subpixel misalignments •Each imposes a set of linear constraints •On the unknown high-resolution intensity values •Enough low-resolution images available Classical Multi-Image
subpixel misalignments •Each imposes a set of linear constraints •On the unknown high-resolution intensity values •Enough low-resolution images available •Set of equations becomes determined & can be solved Classical Multi-Image
subpixel misalignments •Each imposes a set of linear constraints •On the unknown high-resolution intensity values •Enough low-resolution images available •Set of equations becomes determined & can be solved •Limited only to little increase in resolution Classical Multi-Image
subpixel misalignments •Each imposes a set of linear constraints •On the unknown high-resolution intensity values •Enough low-resolution images available •Set of equations becomes determined & can be solved •Limited only to little increase in resolution •By factors smaller than 2 Classical Multi-Image
high resolution images •Learned from database of low-res/high-res image pairs •Usually with a relative scale factor of 2 •Applied to a new low-resolution image Example-based
high resolution images •Learned from database of low-res/high-res image pairs •Usually with a relative scale factor of 2 •Applied to a new low-resolution image •Recover its most likely high-resolution version Example-based
high resolution images •Learned from database of low-res/high-res image pairs •Usually with a relative scale factor of 2 •Applied to a new low-resolution image •Recover its most likely high-resolution version •Exceeds the limits of classical SR Example-based
high resolution images •Learned from database of low-res/high-res image pairs •Usually with a relative scale factor of 2 •Applied to a new low-resolution image •Recover its most likely high-resolution version •Exceeds the limits of classical SR •But not guaranteed to provide the true (unknown) details Example-based
•Magnify (up-scale) while maintaining the edge sharpness •Super-Resolution •Recover new missing high-resolution details •Not explicitly found in any individual low-resolution image
•Magnify (up-scale) while maintaining the edge sharpness •Super-Resolution •Recover new missing high-resolution details •Not explicitly found in any individual low-resolution image •Details beyond the Nyquist frequency
•Magnify (up-scale) while maintaining the edge sharpness •Super-Resolution •Recover new missing high-resolution details •Not explicitly found in any individual low-resolution image •Details beyond the Nyquist frequency •Classical (multi-image)
•Magnify (up-scale) while maintaining the edge sharpness •Super-Resolution •Recover new missing high-resolution details •Not explicitly found in any individual low-resolution image •Details beyond the Nyquist frequency •Classical (multi-image) •High-frequency information are split across multiple low- resolution images (in subpixel aliased form)
•Magnify (up-scale) while maintaining the edge sharpness •Super-Resolution •Recover new missing high-resolution details •Not explicitly found in any individual low-resolution image •Details beyond the Nyquist frequency •Classical (multi-image) •High-frequency information are split across multiple low- resolution images (in subpixel aliased form) •Example-based
•Magnify (up-scale) while maintaining the edge sharpness •Super-Resolution •Recover new missing high-resolution details •Not explicitly found in any individual low-resolution image •Details beyond the Nyquist frequency •Classical (multi-image) •High-frequency information are split across multiple low- resolution images (in subpixel aliased form) •Example-based •High-resolution information are available in patches database, and learned from low-res/high-res pairs
to redundantly recur many times inside the image •within the same scale •Classical SR constraints •and across different scales (coarser) •Provides low-res/high-res pairs for Exampled-based
to redundantly recur many times inside the image •within the same scale •Classical SR constraints •and across different scales (coarser) •Provides low-res/high-res pairs for Exampled-based •Image Completion, Image Re-targeting, Image Denoising
to redundantly recur many times inside the image •within the same scale •Classical SR constraints •and across different scales (coarser) •Provides low-res/high-res pairs for Exampled-based •Image Completion, Image Re-targeting, Image Denoising •Statistically proved later
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale •Generate down-scaled {Is} from I, as scale factor = 1.25s
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale •Generate down-scaled {Is} from I, as scale factor = 1.25s •s = 0, -1, ..., -6 (I0 = I), smallest one is 1.25-6 = 0.26 of input I
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale •Generate down-scaled {Is} from I, as scale factor = 1.25s •s = 0, -1, ..., -6 (I0 = I), smallest one is 1.25-6 = 0.26 of input I •Each patches in I are compared against all patches in {Is}
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale •Generate down-scaled {Is} from I, as scale factor = 1.25s •s = 0, -1, ..., -6 (I0 = I), smallest one is 1.25-6 = 0.26 of input I •Each patches in I are compared against all patches in {Is} •without their DC (average grayscale)
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale •Generate down-scaled {Is} from I, as scale factor = 1.25s •s = 0, -1, ..., -6 (I0 = I), smallest one is 1.25-6 = 0.26 of input I •Each patches in I are compared against all patches in {Is} •without their DC (average grayscale) •Distance function: gaussian-weighted SSD
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale •Generate down-scaled {Is} from I, as scale factor = 1.25s •s = 0, -1, ..., -6 (I0 = I), smallest one is 1.25-6 = 0.26 of input I •Each patches in I are compared against all patches in {Is} •without their DC (average grayscale) •Distance function: gaussian-weighted SSD •Textured patches have much larger SSD errors
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale •Generate down-scaled {Is} from I, as scale factor = 1.25s •s = 0, -1, ..., -6 (I0 = I), smallest one is 1.25-6 = 0.26 of input I •Each patches in I are compared against all patches in {Is} •without their DC (average grayscale) •Distance function: gaussian-weighted SSD •Textured patches have much larger SSD errors •“Good Distance”: SSD with misaligned (0.5px) duplicate
perceive any obvious repetitive structure •vary small patches contain only an edge, a corner, etc •Tested on Berkeley Segmentation Database (BSDS) •Each image I in the database converted to grayscale •Generate down-scaled {Is} from I, as scale factor = 1.25s •s = 0, -1, ..., -6 (I0 = I), smallest one is 1.25-6 = 0.26 of input I •Each patches in I are compared against all patches in {Is} •without their DC (average grayscale) •Distance function: gaussian-weighted SSD •Textured patches have much larger SSD errors •“Good Distance”: SSD with misaligned (0.5px) duplicate •Below “Good Distance” means similar
source patches •with the highest intensity variance •Excludes the uniform and low-frequency patches *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
source patches •with the highest intensity variance •Excludes the uniform and low-frequency patches *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
of low-res images {L1, ..., Ln} given •of the same scene (at subpixel misalignments) •Lj = (H ∗ Bj)↓Sj •↓ denotes a subsampling operation, Sj is subsampling rate
of low-res images {L1, ..., Ln} given •of the same scene (at subpixel misalignments) •Lj = (H ∗ Bj)↓Sj •↓ denotes a subsampling operation, Sj is subsampling rate •Bj(q) is the corresponding Point Spread Function (PSF)
of low-res images {L1, ..., Ln} given •of the same scene (at subpixel misalignments) •Lj = (H ∗ Bj)↓Sj •↓ denotes a subsampling operation, Sj is subsampling rate •Bj(q) is the corresponding Point Spread Function (PSF) •Each low-res pixel p = (x, y) in Lj induces the linear constraint Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q)
of low-res images {L1, ..., Ln} given •of the same scene (at subpixel misalignments) •Lj = (H ∗ Bj)↓Sj •↓ denotes a subsampling operation, Sj is subsampling rate •Bj(q) is the corresponding Point Spread Function (PSF) •Each low-res pixel p = (x, y) in Lj induces the linear constraint Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q) •on the unknown high-res intensity values {H(Qi} within the neighborhood around its corresponding high-res pixel q ∈ H
of low-res images {L1, ..., Ln} given •of the same scene (at subpixel misalignments) •Lj = (H ∗ Bj)↓Sj •↓ denotes a subsampling operation, Sj is subsampling rate •Bj(q) is the corresponding Point Spread Function (PSF) •Each low-res pixel p = (x, y) in Lj induces the linear constraint Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q) •on the unknown high-res intensity values {H(Qi} within the neighborhood around its corresponding high-res pixel q ∈ H •size determined by the support of blur kernel B
image L = (H ∗ B)↓S •Recovering H becomes under-determined •constraints induced by L are fewer than the unknowns in H •Let p be a pixel in L, and P be its surrounding patch (5 x 5)
image L = (H ∗ B)↓S •Recovering H becomes under-determined •constraints induced by L are fewer than the unknowns in H •Let p be a pixel in L, and P be its surrounding patch (5 x 5) •Multiple similar patches P1, ..., Pk in L (at subpixel shifts)
image L = (H ∗ B)↓S •Recovering H becomes under-determined •constraints induced by L are fewer than the unknowns in H •Let p be a pixel in L, and P be its surrounding patch (5 x 5) •Multiple similar patches P1, ..., Pk in L (at subpixel shifts) •Treated as if taken from k different low-res images
image L = (H ∗ B)↓S •Recovering H becomes under-determined •constraints induced by L are fewer than the unknowns in H •Let p be a pixel in L, and P be its surrounding patch (5 x 5) •Multiple similar patches P1, ..., Pk in L (at subpixel shifts) •Treated as if taken from k different low-res images •of the “same scene”, of course
image L = (H ∗ B)↓S •Recovering H becomes under-determined •constraints induced by L are fewer than the unknowns in H •Let p be a pixel in L, and P be its surrounding patch (5 x 5) •Multiple similar patches P1, ..., Pk in L (at subpixel shifts) •Treated as if taken from k different low-res images •of the “same scene”, of course •Each equation induced by Pi is scaled by the degree of similarity of Pi to its patch source P
image L = (H ∗ B)↓S •Recovering H becomes under-determined •constraints induced by L are fewer than the unknowns in H •Let p be a pixel in L, and P be its surrounding patch (5 x 5) •Multiple similar patches P1, ..., Pk in L (at subpixel shifts) •Treated as if taken from k different low-res images •of the “same scene”, of course •Each equation induced by Pi is scaled by the degree of similarity of Pi to its patch source P •for increased numerical stability
image L = (H ∗ B)↓S •Recovering H becomes under-determined •constraints induced by L are fewer than the unknowns in H •Let p be a pixel in L, and P be its surrounding patch (5 x 5) •Multiple similar patches P1, ..., Pk in L (at subpixel shifts) •Treated as if taken from k different low-res images •of the “same scene”, of course •Each equation induced by Pi is scaled by the degree of similarity of Pi to its patch source P •for increased numerical stability •patches of high similarity to P have stronger influence
pixel in L, find its k nearest patch neighbors in same L •eg. Approximate Nearest Neighbor with k = 9 *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
pixel in L, find its k nearest patch neighbors in same L •eg. Approximate Nearest Neighbor with k = 9 •compute their subpixel alignment (at 1/s pixel shifts) *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
pixel in L, find its k nearest patch neighbors in same L •eg. Approximate Nearest Neighbor with k = 9 •compute their subpixel alignment (at 1/s pixel shifts) •(where s is the scale factor) *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
pixel in L, find its k nearest patch neighbors in same L •eg. Approximate Nearest Neighbor with k = 9 •compute their subpixel alignment (at 1/s pixel shifts) •(where s is the scale factor) •assuming sufficient neighbors are found *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
pixel in L, find its k nearest patch neighbors in same L •eg. Approximate Nearest Neighbor with k = 9 •compute their subpixel alignment (at 1/s pixel shifts) •(where s is the scale factor) •assuming sufficient neighbors are found •determined set of linear constraints on the intensity in H *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
pixel in L, find its k nearest patch neighbors in same L •eg. Approximate Nearest Neighbor with k = 9 •compute their subpixel alignment (at 1/s pixel shifts) •(where s is the scale factor) •assuming sufficient neighbors are found •determined set of linear constraints on the intensity in H •scale each equation by its reliability (patch similarity) *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
pixel in L, find its k nearest patch neighbors in same L •eg. Approximate Nearest Neighbor with k = 9 •compute their subpixel alignment (at 1/s pixel shifts) •(where s is the scale factor) •assuming sufficient neighbors are found •determined set of linear constraints on the intensity in H •scale each equation by its reliability (patch similarity) *NBHF$SFEJU(MBTOFSFUBM l4VQFS3FTPMVUJPOGSPNB4JOHMF*NBHFz
and L = (H ∗ B)↓S •I0, I1, ..., In denote a cascade of unknown images of increasing resolutions (scales) ranging from L to H •I0 = L, In = H, with corresponding blur functions B0, ..., Bn
and L = (H ∗ B)↓S •I0, I1, ..., In denote a cascade of unknown images of increasing resolutions (scales) ranging from L to H •I0 = L, In = H, with corresponding blur functions B0, ..., Bn •Bn = B is the PSF relating H to L; B0 is the δ function
and L = (H ∗ B)↓S •I0, I1, ..., In denote a cascade of unknown images of increasing resolutions (scales) ranging from L to H •I0 = L, In = H, with corresponding blur functions B0, ..., Bn •Bn = B is the PSF relating H to L; B0 is the δ function •Ij satisfies L = (Ij ∗Bj)↓Sj
and L = (H ∗ B)↓S •I0, I1, ..., In denote a cascade of unknown images of increasing resolutions (scales) ranging from L to H •I0 = L, In = H, with corresponding blur functions B0, ..., Bn •Bn = B is the PSF relating H to L; B0 is the δ function •Ij satisfies L = (Ij ∗Bj)↓Sj •Although {I0, ..., In} are unknown, the cascade blur kernels {B0, ..., Bn} can be assumed to be known
and L = (H ∗ B)↓S •I0, I1, ..., In denote a cascade of unknown images of increasing resolutions (scales) ranging from L to H •I0 = L, In = H, with corresponding blur functions B0, ..., Bn •Bn = B is the PSF relating H to L; B0 is the δ function •Ij satisfies L = (Ij ∗Bj)↓Sj •Although {I0, ..., In} are unknown, the cascade blur kernels {B0, ..., Bn} can be assumed to be known •PSF B can be approximated with a gaussian; Bj = B(sj) are a cascade of gaussians (variances determined by sj )
and L = (H ∗ B)↓S •I0, I1, ..., In denote a cascade of unknown images of increasing resolutions (scales) ranging from L to H •I0 = L, In = H, with corresponding blur functions B0, ..., Bn •Bn = B is the PSF relating H to L; B0 is the δ function •Ij satisfies L = (Ij ∗Bj)↓Sj •Although {I0, ..., In} are unknown, the cascade blur kernels {B0, ..., Bn} can be assumed to be known •PSF B can be approximated with a gaussian; Bj = B(sj) are a cascade of gaussians (variances determined by sj ) •scale factor sj = αj, the constraint Ij = (H ∗ Bn-1)↓Sn-1 still hold for all j
and L = (H ∗ B)↓S •I0, I1, ..., In denote a cascade of unknown images of increasing resolutions (scales) ranging from L to H •I0 = L, In = H, with corresponding blur functions B0, ..., Bn •Bn = B is the PSF relating H to L; B0 is the δ function •Ij satisfies L = (Ij ∗Bj)↓Sj •Although {I0, ..., In} are unknown, the cascade blur kernels {B0, ..., Bn} can be assumed to be known •PSF B can be approximated with a gaussian; Bj = B(sj) are a cascade of gaussians (variances determined by sj ) •scale factor sj = αj, the constraint Ij = (H ∗ Bn-1)↓Sn-1 still hold for all j •uniform scale factor guarantees that Ij and Ij+m are related by the same Bm , regardless of j
a cascade of unknown images of decreasing resolutions (scales) ranging L, using I-j = (L ∗Bj)↓Sj where j = 0, ..., m •Unlike high-res cascade, low-res {I-j} are known (computed from L)
a cascade of unknown images of decreasing resolutions (scales) ranging L, using I-j = (L ∗Bj)↓Sj where j = 0, ..., m •Unlike high-res cascade, low-res {I-j} are known (computed from L) •Let Pj(p) denote a patch in Ij at pixel location p
a cascade of unknown images of decreasing resolutions (scales) ranging L, using I-j = (L ∗Bj)↓Sj where j = 0, ..., m •Unlike high-res cascade, low-res {I-j} are known (computed from L) •Let Pj(p) denote a patch in Ij at pixel location p •For any pixel p ∈ L (L = I0 ) and its surround patch P0(p), we search k nearest similar patches within {I-j} where j > 0
a cascade of unknown images of decreasing resolutions (scales) ranging L, using I-j = (L ∗Bj)↓Sj where j = 0, ..., m •Unlike high-res cascade, low-res {I-j} are known (computed from L) •Let Pj(p) denote a patch in Ij at pixel location p •For any pixel p ∈ L (L = I0 ) and its surround patch P0(p), we search k nearest similar patches within {I-j} where j > 0 •Let P-j(ṕ) denote such matching patch in I-j , it’s high-res “parent” patch Q0(sj· ṕ) can be extracted from any level between L and I-j
a cascade of unknown images of decreasing resolutions (scales) ranging L, using I-j = (L ∗Bj)↓Sj where j = 0, ..., m •Unlike high-res cascade, low-res {I-j} are known (computed from L) •Let Pj(p) denote a patch in Ij at pixel location p •For any pixel p ∈ L (L = I0 ) and its surround patch P0(p), we search k nearest similar patches within {I-j} where j > 0 •Let P-j(ṕ) denote such matching patch in I-j , it’s high-res “parent” patch Q0(sj· ṕ) can be extracted from any level between L and I-j •low-res/high-res pair [P, Q]
a cascade of unknown images of decreasing resolutions (scales) ranging L, using I-j = (L ∗Bj)↓Sj where j = 0, ..., m •Unlike high-res cascade, low-res {I-j} are known (computed from L) •Let Pj(p) denote a patch in Ij at pixel location p •For any pixel p ∈ L (L = I0 ) and its surround patch P0(p), we search k nearest similar patches within {I-j} where j > 0 •Let P-j(ṕ) denote such matching patch in I-j , it’s high-res “parent” patch Q0(sj· ṕ) can be extracted from any level between L and I-j •low-res/high-res pair [P, Q] •high-res parent of low-res P0(p) in L is Qj(sj· ṕ) in unknown Ij ; it can be copied from extracted Q0(sj· ṕ)
a cascade of unknown images of decreasing resolutions (scales) ranging L, using I-j = (L ∗Bj)↓Sj where j = 0, ..., m •Unlike high-res cascade, low-res {I-j} are known (computed from L) •Let Pj(p) denote a patch in Ij at pixel location p •For any pixel p ∈ L (L = I0 ) and its surround patch P0(p), we search k nearest similar patches within {I-j} where j > 0 •Let P-j(ṕ) denote such matching patch in I-j , it’s high-res “parent” patch Q0(sj· ṕ) can be extracted from any level between L and I-j •low-res/high-res pair [P, Q] •high-res parent of low-res P0(p) in L is Qj(sj· ṕ) in unknown Ij ; it can be copied from extracted Q0(sj· ṕ)
a cascade of unknown images of decreasing resolutions (scales) ranging L, using I-j = (L ∗Bj)↓Sj where j = 0, ..., m •Unlike high-res cascade, low-res {I-j} are known (computed from L) •Let Pj(p) denote a patch in Ij at pixel location p •For any pixel p ∈ L (L = I0 ) and its surround patch P0(p), we search k nearest similar patches within {I-j} where j > 0 •Let P-j(ṕ) denote such matching patch in I-j , it’s high-res “parent” patch Q0(sj· ṕ) can be extracted from any level between L and I-j •low-res/high-res pair [P, Q] •high-res parent of low-res P0(p) in L is Qj(sj· ṕ) in unknown Ij ; it can be copied from extracted Q0(sj· ṕ) P0(p) find NN P-j(ṕ) parent Q0(sj· ṕ) copy Qj(sj· p)
pixel in L •large collection of high-res patch {Qj} between L and H •such “learned” high-res patch Qj induces linear constraints on the unknown target resolution H in the form of Classical SR
pixel in L •large collection of high-res patch {Qj} between L and H •such “learned” high-res patch Qj induces linear constraints on the unknown target resolution H in the form of Classical SR •Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q)
pixel in L •large collection of high-res patch {Qj} between L and H •such “learned” high-res patch Qj induces linear constraints on the unknown target resolution H in the form of Classical SR •Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q) •with more compactly supported blur kernel than B = PSF
pixel in L •large collection of high-res patch {Qj} between L and H •such “learned” high-res patch Qj induces linear constraints on the unknown target resolution H in the form of Classical SR •Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q) •with more compactly supported blur kernel than B = PSF •Solving Coarse-to-Fine
pixel in L •large collection of high-res patch {Qj} between L and H •such “learned” high-res patch Qj induces linear constraints on the unknown target resolution H in the form of Classical SR •Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q) •with more compactly supported blur kernel than B = PSF •Solving Coarse-to-Fine •use constant factor α = 1.25 (namely, sj = 1.25j)
pixel in L •large collection of high-res patch {Qj} between L and H •such “learned” high-res patch Qj induces linear constraints on the unknown target resolution H in the form of Classical SR •Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q) •with more compactly supported blur kernel than B = PSF •Solving Coarse-to-Fine •use constant factor α = 1.25 (namely, sj = 1.25j) •when solving equations for image Ij+1
pixel in L •large collection of high-res patch {Qj} between L and H •such “learned” high-res patch Qj induces linear constraints on the unknown target resolution H in the form of Classical SR •Lj(p) = (H ∗ Bj) (q) = ∑qi ∈ Support(Bj) H(qi) Bj(qi - q) •with more compactly supported blur kernel than B = PSF •Solving Coarse-to-Fine •use constant factor α = 1.25 (namely, sj = 1.25j) •when solving equations for image Ij+1 •employ newly recovered high-res images so far (I0, ..., Ij )
and H. C. Andrews, “Cubic Splines for Image Interpolation and Digital Filtering,” IEEE Trans. On Acoustics, Speech, and Signal Processing, 1981 • NEDI: X. Li and M. T. Orchard, “New Edge-Directed Interpolation,” IEEE Trans. On Image Processing, 2001 • GPP: J. Sun, J. Sun, Z. B. Xun and H. Y. Shum, “Image Super-resoluton using Gradient Profile Prior,” CVPR, 2008 • IUES: R. Fattal, “Image upsampling via imposed edge statistics,” Proceeding of ACM SIGGRAPH, 2007 • SRSI: D. Glasner, S. Bagon, M. Irani, “Super-Resolution from a Single Image,” ICCV, 2009
quality •PSNR & MSSIM of bilinear & cubic convolution are larger than new-edge directed interpolation & unified SR •Visual evaluation •primary metric •PSNR & MSSIM don’t provide accurate measure of visual quality in such case
Streaming on Mobile Devices Hong-Han Shuai, De-Nian Yang, Wen-Huang Cheng, Ming-Syan Chen Academia Sinica, National Taiwan University IEEE Transactions on Multimedia (TMM), 2011
danchen
fta98
masatoto
matsuolab
joisino
sgk
ksudoh
kentaitakura
daimoriwaki
masakat0
kuri8ive
keisuke198619
trishagee
jponch
mojombo
aarron
hannesfritz
jrom
notwaldorf
aleyda
chriscoyier
danielanewman
lynnandtonic
