CN109871852A

CN109871852A - A kind of no reference tone mapping graph image quality evaluation method

Info

Publication number: CN109871852A
Application number: CN201910009929.3A
Authority: CN
Inventors: 侯春萍; 李�浩; 岳广辉
Original assignee: Tianjin University
Current assignee: Tianjin University
Priority date: 2019-01-05
Filing date: 2019-01-05
Publication date: 2019-06-11
Anticipated expiration: 2039-01-05
Also published as: CN109871852B

Abstract

The present invention relates to a kind of no reference tone mapping graph image quality evaluation methods, for image I to be evaluated, comprising the following steps: (1) global characteristics extract part: calculating color moment characteristics and be denoted as f₁；Image I to be evaluated is converted into grayscale image, all pixel ratios fallen in dark space and clear zone of an image is counted respectively, the light and shade distribution characteristics f of image is finally indicated with dark space ratio Φ and clear zone ratio Ω₂；Global entropy feature is denoted as f₃；(2) image I: being divided into the localized mass of 16x16 in airspace by local shape factor, and max pixel value I is calculated in block_max,BWith minimum pixel value I_min,BAnd poor ratio, be denoted as contrast f₄；The local entropy for calculating image block is denoted as f₅；Image I is divided into the block of 16x16 in wavelet field, with the coefficient of wavelet decomposition on approximate, horizontal, the vertical and diagonal four direction of C unified representation, and energy calculation is carried out on coefficient C, wavelet energy is denoted as f₆；(3) quality evaluation.

Description

A kind of no reference tone mapping graph image quality evaluation method

Technical field

The invention belongs to field of image processings, more particularly, to a kind of reference-free quality evaluation side of tone mapping image Method.

Background technique

In recent years, high dynamic range (High Dynamic Range, HDR) shows that equipment gradually increases, but low at present Dynamic range (Low Dynamic Range, LDR) shows that equipment still occupies large quantity.For adapt to LDR equipment demand, It needs the high dynamic range compression of real world to low-dynamic range, scholars propose various tone mapping operators to realize The conversion of HDR to LDR.But in conversion process also along with various distortion phenomenons, such as color imbalance and improper exposure Deng causing uncomfortable visual experience to viewer.For this purpose, the quality of evaluation tone mapping image become one it is urgently to be resolved Task.Existing tone mapping image quality evaluation algorithm performance is limited, and there is very big rooms for promotion.The present invention is directed to color Mapping image fault feature is adjusted, local feature and global characteristics is merged, proposes efficient non-reference picture quality appraisement method.

Summary of the invention

The present invention is directed to the distorted characteristic of tone mapping image, proposes a kind of merge locally with global characteristics without reference matter Evaluation method is measured, this method and human eye subjective assessment have very high consistency.Technical solution is as follows:

A kind of no reference tone mapping graph image quality evaluation method, for image I to be evaluated, comprising the following steps:

(1) global characteristics extract

It calculates color moment characteristics and is denoted as f₁；Image I to be evaluated is converted into grayscale image, defines pixel coverage in [085] [170255] image-region is dark space and clear zone, and counts all pixels fallen in dark space and clear zone of an image respectively Ratio finally indicates the light and shade distribution characteristics f of image with dark space ratio Φ and clear zone ratio Ω₂；Global entropy feature is denoted as f₃；Total is complete Office's character representation is f_G；

(2) local shape factor

Image I is divided into the localized mass of 16x16 in airspace, max pixel value I is calculated in block_max,BWith minimum pixel value I_min,BAnd poor ratio, be denoted as contrast f₄；The local entropy for calculating image block is denoted as f₅；Image I is divided into 16x16's in wavelet field Block with the coefficient of wavelet decomposition on approximate, horizontal, the vertical and diagonal four direction of C unified representation, and carries out energy on coefficient C Operation is measured, wavelet energy is denoted as f₆；Total local feature is denoted as f_L；

(3) quality evaluation

By global characteristics and Local Feature Fusion, it is denoted as f=[f_G f_L].Extract training set characteristics of image f, using support to Amount regression model training obtains quality prediction model, feature f is extracted equally on test set, being input in prediction model can be pre- Altimetric image quality.

No reference tone mapping graph image quality evaluation method provided by the invention, combines part and global characteristics, can The mass fraction of efficient forecast image.

Detailed description of the invention

Fig. 1 algorithm block diagram

Specific embodiment

The present invention proposes a kind of no reference tone mapping graph image quality evaluation method, and algorithm block diagram is as shown in Figure 1.

1, global characteristics extract

(1) color moment

An image I is given, in RGB color, decomposes obtain each Color Channel first, it is then logical in each color Road calculates separately the mean value of image, standard deviation, the degree of bias, and calculation formula is as follows:

Wherein, p_ijIndicate j-th of pixel of i-th of Color Channel, N indicates total pixel number in Color Channel, and formula is used respectively In averaging, color moment characteristics are denoted as f by standard deviation and the degree of bias₁。

(2) light and shade is distributed

Image I is converted into grayscale image, and dark space and clear zone are classified as according to pixel value.

1) dark space: pixel value is distributed in the image-region of [085]；

2) clear zone: pixel value is distributed in the image-region of [170255]

Defining clear zone (dark space) pixel number and accounting for the ratio of total pixel number is that clear zone ratio (dark space ratio) is bright for reacting image It is as follows to be denoted as Ω and Φ calculation formula respectively for (dark) degree:

Ω and Φ are calculated separately on grayscale image, and obtained light and shade distribution characteristics is denoted as f₂.

(3) global entropy

Quantify the information content of tone mapping image using comentropy:

In formula, H (D) indicates global entropy, P_l(D) probability density in first of gray level, for 8bit figure, l's are indicated Maximum value is 255, and global entropy feature is denoted as f₃。

2, local shape factor

(1) local contrast

There is different contrasts in different images region, and it is as follows to define block-based local contrast:

Wherein, setting block size is 16x16, and P, Q are respectively longitudinal block number and lateral block number after image segmentation, I_max,BWith I_min,BMax pixel value and minimum pixel value in image block are respectively indicated, C is set₁=C₂=1. in tri- Color Channels of R, G, B The value for inside calculating separately LC obtains local contrast feature, is denoted as f₄.

(2) local entropy

The probability of each gray value is ρ in image block_ij:

Wherein, M, N indicate the width and height of image block, and the present invention is disposed as the ash at 16, f (i, j) expression image (i, j) Degree, by probability ρ_ijSubstituting into formula can be obtained the local entropy H an of image block_LAll pieces of H is schemed to one_LCalculate average value And standard deviation, local entropy feature is obtained, f is denoted as₅。

(3) wavelet energy

In order to further indicate that the local message of image, dividing the image into size in wavelet field is MxN block, (M=N=16) With approximate, horizontal, in vertical and diagonal direction the coefficient of wavelet decomposition of C unified representation, the decomposition coefficient of image block is calculated, and Energy calculation is carried out on coefficient

Wavelet energy can indicate are as follows:

Wherein, E is the wavelet energy of an image block, contains approximate, horizontal, vertical and diagonal direction energy, right The wavelet energy of one all image block of figure takes mean value and standard deviation, and note this feature is f₆。

In summary, global characteristics are expressed as f_G, then f_G=[f₁,f₂,f₃].Local feature is expressed as f_L, then f_L= [f₄,f₅,f₆] passes through global characteristics and local shape factor, remember that total feature vector is f, then f=[f_G f_L]。

3, quality evaluation

After feature extraction and fusion, part and global characteristics are obtained, using support vector regression (Support Vector Regression, SVR) fused Feature Mapping is mass fraction by algorithm.Specifically, in conjunction with Fig. 1, by the master of training set Appearance quality score and the feature f of extraction, which are sent into SVR, to be trained, and a quality prediction model is obtained.Test set equally into Row feature extraction, in input prediction model, to predict the mass fraction of test set picture.It is effective for verification method Property, select Pearson's linearly dependent coefficient (Pearson Linear Correlation Coefficient, PLCC), this Pierre Graceful rank correlation coefficient (Spearman Rank-order Correlation Coefficient, SRCC), Ken Deer rank correlation system Number (Kendall ' s Rank Correlation Coefficient, KRCC) and root-mean-square error (Root Mean Squared Error, RMSE) it is used as standard.PLCC, SRCC and KRCC are bigger, and RMSE is smaller, illustrate that algorithm performance is better.

Select ESPL-LIVE HDR database that database is divided into training set and (is accounted for guarantee reasonability as platform 80%) 1000 intermediate values finally and test set (accounting for 20%), and random division 1000 times, are taken as a result, being shown in Table 1.

1 algorithm performance of table

As it can be seen from table 1 non-reference evaluation method of the invention has bigger PLCC, SRCC and KRCC value, RMSE is smaller, shows that algorithm and the quality of human eye subjective judgement image have very high consistency, can be used for evaluating tone mapping The quality of image.

Claims

1. a kind of no reference tone mapping graph image quality evaluation method, for image I to be evaluated, comprising the following steps:

(1) global characteristics extract

It calculates color moment characteristics and is denoted as f₁；Image I to be evaluated is converted into grayscale image, define pixel coverage in [085] and [170255] image-region is dark space and clear zone, and counts all pixel ratios fallen in dark space and clear zone of an image respectively Example, finally indicates the light and shade distribution characteristics f of image with dark space ratio Φ and clear zone ratio Ω₂；Global entropy feature is denoted as f₃；Total overall situation Character representation is f_G；

(2) local shape factor

Image I is divided into the localized mass of 16x16 in airspace, max pixel value I is calculated in block_max,BWith minimum pixel value I_min,B's With poor ratio, it is denoted as contrast f₄；The local entropy for calculating image block is denoted as f₅；Image I is divided into the block of 16x16 in wavelet field, uses C Coefficient of wavelet decomposition on approximate, horizontal, the vertical and diagonal four direction of unified representation, and energy calculation is carried out on coefficient C, Wavelet energy is denoted as f₆；Total local feature is denoted as f_L；

(3) quality evaluation

By global characteristics and Local Feature Fusion, it is denoted as f=[f_G f_L].Training set characteristics of image f is extracted, is returned using supporting vector Return model training to obtain quality prediction model, feature f is extracted equally on test set, being input in prediction model can prognostic chart Image quality amount.