No-Reference Image Quality Assessment Based on Localized Gradient Statistics: Application to JPEG and JPEG2000

Transcription

1 No-Reference Image Quality Assessment Based on Localized Gradient Statistics: Application to JPEG and JPEG2000 Hantao Liu* a, Judit Redi b, Hani Alers a, Rodolfo Zunino b, Ingrid Heynderickx a,c a Department of Mediamatics, Delft University of Tecnology, Delft, Te Neterlands b Department of Biopysical and Electronic Engineering, University of Genoa, Genoa, Italy c Group Visual Experiences, Pilips Researc Laboratories, Eindoven, Te Neterlands ABSTRACT Tis paper presents a novel system tat employs an adaptive neural network for te no-reference assessment of perceived quality of JPEG/JPEG2000 coded images. Te adaptive neural network simulates te uman visual system as a black box, avoiding its explicit modeling. It uses image features and te corresponding subjective quality score to learn te unknown relationsip between an image and its perceived quality. Related approaces in literature extract a considerable number of features to form te input to te neural network. Tis potentially increases te system s complexity, and consequently, may affect its prediction accuracy. Our proposed metod optimizes te feature-extraction stage by selecting te most relevant features. It sows tat one can largely reduce te number of features needed for te neural network wen using gradient-based information. Additionally, te proposed metod demonstrates tat a common adaptive framework can be used to support te quality estimation for bot compression metods. Te performance of te metod is evaluated wit a publicly available database of images and teir quality score. Te results sow tat our proposed no-reference metod for te quality prediction of JPEG and JPEG2000 coded images as a comparable performance to te leading metrics available in literature, but at a considerably lower complexity. Keywords: Image quality assessment, objective metric, JPEG, JPEG2000, neural network. INTRODUCTION Te development of electronic imaging and multimedia tecniques as pused te demand for reliable quality assessment. Understanding and evaluating image quality as become increasingly important for a broad range of applications, suc as te optimization of digital imaging systems, te bencmarking of image and video coding algoritms, and te quality monitoring and control in displays []. Traditionally, image quality as been evaluated by uman subjects, and a mean opinion score (MOS) as been used to represent te image quality perceived by an averaged viewer. Wen conducted properly, subjective experiments are considered as te most reliable means of assessing image quality. However, performing subjective experiments is very time-consuming, and as a consequence, very expensive and too slow to be useful in real-world applications. Terefore, during te last decades, a lot of researc effort as been devoted to te development of objective metrics tat automatically and quantitatively can predict image quality as perceived by an averaged viewer. Objective metrics reported in literature range from dedicated metrics tat measure a specific image distortion to general metrics tat assess te overall perceived quality. Bot te dedicated and general metrics can be classified into fullreference (FR) or no-reference (NR) metrics, depending on weter te distorted image is compared to te original image or video. FR metrics measure te similarity or fidelity between te distorted image and its original version, were te latter is considered as a distortion-free reference. Te most widely used FR metrics are te mean squared error (MSE) and te peak signal-to-noise ratio (PSNR), bot aiming for an overall quality assessment. Improved alternatives of tese two basic general metrics include e.g. te structural similarity (SSIM) index [2] and te visual information fidelity (VIF) index [3]. Since FR metrics require te access to te original, wic is not available in (most) real-time systems, teir applicability is limited to in-lab (off-line) testing of image and video processing algoritms. Instead for real-time applications, NR metrics, in wic te quality prediction is based on te distorted image only, i.e. witout any reference, are more practical. Designing NR metrics, owever, is still callenging mainly due to te limited understanding of ow te uman visual system (HVS) affects image quality assessment. Human Vision and Electronic Imaging XV, edited by Bernice E. Rogowitz, Trasyvoulos N. Pappas, Proc. of SPIE-IS&T Electronic Imaging, SPIE Vol. 7527, 7527F 200 SPIE-IS&T CCC code: X/0/$8 doi: 0.7/ SPIE-IS&T/ Vol F-

2 During te last decades, tere is considerable progress in te development of NR metrics, as can be seen from some successful metods reported in te literature [4]-[]. A large number of NR metrics, suc as e.g. in [4]-[7], are dedicated metrics measuring a specific type of artifact created by a specific image distortion process, suc as a metric measuring sensor noise, ringing or blockiness as a consequence of signal compression, or blur caused during acquisition. In suc a scenario, te design of te NR metric can make use of te specific caracteristics of te artifact, and terefore, generally obtains a iger reliability wit respect to te perceived quality degradation. Te design of specific NR metrics is particularly beneficial for e.g. video cain optimization [2]. For te prediction of te overall perceived quality, approaces suc as pooling te local distortions into an overall quality score, and combining different artifacts inerent in an image ave been reported [6] and [0], but are studied only to a limited extent. In [], natural scene statistics were used to blindly measure te overall quality of images compressed by JPEG2000. Te approac relied on te assumption tat typical natural images exibit strong statistical regularities, and terefore, reside in a tiny area of te space containing all possible images. Based on tis assumption, te approac quantified image quality by detecting variations in te statistics of image features in te wavelet domain. Instead of precisely modeling specific types of artifacts or natural scene statistics, some approaces suc as e.g. in [8] and [9] attempt to formulate NR image quality assessment as a macine learning problem. Tey treat te HVS as a black box, wose input-output relationsip between image caracteristics and a quality rating is to be learned by computational intelligent tools, suc as neural networks. Te problem is generally formulated as a regression or function approximation approac, and te training data are obtained from extensive subjective experiments. Te goal is to train te model so tat te error between te desired output (i.e. te subjective quality rating) and te model prediction is minimized. Te approac was proved to be effective for te overall quality prediction of JPEG compressed images (see [8] and [9]), but at te expense of te extraction of a considerable number of image features, suc as general pixel-based features in [8] or HVS-based features in [9], as input to te neural network. Tis potentially increases te model s complexity, and consequently, may affect te prediction accuracy of te metric. Because of te widespread use of compression, a lot of researc is devoted to measuring te image quality after JPEG and JPEG2000 coding and decoding. In tis paper, we furter rely on te approac of using neural networks for te NR assessment of perceived overall quality of JPEG/JPEG2000 coded images. We optimize te feature-extraction stage by efficiently selecting and calculating te most relevant features, tus providing a simple yet efficient alternative for realtime implementation. It sould be noted tat te wole process is built up on te luminance component of images only in order to furter reduce te computational load. 2. FEATURE EXTRACTION Figure. Illustration of blocking and blur artifacts in a JPEG and a JPEG2000 compressed image, respectively. Literature as sown tat te image quality of JPEG compressed images is igly correlated wit te occurrence of blocking artifacts [7]; wile te quality of JPEG2000 compressed images is igly correlated wit te occurrence of blur [0]. A blocking artifact manifests itself as an artificial discontinuity in te image content, wic is a direct consequence of te fact tat te quantization in JPEG is block-based and tat te blocks are quantized independently [3]. A blur artifact occurs in JPEG2000 mainly due to te loss of ig frequency transform coefficients in te wavelet-based coding, SPIE-IS&T/ Vol F-2

3 as a result of wic te image signal is smootened [0]. Figure illustrates te occurrence of blocking artifacts in a JPEG compressed image, and of blur artifacts in a JPEG2000 compressed image. Quality degradation as a consequence of compression would ence easily be predictable from te extraction of blockiness and blur related image features, provided tat an adaptive tool suc as a neural network is used to empirically learn te igly non-linear relationsip between tese features and te quality rating. To efficiently caracterize te local beavior of artifacts and tus to feed te neural network wit relevant features for image quality prediction, a gradientbased feature-extraction sceme is proposed. It contains tree basic components: () te localization of artifacts, (2) te local feature extraction using local gradients in relation to teir neigborood, and (3) te assembling of global statistical descriptors as inputs to te neural network. How to implement eac of tese steps for te quality prediction of JPEG and JPEG2000 compressed images is detailed below. 2. Local feature extraction: JPEG Due to te underlying algoritm for JPEG compression, te spatial location of blocking artifacts is very regular. In principle, tey occur on a grid of blocks of 8 8 pixels, starting at te top-left corner of an image. In real-life applications, owever, te grid size may be different and its starting position may be sifted due to deviations in te incoming signal or as a consequence of spatial scaling. As a consequence, a NR metric runs te risk of calculating blockiness at wrong pixel positions, wic consequently degrades its accuracy. To ensure tat te metric is calculated at te exact position of te block boundaries, a grid detector is adopted. Te blocking grid detection metod proposed in [7] is implemented in tis paper, but it sould be noted tat te feature extraction approac proposed ere is independent of te particular coice of grid detector, and so, any oter alternative can be conveniently used instead. Te blocking grid detection metod of [7] first maps an image onto a -D signal profile, in wic te periodic property of blocking artifacts is maintained. Ten te exact block size as well as te grid offset are easily extracted from te discrete Fourier transform (DFT) of tis -D signal profile. Figure 2. Illustration of te template for calculating te local blockiness: (a) two adjacent 8 8 blocks (i.e. A and B) extracted from a real JPEG image, and (b) te gradient profile of te image patc of (a). Wen te blocking artifacts are (exactly) located, teir local beavior can be explicitly modeled. In tis paper, te local feature extraction for te JPEG compressed images is based on calculating te visual strengt of a blocking artifact witin a local area of te image content [7]. Since a blocking artifact is a local edge tat stands out from its spatial vicinity, it can be simply defined relating te energy present in te gradient at te artifact to te energy present in te gradient in its neigboring pixels. Wen we denote te luminance cannel of an image of M N (eigt widt) pixels as I(i, j) for iϵ[, M], jϵ[, N], te local blockiness L blockiness- along te orizontal direction at location (i, j) is quantified as L blockiness- (i, j)= 2n G ( i, j) x= n,..., n, x 0 G ( i, j + x) (i, j) ϵ {location of blocking} () were G (i, j) indicates te gradient map along te orizontal direction, and it is computed as G (i, j)= I ( i, j + ) I( i, j), j [, N ] (2) SPIE-IS&T/ Vol F-3

4 An example of te template for calculating te L blockiness- is sown in Figure 2, were two adjacent blocks of 8 8 pixels (i.e. A and B) are extracted from a real JPEG image. Te local blockiness along te vertical direction L blockiness-v can be similarly calculated. Te iger te values of L blockiness- and L blockiness-v te larger te distortion of te blocking artifact is. It sould, owever, be noted tat tis does not necessarily mean tat te blocking artifact is also more visible. Te local visibility of a blocking artifact may be affected by texture and luminance masking, wic typically occur in te HVS. It as been sown in literature tat taking into account tese masking effects can be greatly beneficial for te prediction performance of a NR blocking metric [7]. However, modeling te HVS introduces more computational power. So, in tis paper we avoid te calculation of masking, and rely on te neural network to learn te unknown functional relationsip between te extracted gradient-based features and te rating of overall image quality. 2.2 Local feature extraction: JPEG2000 In JPEG2000 compression, blur artifacts are perceptually prominent along edges or in textured areas. Measuring te smooting or smearing effect on strong edges as been proved to be an effective approac to approximate te overall perceived quality of JPEG2000 compressed images [0]. In tis paper, te local feature extraction for te JPEG2000 compressed images is built upon calculating te degree of blur at an edge witin a local area of image content. To detect strong edges, and consequently to identify te spatial location of blur artifacts, a variety of tecniques as been proposed in literature (e.g. PEM in [5]). However, to maintain a low complexity of te system, just a Sobel edge detector is adopted ere. Te location of te strong edges is ten extracted by applying a tresold to te resulting gradient image (i.e. by removing noise and insignificant edges) OLTTJJ Figure 3. Illustration of te calculation of te local blur: (a) te intensity profile over a detected edge (i.e. at location (3, 259) in Figure (b)), (b) te intensity profile over te corresponding pixels of (a) in te original image of Figure (b), (c) te gradient profile of (a), and (d) te gradient profile of (b). SPIE-IS&T/ Vol F-4

5 For all pixels located along te detected edges, local blur is defined as te sarpness of te edge in te gradient domain instead of calculating te distance between te start and end position of an edge (i.e. te edge spread along te orizontal/vertical direction) as proposed in [0]. Wen describing blur simply as te relative gradient energy of an edge compared to its direct vicinity, it can be quantified in te same manner as used in (), i.e.: L blur- (i, j)= 2n G ( i, j) x= n,..., n, x 0 G ( i, j + x) (i, j) ϵ {location of blur} (3) were L blur- indicates te local blur along te orizontal direction. L blur-v, i.e. te local blur in te vertical direction, can be calculated similarly. Te lower te value of L blur- and L blur-v, te larger te distortion of te blur artifact is. Figure 3 explains te reasoning beind te proposed approac of using gradient energy to detect blur. Figure 3(a) sows te intensity profile over a detected edge (i.e. at location (3, 259) of te image in Figure (b)), wile Figure 3(b) sows te corresponding pixel intensity values in te original image of Figure (b). Te difference in sarpness between te two edges is clearly revealed in te gradient domain (see Figure 3(c) and (d)). In correspondence, te values of L blur- indicate tat te edge of Figure 3(a) is more blurred tan te edge of Figure 3(b). 2.3 Global descriptor of te image features To gater te local distortion information in a more compact format, in order to build an informative global descriptor to feed te neural network, a refined pooling strategy of te one proposed in [8] is adopted. It results in a vector tat represents eac image. Tis vector includes a statistical representation of te distortion distribution in te image. Having computed te feature values m i (i=,..., N M ) per image (i.e. L blockiness calculated on te blocking grid or L blur calculated on te detected edges), tese values are sorted in ascending order of magnitude. Te envelope of te obtained distribution is ten expressed in te global distortion descriptor m by taking of its percentiles φ: m = { ϕα ; α { 0,0,20,30,40,50,60,70,80,90,00 }; ϕα = α (4) Compared to simply taking te average of te feature values, tis spatial pooling strategy allows feeding te non-linear regression wit a more complete overview of te amount and beavior of te considered distortion in te image. N M 3. NR IMAGE QUALITY ESTIMATOR Te image quality assessment community is accustomed to make use of non-linear mapping strategies to improve te correlation between objective metric predictions and uman quality assessments []. Several approaces can be used to tis end, ranging from logistic fitting [] to macine learning (ML) [8], [9] and [4]. In te case of perception-related problems, macine learning metods are powerful tools. Even toug tey don t require any a-priori information, ML metods still allow inferring igly non-linear relationsips between numerical descriptions of images and teir subjective assessment, were te mapping is learnt from examples (i.e. real-world observations). Moreover, wen te computational constraints are tigt, suc as in real-time applications, te accuracy in te mapping granted by ML metods allows designing simple metrics wit low computational requirements. In te system proposed ere, a feed-forward neural network maps te blockiness and blur features into te associated estimates of perceived quality. Neural networks learn te non-linear dependency of target values tˆ from inputs x as a series expansion of n basis functions a, often sigmoids [5]. n t = y = ( x) = ( x) ˆ w a (5) Te MultiLayer Perceptron (MLP) paradigm [5] belongs to tis class of networks, and as been proved to perform effectively in problems were te target-mapping function can be retrieved by a few computing units endowed wit global scope. MLPs aim at implementing a stimulus-response beavior by arranging suc computing units ( neurons ) into a layered network: eac unit involves a non-linear transformation of weigted inputs; eac layer outputs are furter non-linearly transformed by te neurons in te next layer. SPIE-IS&T/ Vol F-5

6 Te Circular Back Propagation (CBP) network [6] is an extension of te MLP paradigm, based on te addition of one more input value, being te sum of te squared values of all te network inputs. It as been proven tat tis addition allows te network to switc between te classic, sigmoidal beavior, and a smooter, bell-saped radial function [6]. Te modeling process is entirely data-driven, and fully defined in te training pase. Tanks to tis adaptive beavior, CBP networks are quite appropriate for perception modeling, were often te underlying cognitive mecanisms are not clear. Te CBP arcitecture can be formally described as follows. Given an input vector x={x,, x ni } for a given stimulus and its corresponding target value t of dimensionality n o, te input layer is structured in n i neurons, eac connected to every neuron in te following idden layer. Te j-t idden neuron performs a non-linear transformation of a weigted combination of te input values, wit coefficients w j,i ( j=,, n ; i=,, n i ): a j ni ni 2 = sigm wj, 0 + wj, ixi + wj, n + x i i (6) i= i= were sigm(z)=(+e -z ) -, and a j is te neuron activation (i.e. te output of te basis function). Te output layer provides te actual network predictions, y k, (k =,, n o ), in a similar way: y k n = sigm wk, 0 + wk, ja j (7) j= In te case of quality assessment, n o =, and te target value t is just te quality score represented by a scalar value. Te structural CBP modification does not interfere wit te possibility of using te well-known back propagation algoritm [5] for model training. Te cost to be minimized is expressed troug a quadratic function comparing te target and te predicted value: E = n n o n p no ( l) ( l) ( t y ) k k p l= k= were n p is te number of training patterns. Togeter wit te weigting coefficients w j,i and w k,j, te oter degree of freedom of te neural network tat as to be set is te number of idden units, n. For tis task, te present researc followed an empirical approac [7] mainly because of its simplicity and proved effectiveness. Te wole parameter setting process is completed in te training pase, and ence does not entail computational overead at runtime. 4. EXPERIMENTAL RESULTS To evaluate te performance of te proposed approac, te LIVE image quality assessment database [8] was used. It consists of a set of twenty nine ig-resolution and ig-quality color source images tat reflect adequate diversity in image content. Tese images were compressed using JPEG at a bit rate ranging from 0.5 bits per pixel (bpp) to 3.34 bpp, resulting in a database of 233 JPEG compressed stimuli (including te originals). Te same source images were also compressed using JPEG2000 at a bit rate ranging from bpp to 3.5 bpp, yielding a database of 227 JPEG2000 compressed stimuli (including te originals). An extensive psycovisual experiment was conducted to assign a difference mean opinion score (DMOS) to eac stimulus. Te DMOS was measured on a continuous linear scale tat was divided into five intervals marked wit te adjectives Bad, Poor, Fair, Good and Excellent. In our performance evaluation, te source images were divided into two groups, i.e. 20 out of 29 source images were used for training and te remaining 9 images were used for testing. Tis resulted in 6 stimuli for training and 72 stimuli for testing in te JPEG database for evaluating our JPEG metric, and 56 stimuli for training and 7 stimuli for testing in te JPEG2000 database for evaluating our JPEG2000 metric. For eac stimulus, a vector containing eleven percentiles of te distribution of te local blockiness/blur features was taken as te input to te neural network, wic was equipped wit 3 idden neurons. Figure 4 sows te scatter plots of te DMOS versus te quality prediction based on our neural network approac, for te JPEG and JPEG2000 database of test images. 2 (8) SPIE-IS&T/ Vol F-6

7 Te performance of an objective metric can be evaluated wit respect to its ability to predict subjective quality ratings (te DMOS). Two values can be employed to caracterize tis ability: i.e. te Pearson linear correlation coefficient, and te root mean square error (RMSE) [9]. Based on tese two values te JPEG/JPEG2000 metrics proposed in tis paper are compared to state-of-te-art NR metrics, including four metrics for JPEG [6]-[9] and two metrics for JPEG2000 [0] and []. Table and 2 list te Pearson correlation coefficient and te RMSE for our metrics as well as for te state-ofte-art metrics. It sould be noted tat we ad no access to te test environment of te metrics in [6]-[], and te comparison among metrics migt be biased due to e.g. a different selection of te disjoint sets for training and testing. However, te metrics mentioned above were all evaluated wit te LIVE database; ence te performance comparison is considered fairly valuable. For te NR metrics for JPEG compression, our proposed metric outperforms te metrics of [6] and [7] in predicting te overall perceived quality. Its performance is comparable to tat of te metrics of [8] and [9], wic are also based on a neural network approac. Te advantage of our metric lies in its simplicity in te extracted features, compared to te metrics of [8] and [9]. For te NR metrics for JPEG2000 compression, our proposed metric clearly outperforms te metric in [0] mainly due to te powerful neural network, wic can efficiently approximate te functional relationsip between te extracted local blur features and te rating of overall quality. Te performance of our proposed JPEG2000 metric is comparable to te metric in [], but it is indeed a simple yet efficient alternative useful for real-time implementation. 0 9 JPEG 8 7 DMOS Proposed Metric 0 9 JPEG DMOS Proposed Metric Figure 4: Scatter plots of DMOS versus te proposed metric (based on te neural network approac) for JPEG and JPEG2000. SPIE-IS&T/ Vol F-7

8 Table. Performance comparison of our approac wit state-of-te-art metrics for JPEG. NR Metric (JPEG) Pearson Correlation Coefficient Root Mean Square Error (RMSE) Proposed (scale [-, ]) Wang et al [6] 0.93 N/A Liu et al [7] 0.98 N/A Gastaldo et al [8] (scale [-,]) Babu et al [9] N/A (scale [, 0] Table 2. Performance comparison of our approac wit state-of-te-art metrics for JPEG2000. NR Metric (JPEG2000) Pearson Correlation Coefficient Root Mean Square Error (RMSE) Proposed (scale [-, ]) Marziliano et al [0] 0.85 N/A Seik et al [] 0.93 N/A 5. CONCLUSIONS In tis paper, we present a novel approac to assess te overall perceived quality of JPEG/JPEG2000 compressed images, witout te access to teir original version. Te approac extracts te most relevant features tat are representative for te quality degradation as a consequence of compression, i.e. te local blockiness in te JPEG compressed images, and te local blur in te JPEG2000 compressed images. Te igly non-linear relationsip between te extracted features and te quality rating is empirically learned by a CBP neural network. Tis approac intrinsically takes advantage of te prior knowledge on te specific caracteristics of JPEG/JPEG2000 compression artifacts, tus optimizing te metric by efficiently calculating te image features. Te performance of te proposed NR JPEG/JPEG2000 metrics is evaluated wit te use of te LIVE database, and compared to several leading alternatives in literature. Experimental results sow tat our metrics result in a strong correlation wit subjective data at a reduced computational load. As suc, te proposed approac is promising in terms of bot computational efficiency and practical reliability for real-time applications. REFERENCES [] Wang, Z. and Bovik, A. C., [Modern Image Quality Assessment], Syntesis Lectures on Image, Video & Multimedia Processing, Morgan & Claypool Publisers (2006). [2] Wang, Z., Bovik, A. C., Seik, H. R. and Simoncelli, E. P., "Image quality assessment: from error visibility to structural similarity," IEEE Transactions on Image Processing, vol.3, no.4, pp (2004). [3] Seik, H. R. and Bovik, A. C., "Image information and visual quality," IEEE Transactions on Image Processing, vol.5, no.2, pp (2006). [4] Ferzli, R. and Karam, L. J., A no-reference objective image sarpness metric based on te notion of just noticeable blur (JNB), IEEE Transactions on Image Processing, vol. 8, pp (2009). [5] Liu, H., Klomp, N. and Heynderickx, I., "A No-Reference Metric for Perceived Ringing," in Proc. Fourt International Worksop on Video Processing and Quality Metrics for Consumer Electronics (2009). [6] Wang, Z., Seik, H. R. and Bovik, A. C., "No-reference perceptual quality assessment of JPEG compressed images", Proc. IEEE International Conference on Image Processing (2002). SPIE-IS&T/ Vol F-8

9 [7] Liu, H. and Heynderickx, I., A Perceptually Relevant No-Reference Blockiness Metric Based on Local Image Caracteristics, EURASIP Journal on Advances in Signal Processing, vol (2009). [8] Gastaldo, P. and Zunino, R., Neural networks for te no-reference assessment of perceived quality, Journal of. Electronic Imaging, 4 (3), (2005). [9] Babu, R. V., Sures, S. and Perkis, A., No-reference JPEG-image quality assessment using GAP-RBF, Signal Processing, vol. 87, no.6, pp (2007). [0] Marziliano, P., Dufax, F., Winkler, S. and Ebraimi, T., Perceptual blur and ringing metrics: Application to JPEG2000, Signal Processing: Image Communication, vol. 9, pp (2004). [] Seik, H. R., Bovik, A. C. and Cormack, L. K., "No-Reference Quality Assessment Using Natural Scene Statistics: JPEG2000," IEEE Transactions on Image Processing, vol. 4, no. 2 (2005). [2] Zon, K. and Ali, W., Automated video cain optimization, IEEE Transactions on Consumer Electronics, vol. 47, pp (200). [3] Yuen, M. and Wu, H. R., A survey of ybrid MC/DPCM/DCT video coding distortions, Signal Processing, vol. 70, no. 3, pp (998). [4] Redi, J., Gastaldo, P., Zunino, R. and Heynderickx, I., Reduced reference assessment of perceived quality by exploiting color information, Fourt International Worksop on Video Processing and Quality Metrics for Consumer Electronics (2009). [5] Rumelart, D. E. and McClelland, J. L., [Parallel distributed processing], MIT Press, Cambridge, MA (986). [6] Ridella, S., Rovetta, S. and Zunino, R., Circular back-propagation networks for classification, IEEE Trans. on Neural Networks, vol. 8, pp (997). [7] Widrow, B. and Ler, M. A., 30 Years of Adaptive Neural Networks: Perceptron, Madaline and Back Propagation, Proc. IEEE. 78(9), (990). [8] Seik, H. R., Wang, Z., Cormack, L. and Bovik, A. C., "LIVE Image Quality Assessment Database Release 2," ttp://live.ece.utexas.edu/researc/quality [9] VQEG (2003, Aug.): Final report from te video quality experts group on te validation of objective models of video quality assessment. Available: ttp:// SPIE-IS&T/ Vol F-9