Method for evaluating quality of tone-mapping image based on exposure analysis

A method for evaluating quality of a tone-mapping image based on exposure analysis, comprising steps of:

• (1) denoting SHDR as a high dynamic range image which is unprocessed with a width W and a height H, SHDR is an input signal of a conventional low dynamic range display devices;
• denoting STM as a tone-mapping image generated from SHDR after processing with a tone-mapping operator, wherein STM serves as a tone-mapping image to be evaluated;
• (2) performing pre-exposure processing on a luminance component of SHDR under different exposure degrees, so as to generate an overexposed image and an underexposed image of the luminance component of SHDR, which is respectively denoted as EIover and EIunder;
• (3) dividing EIover into
• ⌊
• W
• 2
• u
• ⌋
• ×
• ⌊
• H
• 2
• v
• ⌋
• non-overlapped image blocks with a size of 2u×2v;
• then finding out overexposed image blocks from all image blocks of EIover;
• forming all of the overexposed image blocks in EIover into an overexposed area which is denoted as RoverExposure;
• wherein └ ┘ is a floor operation symbol, u and v are identical integer selected from an interval of [2, 5];
• dividing EIunder into
• ⌊
• W
• 2
• u
• ⌋
• ×
• ⌊
• H
• 2
• v
• ⌋
• non-overlapped image blocks with a size of 2u×2v;
• then finding out underexposed image blocks from all image blocks of EIunder;
• forming all of the underexposed image blocks in EIunder into an underexposed image area which is denoted as RunderExposure;
• (4) dividing SHDR into
• ⌊
• W
• 2
• u
• ⌋
• ×
• ⌊
• H
• 2
• v
• ⌋
• non-overlapped image blocks with a size of 2u×2v;
• then, according to RoverExposure and RunderExposure, dividing SHDR into an easy overexposed area, an easy underexposed area and an easy normal-exposed area, which are respectively denoted as Rover, Runder and Rnormal;
• (5) dividing STM into
• ⌊
• W
• 2
• u
• ⌋
• ×
• ⌊
• H
• 2
• v
• ⌋
• non-overlapped image blocks with a size of 2u×2v;
• then denoting an area corresponding to Rover in STM as a tone-mapping easy overexposed area R1;
• then denoting an area corresponding to Runder in STM as a tone-mapping easy underexposed area R2;
• denoting an area corresponding to Rnormal in STM as a tone-mapping easy normal-exposed area R3;
• (6) judging if there is overexposure in each of the image blocks in R1, if yes, determining image blocks which are overexposed as tone-mapping overexposed blocks, then counting a total number of the tone-mapping overexposed blocks which is denoted as NoverTM;
• judging if there is underexposure in each of the image blocks in R2, if yes, determining image blocks which are underexposed as tone-mapping underexposed blocks, then counting a total number of the tone-mapping underexposed blocks which is denoted as NunderTM;
• (7) according to NoverTM and NunderTM, calculating an abnormal exposure rate of STM which is denoted as ηabnormal,
• η
• abnormal
• =
• N
• over
• TM
• +
• N
• under
• TM
• N
• R
• 1
• +
• N
• R
• 2
• ;
• wherein NR1 represents a total number of image blocks in R1;
• NR2 represents a total number of image blocks in R2;
• (8) calculating an overexposed residual energy of R1, which is denoted as E1, wherein
• E
• 1
• =
• ∑
• n
• =
• 1
• N
• R
• 1
• 
• (
• μ
• 1
• ,
• n
• -
• L
• over
• E
• )
• 2
• N
• R
• 1
• ;
• calculating an underexposed residual energy of R2, which is denoted as E2, wherein
• E
• 2
• =
• ∑
• n
• ′
• =
• 1
• N
• R
• 2
• 
• (
• μ
• 2
• ,
• n
• ′
• -
• L
• under
• E
• )
• 2
• N
• R
• 2
• ;
• wherein μ1,n represents an average value of pixel values of all pixels in a corresponding area of an nth image blocks in R1 in an luminance component of STM;
• μ2,n′ represents an average value of pixel values of all pixels in a corresponding area of an n'th image blocks in R2 in the luminance component of STM;
• LoverE is an extreme overexposure brightness value, LoverE=255;
• LunderE is an extreme underexposure brightness value, LunderE=0;
• (9) converting R3 from an RGB (red green blue) color space into an opponent color space, which is denoted as R3′;
• then calculating an average value and a variance of a component value of a red-green channel of all pixels in R3′;
• which are respectively denoted as μrg and σrg;
• calculating an average value and a variance of a component value of yellow-blue channel of all pixels in R3′, which are respectively denoted as μyb and σyb;
• then calculating an exposure color index of R3′, which is denoted as C3, C3=√{square root over (σrg2+σyb2)}+ωc×√{square root over (μrg2+μyb2)};
• wherein ωc represents weighing of an average color value;
• (10) obtaining a characteristic vector of STM, which is denoted as FTM, FTM=[ηabnormal, E1,E2,C3], wherein symbol [ ] is a vector symbol;
• (11) testing FTM according to a support vector regression training model to obtain an objective quality evaluation predictive value of STM, which is denoted as Q=f(Xdis), f(Xdis)=(Vbest)Tφ(Xdis)+bbest;
• wherein Q is a function of Xdis, f( ) is an expression form of function, Xdis is inputting for representing FTM, Vbest and bbest are an optimal weight vector and an optimal bias of the support vector regression training model, (Vbest)T is a transposition of Vbest, φ(Xdis) is a linear function of Xdis.