| 主权项 |
1. An objective assessment method for a stereoscopic image quality combined with manifold characteristics and binocular characteristics, comprising steps of:
{circle around (1)} selecting multiple undistorted natural scene plane images, and extracting a luminance component from each undistorted natural scene plane image; then, dividing the luminance component of each undistorted natural scene plane image into image blocks which are not overlapping mutually and have a size of 8×8; next, randomly selecting N image blocks from all the image blocks of the luminance components of all the undistorted natural scene plane images; adopting each selected image block as a training sample; and denoting an ith training sample as xi; wherein: 5000≦N≦20000, and 1≦i≦N; subsequently, forming a gray vector through arranging pixel values of all pixels in each training sample; and denoting a gray vector formed through arranging pixel values of all pixels in the xi as xicol, wherein: the xicol has a dimensionality of 64×1; and, a value of a 1st element to a 64th element in the xicol respectively correspond to the pixel value of each pixel in the xi in a line-by-line scanning manner; afterwards, for each gray vector corresponding to each training sample, subtracting a mean value of values of all elements in the gray vector from a value of each element in the gray vector, so as to centralize the gray vector corresponding to each training sample; and denoting a gray vector obtained after centralizing the xicol as {circumflex over (x)}icol; and finally, denoting a matrix formed by all obtained centralized gray vectors as X, X=[{circumflex over (x)}icol, {circumflex over (x)}2col, . . . , {circumflex over (x)}Ncol], wherein: the X has a dimensionality of 64×N; the {circumflex over (x)}1col, {circumflex over (x)}2col, . . . , {circumflex over (x)}Ncol respectively represent a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 1st training sample, a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 2nd training sample, . . . , and a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in an Nth training sample; and a symbol “[ ]” is a vector representation symbol; {circle around (2)} processing the X with dimensionality reduction and whitening through a principal component analysis, and denoting an obtained matrix after the dimensionality reduction and whitening as XW, wherein: the XW has a dimensionality of M×N; and the M is a set low dimensionality, 1<M<64; {circle around (3)} training N column vectors in the XW through an orthogonal locality preserving projection algorithm, and obtaining a best mapping matrix of eight orthogonal bases of the XW, denoted as JW, wherein the JW has a dimensionality of 8×M; and then, according to the JW and a whitened matrix, calculating a best mapping matrix of an original sample space, denoted as J, J=JW×W, wherein: the J has a dimensionality of 8×64; the W represents the whitened matrix; and the W has a dimensionality of M×64; {circle around (4)} representing an original undistorted natural scene stereoscopic image having a width of W′ and a height of H′ by Iorg; respectively denoting a left viewpoint image and a right viewpoint image of the Iorg as IorgL and IorgR, and extracting luminance components respectively from the IorgL and the IorgR; representing a distorted stereoscopic image of the Iorg after distortion by Idis; adopting the Idis as a distorted stereoscopic image to be assessed; respectively denoting a left viewpoint image and a right viewpoint image of the Idis as IdisL and IdisR, and extracting luminance components respectively from the IdisL and the IdisR; then, dividing the luminance components of the IorgL, the IorgR, the IdisL, and theIdisRrespectivelyinto⌊W′8⌋×⌊H′8⌋image blocks which are not overlapping mutually and have a size of 8×8; denoting a jth image block in the luminance component of the IorgL as xjref,L; denoting a jth image block in the luminance component of the IorgR as xjref,R; denoting a jth image block in the luminance component of the IdisL as xjdis,L; and denoting a jth image block in the luminance component of the IdisR as xjdis,R; wherein a symbol “└ ┘” is a floor symbol; 1≦j≦N′; andN′=⌊W′8⌋×⌊H′8⌋; next, forming a gray vector through arranging pixel values of all pixels in each image block of the luminance component of the IorgL, and denoting a gray vector formed through arranging pixel values of all pixels in the xjref,L as xjref,L,col; forming a gray vector through arranging pixel values of all pixels in each image block of the luminance component of the IorgR, and denoting a gray vector formed through arranging pixel values of all pixels in the xjref,R as xjref,R,col; forming a gray vector through arranging pixel values of all pixels in each image block of the luminance component of the IdisL, and denoting a gray vector formed through arranging pixel values of all pixels in the xjdis,L as xjdis,L,col; forming a gray vector through arranging pixel values of all pixels in each image block of the luminance component of the IdisR, and denoting a gray vector formed through arranging pixel values of all pixels in the xjdis,R as xjdis,R,col; wherein: the xjref,L,col, the xjref,R,col, the xjdis,L,col, and the xjdis,R,col all have a dimensionality of 64×1; a value of a 1st element to a 64th element in the xjref,L,col respectively correspond to the pixel value of each pixel in the xjref,L in the line-by-line scanning manner; a value of a 1st element to a 64th element in the xjref,R,col respectively correspond to the pixel value of each pixel in the xjref,R in the line-by-line scanning manner; a value of 1st element to a 64th element in the xjdis,L,col respectively correspond to the pixel value of each pixel in the xjdis,L in the line-by-line scanning manner; and, a value of a 1st element to a 64th element in the xjdis,R,col respectively correspond to the pixel value of each pixel in the xjdis,R in the line-by-line scanning manner; afterwards, for each gray vector corresponding to each image block of the luminance component of the IorgL, subtracting a mean value of values of all elements in the gray vector from a value of each element in the gray vector, so as to centralize the gray vector corresponding to each image block of the luminance component of the IorgL, and denoting a gray vector obtained after centralizing the xjref,L,col as {circumflex over (x)}jref,L,col; for each gray vector corresponding to each image block of the luminance component of the IorgR, subtracting a mean value of values of all elements in the gray vector from a value of each element in the gray vector, so as to centralize the gray vector corresponding to each image block of the luminance component of the IorgR, and denoting a gray vector obtained after centralizing the xjref,R,col as {circumflex over (x)}jref,R,col; for each gray vector corresponding to each image block of the luminance component of the IdisL, subtracting a mean value of values of all elements in the gray vector from a value of each element in the gray vector, so as to centralize the gray vector corresponding to each image block of the luminance component of the IdisL, and denoting a gray vector obtained after centralizing the xjdis,L,col as {circumflex over (x)}jdis,L,col; for each gray vector corresponding to each image block of the luminance component of the IdisR, subtracting a mean value of values of all elements in the gray vector from a value of each element in the gray vector, so as to centralize the gray vector corresponding to each image block of the luminance component of the IdisR, and denoting a gray vector obtained after centralizing the xjdis,R,col as {circumflex over (x)}jdis,R,col; and finally, denoting a matrix formed by all obtained centralized gray vectors corresponding to the luminance component of the IorgL as Xref,L, Xref,L=[{circumflex over (x)}1ref,L,col, {circumflex over (x)}2ref,L,col, . . . , {circumflex over (x)}N′ref,L,col]; denoting a matrix formed by all obtained centralized gray vectors corresponding to the luminance component of the IorgR as Xref,R, Xref,R=[{circumflex over (x)}1ref,R,col, {circumflex over (x)}2ref,R,col, . . . , {circumflex over (x)}N′ref,R,col]; denoting a matrix formed by all obtained centralized gray vectors corresponding to the luminance component of the IdisL as Xdis,L, Xdis,L=[{circumflex over (x)}1dis,L,col, {circumflex over (x)}2dis,L,col, . . . , {circumflex over (x)}N′dis,L,col]; and, denoting a matrix formed by all obtained centralized gray vectors corresponding to the luminance component of the IdisR as Xdis,R, Xdis,R=[{circumflex over (x)}1dis,R,col, {circumflex over (x)}2dis,R,col, . . . , {circumflex over (x)}N′dis,R,col]; wherein the Xref,L, the Xref,R, the Xdis,L, and the Xdis,R all have a dimensionality of 64×N′; the {circumflex over (x)}1ref,L,col, {circumflex over (x)}2ref,L,col, . . . , {circumflex over (x)}N′ref,L,col respectively represent a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 1st image block of the luminance component of the IorgL, a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 2nd image block of the luminance component of the IorgL, . . . , and a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in an N′th image block of the luminance component of the IorgL; the {circumflex over (x)}1ref,R,col, {circumflex over (x)}2ref,R,col, . . . , {circumflex over (x)}N′ref,R,col respectively represent a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 1st image block of the luminance component of the IorgR, a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 2nd image block of the luminance component of the IorgR, . . . , and a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in an N′th image block of the luminance component of the IorgR; the {circumflex over (x)}1dis,L,col, {circumflex over (x)}2dis,L,col, . . . , {circumflex over (x)}N′dis,L,col respectively represent a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 1st image block of the luminance component of the IdisL, a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 2nd image block of the luminance component of the IdisL, . . . , and a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in an N′th image block of the luminance component of the IdisL; the {circumflex over (x)}1dis,R,col, {circumflex over (x)}2dis,R,col, . . . , {circumflex over (x)}N′dis,R,col respectively represent a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 1st image block of the luminance component of the IdisR, . . . , a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in a 2nd image block of the luminance component of the IdisR, . . . , and a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in an N′th image block of the luminance component of the IdisR; and the symbol “[ ]” is the vector representation symbol; {circle around (5)} calculating a structural difference between each column vector in the Xref,L and a corresponding column vector in the Xdis,L, and denoting a structural difference between the {circumflex over (x)}jref,L,col and the {circumflex over (x)}jdis,L,col as AVE({circumflex over (x)}jref,L,col, {circumflex over (x)}jdis,L,col); calculating a structural difference between each column vector in the Xref,R and a corresponding column vector in the Xdis,R, and denoting a structural difference between the {circumflex over (x)}jref,R,col and the {circumflex over (x)}jdis,R,col as AVE({circumflex over (x)}jref,R,col, {circumflex over (x)}jdis,R,col); then, forming a vector having a dimensionality of 1×N′ through orderly arranging N′ structural differences corresponding to the Xref,L and the Xdis,L, denoted as vL; and, forming a vector having a dimensionality of 1×N′ through orderly arranging N′ structural differences corresponding to the Xref,R and the Xdis,R, denoted as vR; wherein: a value of a jth element in the vL is vjL, vjL=AVE({circumflex over (x)}jref,L,col, {circumflex over (x)}jdis,L,col); and a value of a jth element in the vR is vjR, vjR=AVE({circumflex over (x)}jref,R,col, {circumflex over (x)}jdis,R,col); and obtaining an undistorted left viewpoint image block set, a distorted left viewpoint image block set, an undistorted right viewpoint image block set, and a distorted right viewpoint image block set, comprising steps of: a1), setting a left viewpoint image block selection threshold TH1 and a right viewpoint image block selection threshold TH2; a2), extracting all elements having a value larger than or equal to the TH1 from the vL, and extracting all elements having a value larger than or equal to the TH2 from the vR; and a3), adopting a set formed by the image blocks of the luminance component of the IorgL corresponding to the elements extracted from the vL as the undistorted left viewpoint image block set, denoted as Yref,L, Yref,L={xjref,L|AVE({circumflex over (x)}jref,L,col, {circumflex over (x)}jdis,L,col)≧TH1, 1≦j≦N′}; adopting a set formed by the image blocks of the luminance component of the IdisL corresponding to the elements extracted from the vL as the distorted left viewpoint image block set, denoted as Ydis,L, Ydis,L={xjdis,L|AVE({circumflex over (x)}jref,L,col, {circumflex over (x)}jdis,L,col)≧TH1, 1≦j≦N′}; adopting a set formed by the image blocks of the luminance component of the IorgR corresponding to the elements extracted from the vR as the undistorted right viewpoint image block set, denoted as Yref,R, Yref,R={xjref,R|AVE({circumflex over (x)}jref,R,col, {circumflex over (x)}jdis,R,col)≧TH2, 1≦j≦N′}; and, adopting a set formed by the image blocks of the luminance component of the IdisR corresponding to the elements extracted from the VR as the distorted right viewpoint image block set, denoted as Ydis,R, Ydis,R={xjdis,R|AVE({circumflex over (x)}jref,R,col, {circumflex over (x)}jdis,R,col)≧TH2, 1≦j≦N′}; {circle around (6)} calculating a manifold characteristic vector of each image block in the Yref,L, and denoting a manifold characteristic vector of a tth image block in the Yref,L as rtref,L, rtref,L=J×{circumflex over (x)}tref,L,col; calculating a manifold characteristic vector of each image block in the Ydis,L, and denoting a manifold characteristic vector of a tth image block in the Ydis,L as dtdis,L, dtdis,L=J×{circumflex over (x)}tdis,L,col; calculating a manifold characteristic vector of each image block in the Yref,R, and denoting a manifold characteristic vector of a t′th image block in the Yref,R as rt′ref,R, rt′ref,R=J×{circumflex over (x)}t′ref,R,col; calculating a manifold characteristic vector of each image block in the Ydis,R, and denoting a manifold characteristic vector of a t′th image block in the Ydis,R as dt′dis,R, dt′dis,R=J×{circumflex over (x)}t′dis,R,col; wherein: 1≦t≦K, the K represents a total number of the image blocks in the Yref,L, namely a total number of the image blocks in the Ydis,L; 1≦t′≦K′, the K′ represents a total number of the image blocks in the Yref,R, namely a total number of the image blocks in the Ydis,R; the rtref,L, the dtdis,L, the rt′ref,R, and the dt′dis,R all have a dimensionality of 8×1; the {circumflex over (x)}tref,L,col represents a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in the tth image block of the Yref,L; the {circumflex over (x)}tdis,L,col represents a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in the tth image block of the Ydis,L; the {circumflex over (x)}t′ref,R,col represents a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in the t′th image block of the Yref,R; and, the {circumflex over (x)}t′dis,R,col represents a gray vector obtained after centralizing a gray vector formed through arranging pixel values of all pixels in the t′th image block of the Ydis,R; then, forming a matrix by the manifold characteristic vectors of all the image blocks in the Yref,L, denoted as RL; forming a matrix by the manifold characteristic vectors of all the image blocks in the Ydis,L, denoted as DL; forming a matrix by the manifold characteristic vectors of all the image blocks in the Yref,R, denoted as RR; and, forming a matrix by the manifold characteristic vectors of all the image blocks in the Ydis,R, denoted as DR; wherein: the RL and the DL both have a dimensionality of 8×K; the RR and the DR both have a dimensionality of 8×K′; a tth column vector in the RL is the rtref,L; a tth column vector in the DL is the dtdis,L; a t′th column vector in the RR is the rtref,R; and, a t′th column vector in the DR is the dt′dis,R; and calculating a manifold characteristic similarity between the luminance component of the IorgL and the luminance component of the IdisL, denoted as MFS1L,MFS1L=18×K∑m=18∑t=1K2Rm,tLDm,tL+C1(Rm,tL)2+(Dm,tL)2+C1;and, calculating a manifold characteristic similarity between the luminance component of the IorgR and the luminance component of the IdisR, denoted as MFS1R,MFS1R=18×K′∑m=18∑t′=1K′2Rm,t′RDm,t′R+C1(Rm,t′R)2+(Dm,t′R)2+C1;wherein: the Rm,tL represents a value of an element in a mth row and a tth column of the RL; the Dm,tL represents a value of an element in a mth row and a tth column of the DL; the Rm,t′R represents a value of an element in a mth row and a tth column of the RR; the Dm,t′R represents a value of an element in a mth row and a tth column of the DR; and, the C1 is a small constant for guaranteeing a result stability;
{circle around (7)} calculating a luminance similarity between the luminance component of the IorgL and the luminance component of the IdisL, denoted as MFS2L,MFS2L=∑t=1K(μtref,L-μ_ref,L)×(μtdis,L-μ_dis,L)+C2∑t=1K(μtref,L-μ_ref,L)2×∑t=1K(μtdis,L-μ_dis,L)2+C2;and, calculating a luminance similarity between the luminance component of the IorgR, and the luminance component of the IdisR, denoted as MFS2R,MFS2R=∑t′=1K′(μt′ref,R-μ_ref,R)×(μt′dis,R-μ_dis,R)+C2∑t′=1K′(μt′ref,R-μ_ref,R)2×∑t′=1K′(μt′dis,R-μ_dis,R)2+C2;wherein: the μ1ref,L represents a mean value of the pixel values of all the pixels in the tth image block of the Yref,L,μ_ref,L=∑t=1Kμtref,LK;the μtdis,L represents a mean value of the pixel values of all the pixels in the tth image block of the Ydis,L,μ_dis,L=∑t=1Kμtdis,LK;the μt′ref,R represents a mean value of the pixel values of all the pixels in the t′th image block of the Yref,R,μ_ref,R=∑t′=1K′μt′ref,RK′;the μt′dis,R represents a mean value of the pixel values of all the pixels in the t′th image block of the Ydis,R,μ_dis,R=∑t′=1K′μt′dis,RK′;and, the C2 is a small constant for guaranteeing a result stability;
{circle around (8)} processing the MFS1L and the MFS2L with power exponent weighting, so as to obtain a quality score of the IdisL, denoted as MFSL, MFSL=(MFS1L)α×(MFS2L)β; and, processing the MFS1R and the MFS2R with the power exponent weighting, so as to obtain a quality score of the IdisR, denoted as MFSR, MFSR=(MFS1R)α×(MFS2R)β; wherein: the α is for adjusting a relative importance of the MFS1L and the MFS1R; the β is for adjusting a relative importance of the MFS2L and the MFS2R; and, α+β=1; and {circle around (9)} obtaining a weighted value of the IdisL and a weighted value of the IdisR through a binocular rivalry model, respectively denoted as ωL and ωR; weighting the MFSL through the ωL, and weighting the MFSR through the ωR, so as to obtain a quality value of the Idis, denoted as Q, Q=ωL×MFSL+ωR×MFSR; wherein: the ωL is for adjusting a relative importance of the MFSL; the ωR is for adjusting a relative importance of the MFSR; and, ωL+ω2=1. |
