefomance compaisons of eight IQMs fo each of five diffeent distotion types and fo the entie images ae given. Table shows the pefomance of the poposed MSQMs (MSQM N, MSQM U, and MSQM G ) in tems of the Peason CC, RMSE, and SROCC. It can be obseved that the poposed MSQM G gives bette pefomance than the othe two MSQMs fo entie images (All data) and five distotion types. The MSQM U gives bette pefomance than MSQM N wheeas the MSQM G bette than MSQM U. Expeimental esults in Table suppot the desciption in Sec... Thus, aftewad, we egad the poposed MSQM G as ou final poposed IQM and compae the MSQM G with conventional IQMs. In Tables, numbes undelined epesent the best pefomance measues consideed among IQM methods compaed. Table. Pefomance compaison of the poposed MSQMs (Peason CC, RMSE, and SROCC). Peason CC RMSE SROCC JPEG 000 JPEG WN GBlu FF All data MSQM N MSQM U MSQM G MSQM N MSQM U MSQM G MSQM N MSQM U MSQM G

11 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 Table shows the pefomance compaison of diffeent IQMs in tems of the Peason CC afte logistic egession. The Peason CC between the IQMs and the DMOS is calculated to evaluate the pediction accuacy. A high value of the Peason CC signifies that the pefomance of the IQM is good. As shown in Table, fo the entie images (All data) the poposed MSQM G gives bette esults than the othe conventional IQMs, which means that the poposed MSQM G is most simila to the DMOS (high Peason CC) among eight IQMs. Fo the fast fading (FF) distoted images, the MSQM G gives bette pefomance than the othe IQMs. Fo the othe fou distotion types except fo FF, the VIF gives bette pefomance than the othe IQMs. Fo the JPEG 000 and JPEG distotion types, the MSSIM gives the pefomance compaable to the poposed MSQM G. The poposed MSQM G gives the pefomance simila to the VIF in white noise (WN). It gives bette pefomance than the othe conventional IQMs except fo the VIF. Table. Pefomance compaison of diffeent IQMs (Peason CC). JPEG 000 JPEG WN GBlu FF All data PSNR MSVD MSSIM IFC VIF MLU HRQM MSQM G Table lists the pefomance compaison of diffeent IQMs in tems of the RMSE. The RMSE quantifies the amount of the distance between the tue values and the estimates. A small value of the RMSE denotes that the pefomance of the IQM is good. The poposed MSQM G gives bette pefomance than the conventional IQMs fo the entie images (All data), which signifies that the distance between the MSQM G and DMOS is small. Fo the thee distotion types except fo JPEG000 and WN, the poposed MSQM G gives the best esults, and fo JPEG000 and WN, it gives pefomance simila to the MSSSIM and VIF, espectively. Anyhow, the poposed MSQM G gives bette pefomance than othe conventional IQMs fo the entie images and fo the diffeent distotion types. Table. Pefomance compaison of diffeent IQMs (RMSE). JPEG 000 JPEG WN GBlu FF All data PSNR MSVD MSSIM IFC VIF MLU HRQM MSQM G Fig. 7 shows the scatte plots of the poposed MSQM G and the VIF vesus the DMOS fo JPEG 000 and Gaussian blu distotions. In case of the Gaussian blu, the Peason CC of the VIF is lage than that of the poposed MSQM G, as shown in Table, wheeas, the RMSE of the poposed MSQM G is smalle than that of the VIF, as shown in Table. The eason is that the 5

12 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 distibution of the points in the poposed MSQM G is close to the ed line than that of the VIF fo the JPEG 000 compession distotion as shown in Fig. 7(a). The ed line signifies the Peason CC of. Fo the Gaussian blu distotion, the eason is the same as the JPEG 000 compession distotion case, as shown in Fig. 7(b). It is also tue fo the JPEG compession and FF cases. (a) (b) Figue 7. Scatte plots of the poposed MSQM G and VIF vesus DMOS fo JPEG 000 and Gaussian blu distotion types: (a) JPEG 000; (b) Gaussian blu. Table lists the pefomance compaison of diffeent IQMs in tems of the SROCC. The SROCC can eflect the monotonicity of the IQM. A high value of the SROCC signifies that the pefomance of the IQM is good. It can be obseved that the poposed MSQM G gives the bette pefomance than othe IQMs fo entie images (All data) and the simila esults to the VIF fo WN. In WN, the PSNR gives the esults compaable to the VIF. Table. Pefomance compaison of diffeent IQMs (SROCC). JPEG 000 JPEG WN GBlu FF All data PSNR MSVD MSSIM IFC VIF MLU HRQM MSQM G As shown in Tables, the poposed MSQM G gives good pefomance fo the entie images (All data), which signifies that the poposed MSQM G gives the tend simila to the DMOS (high Peason CC), the distance between the MSQM G and DMOS is small, and the ank accoding to the MSQM G is simila to that accoding to the DMOS. Fig. 8 shows scatte plots of eight IQMs (seven conventional IQMs and the poposed MSQM G ) vesus the DMOS. As shown in Fig. 7, the poposed MSQM G shows simila pefomance to the VIF, whee the VIF gives bette pefomance than the othe conventional IQMs. It can be obseved that the scatte plot of the poposed MSQM G is clusteed moe closely to the ed line than othe IQMs, and that the distance between the line and the points is smalle than those of othe IQMs. Thus, the poposed MSQM G gives bette pefomance fo entie images (All data) than othe IQMs in tems of the Peason CC and RMSE as shown in Tables and. 5

13 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 Figue 8. Scatte plots of diffeent IQMs vesus DMOS Fig. 9 shows the scatte plots of five distotion types of the poposed MSQM G vesus the DMOS. It can be obseved that the distibution of the points is close to the ed line fo the white noise and Gaussian blu distotions than fo the othe thee distotion types. Thus, the RMSE values of the two distotion types ae smalle than those of the othe thee distotion types, especially, fo the Gaussian blu distotion. Figue 9. Scatte plots of five distotion types of the poposed MSQM G vesus DMOS In summay, it can be obseved that the pefomance of the poposed MSQM G is bette than those of conventional IQMs. Figs. 7 9 suppot the expeimental esults shown in Tables.. CONCLUSIONS In this pape, we popose a new stuctual infomation based IQM, which uses the block motif scan of edge pixels of an image. Instead of using the whole pixels of an image, only edge pixels of an image extacted by Sobel masks ae used. Thee diffeent types of MSQMs ae descibed and thei pefomance is compaed: MSQM N, MSQM U, and MSQM G. Though expeiments with the LIVE images, it can be obseved that the poposed MSQM G gives the bette pefomance than othe conventional IQMs in tems of the Peason CC, RMSE, and SROCC. Futue wok will focus on the extension of the poposed MSQM G that uses colo infomation. ACKNOWLEDGEMENTS This wok was suppoted by the Second Bain Koea Poject. REFERENCES [] Z. Wang, A. C. Bovik, Mean squaed eo: love it o leave it? A new look at signal fidelity measues, IEEE Signal Pocessing Magazine 6 (009) [] Z. Wang and A. C. Bovik, A univesal image quality index, IEEE Signal Pocessing Lettes 9 (00) 8 8. [] Z. Wang and A. C. Bovik, Image quality assessment: Fom eo visibility to stuctual similaity, IEEE Tansaction on Image Pocessing (00)

14 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 [] G.-H. Chen, C.-L. Yang, and S.-L. Xie, Edge-based stuctual similaity fo image quality assessment, in: Poceedings of Intenational Confeence on Acoustics, Speech, and Signal Pocessing, Toulouse, Fance, 006, pp. 9. [5] G.-H. Chen, C.-L. Yang, and S.-L. Xie, Gadient-based stuctual similaity fo image quality assessment, in: Poceedings of Intenational Confeence on Image Pocessing, Atlanta, GA, 006, pp [6] Z. Wang, E. P. Simoncelli, and A. C. Bovik, Multiscale stuctual similaity fo image quality assessment, in: Poceedings of IEEE Asiloma Confeence on Signals, Systems, and Computes, Pacific Gove, CA, 00, pp [7] F. Wei, X. Gu, and Y. Wang, Image quality assessment using edge and contast similaity, in: Poceedings of IEEE Intenational Joint Confeence on Neual Netwoks, Hong Kong, China, 008, pp [8] G. Zhai, W. Zhang, X. Yang, and Y. Xu, Image quality assessment metics based on multi-scale edge pesentation, in: Poceedings of IEEE Wokshop Signal Pocessing System Design and Implementation, Athens, Geece, 005, pp. 6. [9] C.-L. Yang, W.-R. Gao, and L.-M. Po, Discete wavelet tansfom-based stuctual similaity fo image quality assessment, in: Poceedings of IEEE Intenational Confeence on Image Pocessing, San Diego, CA, 008, pp [0] A. Shnaydeman, A. Gusev, and A. M. Eskicioglu, An SVD-based gayscale image quality measue fo local and global assessment, IEEE Tansaction on Image Pocessing 5 (006) 9. [] H.-S. Han, D.-O Kim, and R.-H. Pak, Stuctual infomation-based image quality assessment using LU factoization, IEEE Tansaction on Consume Electonics 55 (009) [] D.-O Kim and R.-H. Pak, New image quality metic using the Hais esponse, IEEE Signal Pocessing Lettes 6 (009) [] D.-O Kim and R.-H. Pak, Joint featue-based visual quality assessment, Electonics Lettes (007) 5. [] L. Cui and A. R. Allen, An image quality metic based on cone, edge and symmety maps, in: Poceedings of Bitish Machine Vision Confeence, Leeds, UK, 008. [5] R. Fezli and L. J. Kaam, A no-efeence objective image shapness metic based on just-noticeable blu and pobability summation, in: Poceedings of Intenational Confeence on Image Pocessing, San Antonio, TX, 007, pp [6] Z. Wang, H. R. Sheikh, and A. C. Bovik, No-efeence peceptual quality assessment of JPEG compessed images, in: Poceedings of Intenational Confeence on Image Pocessing, Rocheste, NY, 00, pp [7] N. Damea-Venkata, T. Kite, W. Geisle, B. Evans, and A. C. Bovik, Image quality assessment based on a degadation model, IEEE Tansaction on Image Pocessing 9 (000) [8] A. B. Watson, DCTune: A technique fo visual optimization of DCT quantization matices fo individual images, in: Digest of Technical Papes of Society fo Infomation Display, Seattle, WA, 99, pp [9] ITU-R Recommendation J., Objective peceptual video quality measuement techniques fo digital cable television in the pesence of a full efeence, Intenational Telecommunication Union, 00. [0] D. M. Chandle and S. S. Hemami, VSNR: A wavelet-based visual signal-to-noise atio fo natual images, IEEE Tans. Image Pocessing 6 (007) [] Z. Liu and R. Laganiee, On the use of phase conguency to evaluate image similaity, in: Poceedings of Intenational Confeence on Acoustics, Speech, Signal Pocessing, Toulouse, Fance, 006, pp [] G. Zhai, W. Zhang, Y. Xu, and W. Lin, LGPS: Phase based image quality assessment metic, in: Poceedings of IEEE Wokshop Signal Pocessing Systems, Shanghai, China, 007, pp [] P. Skuowski and A. Guca, Image quality assessment using phase spectum coelation, Lectue Notes in Compute Science, Compute Vision and Gaphics, Eds. G. Goos et al., Spinge-Velag Belin Heidelbeg, (008) [] X. Feng, T. Liu, D. Yang, and Y. Wang, Saliency based objective quality assessment of decoded video affected by packet losses, in: Poceedings of Intenational Confeence on Image Pocessing, San Diego, CA, 008, pp [5] Z. You, A. Pekis, M. M. Hannuksela, and M. Gabbouj, Peceptual quality assessment based on visual attention analysis, in: Poceedings of ACM Intenational Confeence on Multimedia, Beijing, China, 009, pp

15 Signal & Image Pocessing : An Intenational Jounal (SIPIJ) Vol., No.5, Octobe 0 [6] H. R. Sheikh, A. C. Bovik, and G. de Veciana, An infomation fidelity citeion fo image quality assessment using natual scene statistics, IEEE Tansaction on Image Pocessing (005) 7 8. [7] H. R. Sheikh and A. C. Bovik, Image infomation and visual quality, IEEE Tansaction on Image Pocessing 5 (006) 0. [8] N. Jhanwa, S. Chaudhui, G. Seethaaman, and B. Zavidovique, Content based image etieval using motif cooccuence matix, Image and Vision Computing (00) 0. [9] T.-X. Lin and C.-S. Hung, Quadant motif appoach fo image etieval, in: Poceedings of the th Intenational Confeence on Cental Euope on Compute Gaphics, Visualization and Compute Vision, Plzen, Czech Republic, 006, pp [0] A. Hafiane, S. Chaudhui, G. Seethaaman, and B. Zavidovique, Region-based CBIR in GIS with local space filling cuves to spatial epesentation, Patten Recognition Lettes 7 (006) [] C.-H. Lin, R.-T. Chen, and Y.-K. Chan, A smat content-based image etieval system based on colo and textue featue, Image and Vision Computing (008) 0. [] Y.-F. Chen, Y.-K. Chan, G.-U. Chang, M.-C. Tsao, Y.-J. Syu, and C.-H. Lin, Image etieval using modified colo vaiation co-occuence matix, in: Poceedings of th Intenational Confeence on Industial, Engineeing and Othe Application of Applied Intelligent System, Woclaw, Poland, 008, pp. 5. [] VQEG, Daft final epot fom the video quality expets goup on the validation of objective models of multimedia quality assessment, phase I, [] H.R. Sheikh, Z. Wang, L. Comack, and A.C. Bovik, LIVE image quality assessment database elease, [5] H. R. Sheikh, M. F. Sabi, and A. C. Bovik, A statistical evaluation of ecent full efeence quality assessment algoithms, IEEE Tansaction on Image Pocessing 5 (006) 0 5. AUTHORS Z. Cui eceived the M.S. degee in electonic engineeing fom Sogang Univesity in 009. He cuent eseach inteests ae image quality assessment and image pocessing. Dong-O Kim eceived the B.S. and M.S. degees in electonic engineeing fom Sogang Univesity, Seoul, Koea, in 999 and 00, espectively. Cuently, he is woking towad the Ph.D. degee in electonic engineeing at Sogang Univesity. His cuent eseach inteests ae image quality assessment and physics-based compute vision fo compute gaphics. Rae-Hong Pak was bon in Seoul, Koea, in 95. He eceived the B.S. and M.S. degees in electonics engineeing fom Seoul National Univesity, Seoul, Koea, in 976 and 979, espectively, and the M.S. and Ph.D. degees in electical engineeing fom Stanfod Univesity, Stanfod, CA, in 98 and 98, espectively. In 98, he joined the faculty of the Depatment of Electonic Engineeing, School of Engineeing, Sogang Univesity, Seoul, Koea, whee he is cuently a Pofesso. In 990, he spent his sabbatical yea as a Visiting Associate Pofesso with the Compute Vision Laboatoy, Cente fo Automation Reseach, Univesity of Mayland at College Pak. In 00 and 00, he spent sabbatical semestes at Digital Media Reseach and Development Cente, Samsung Electonics Co., Ltd. (DTV image/video enhancement). His cuent eseach inteests ae compute vision, patten ecognition, and video communication. He seved as Edito fo the Koea Institute of Telematics and Electonics (KITE) Jounal of Electonics Engineeing fom 995 to 996. D. Pak was the ecipient of a 990 Post-Doctoal Fellowship pesented by the Koea Science and Engineeing Foundation (KOSEF), the 987 Academic Awad pesented by the KITE, and the 000 Haedong Pape Awad pesented by the Institute of Electonics Enginees of Koea (IEEK), the 997 Fist Sogang Academic Awad, and the 999 Pofesso Achievement Excellence Awad pesented by Sogang Univesity. 55