阅读说明：本技术 一种无参考非对称失真立体图像质量评价方法 (It is a kind of without refer to asymmetric distortion stereo image quality evaluation method ) 是由 陈勇 朱凯欣 刘焕淋 于 2019-07-12 设计创作，主要内容包括：本发明涉及一种无参考非对称失真立体图像质量评价方法,包括以下步骤：首先,根据双目视觉特性判别不同的双目组合行为,并构建相应的独眼图,而非使用固定的独眼图模拟不同的双目组合行为。然后,在左右视图及独眼图上分别提取单目和双目视觉特征。另外,为准确度量立体图像的深度信息,本发明在加权视差图和左右视图相关性图上,提取能够反映水平深度和纵向深度信息的质量感知特征。最后,通过训练极限学习机(Extreme Learning Machine,ELM)算法,构建从立体图像质量感知特征域到质量分数域的映射关系模型。(The present invention relates to a kind of without with reference to asymmetric distortion stereo image quality evaluation method, the following steps are included: first, differentiate that different binoculars combines behavior according to binocular vision characteristic, and construct corresponding one-eyed figure, rather than intends different binoculars using fixed one-eyed artwork and combine behavior.Then, monocular and binocular vision feature are extracted respectively on left and right view and one-eyed figure.In addition, being the depth information of accuracy amount stereo-picture, the present invention extracts the quality Perception Features for being able to reflect horizontal depth and longitudinal depth information on weighting disparity map and left and right drawings relativity figure.Finally, constructing the mapping relations model from stereo image quality Perception Features domain to mass fraction domain by training extreme learning machine (Extreme Learning Machine, ELM) algorithm.)

1. a kind of without with reference to asymmetric distortion stereo image quality evaluation method, it is characterised in that: fully consider binocular vision spy Property and depth quality metric, complete asymmetric stereo image quality measurement.The following steps are included:

Step 1: according to the size of structural similarity and useful information amount between the view of left and right, to the binocular of stereo pairs Combination behavior judged, and construct corresponding one-eyed figure come simulate human visual system (Human Visual System, HVS the binocular in) combines behavior；

Step 2: the Statistical Distribution Characteristics and texture that can characterize image fault are extracted on left and right view and above-mentioned one-eyed figure Feature；

Step 3: extracting horizontal depth Perception Features on weighting disparity map, longitudinal depth is extracted on the view related coefficient figure of left and right Perception Features；

Step 4: limit of utilization learning machine (Extreme Learning Machine, ELM) is trained, building from property field to The mapping relations in mass fraction domain.

2. according to claim 1 a kind of without with reference to asymmetric distortion stereo image quality evaluation method, it is characterised in that: The step 1 is realized especially by following manner: the structural similarity value of left and right view is calculated using structural similarity model S, and be arranged by threshold value, combination behavior of the stereo pairs in HVS is divided into binocular fusion and binocular competition (binocular inhibition It is a kind of special circumstances of binocular competition).Then, the useful information amount for calculating left and right view is judged to binocular if difference is smaller Competition；If differing greatly, it is judged to binocular inhibition, and chooses the view to contain much information and dominates entire fusion process, because there is master It sees experiments have shown that the view to contain much information is easier to attract the attention of HVS.Finally, being constructed according to different binocular behaviors corresponding One-eyed figure combine behavior to simulate the binocular of HVS, rather than fixed one-eyed artwork is used to intend three kinds of different binocular behaviors.

3. according to claim 1 a kind of without with reference to asymmetric distortion stereo image quality evaluation method, it is characterised in that: The step 3 is realized especially by following manner: binocular parallax causes left and right view to occur mismatching information, this mismatch Information is the main reason for bringing depth information for stereo-picture, and still, the mismatch information in not all region is to human eye Attraction is the same, and the present invention is parallax value d (i, j) distribution power at pixel (i, j) according to human eye vision attention mechanism Weight ω (i, j), to obtain shown in formula (1), for measuring the weighting disparity map D of stereo-picture depth quality_ω。

The statistical nature of disparity map can be changed due to being distorted, the present invention chooses some simple statistical natures, such as: kurtosis k With degree of skewness s, as horizontal depth feature, such as formula (2-3).

D_ω(i, j)=ω (i, j) * d (i, j) (1)

K=E [(D_ω-E(D_ω))⁴]/E[(D_ω-E(D_ω))²] (2)

S=E [(D_ω-E(D_ω))³]/(E[(D_ω-E(D_ω))²])^3/2 (3)

4. according to claim 1 a kind of without with reference to asymmetric distortion stereo image quality evaluation method, it is characterised in that: The step 3 is realized especially by following manner: although binocular parallax is the main reason for generating depth information, it is anti- What is reflected is the horizontal position difference between the left and right retinal projection of set point in space, is led to have ignored by asymmetric distortion The left and right view longitudinal difference of cause influences depth information bring.Therefore, the present invention is left by column count using auto-correlation function View I_lWith right parallax compensation figure I_drAnd right view I_rWith left parallax compensation figure I_dlThe related coefficient figure of longitudinal information change, Then statistical nature is extracted, characterization left and right view longitudinally mismatches influence of the information to depth quality.

I_dl(i, j)=I_l((i,j)+d_l(i,j)) (4)

I_dr(i, j)=I_r((i,j)+d_r(i,j)) (5)

Wherein, d_l(x, y) is indicated using left view as main view, is looked for it from right view based on structural similarity quantity algorithm (SSIM) Maximum similitude, and calculate the difference between two o'clock, d_r(x, y) indicates to be based on SSIM algorithm from a left side using right view as main view Its maximum similitude is looked in view, and calculates the difference between two o'clock.

5. according to claim 1 a kind of without with reference to asymmetric distortion stereo image quality evaluation method, it is characterised in that: The step 4 is realized especially by following manner: limit of utilization learning machine (Extreme Learning Machine, ELM) It is trained, learns the mapping relations model from property field to mass fraction domain.The feature of ELM maximum is for traditional nerve Network, especially Single hidden layer feedforward neural networks (SLFNs), than traditional learning algorithm speed under the premise of guaranteeing to learn precision Degree is faster.

Technical field

The invention belongs to stereo image quality assessment technique fields, are related to a kind of without with reference to asymmetric distortion perspective view image quality Measure evaluation method.

Background technique

With the development of 3D technology, 3D film and TV are increasingly by global concern.3D rendering is in acquisition, compression, transmission During storage, left and right view will receive different degrees of and different types of distortion, to influence the viewing quality of human eye. Moreover, there are certain parallaxes between stereo-picture or so view compared with 2D image, depth perception can be provided for human eye, To bring more true visual experience, but excessive parallax can also cause uncomfortable experience to human eye, equally It will affect judgement of the human eye to picture quality.Therefore, to bring good visual experience, it is necessary to construct one and people The eye consistent stereo image quality evaluation model of subjective scoring.

Different from traditional 2D visual perception, the visual perception of 3D rendering is highly dependent on the quality of left and right view.According to view Feel that psychological study shows that the different views that right and left eyes obtain can be converted into single width view by human visual system, but It is not to be simply added left and right Eye irritation to can be obtained single width view during this conversion, especially asymmetric distortion Image.Due to the binocular vision mechanism of HVS system complex, current research still can not comprehensively simulate binocular vision characteristic.

Other than binocular vision characteristic, it is also necessary to consideration another influence stereo image quality an important factor for be exactly Depth information.Horizontal parallax is the principal element of depth information to be provided for human eye, but excessive parallax will also result in dizzy mesh Dizzy, discomfort influences judgement of the human eye to picture quality.The main problem that stereo image quality evaluation at present faces is exactly such as What simulates the binocular vision mechanism of complexity and the measurement to depth quality.

Summary of the invention

In view of this, the purpose of the present invention is to provide a kind of without with reference to asymmetric distortion stereo image quality evaluation side Method, in order to which more comprehensively the binocular vision characteristic of simulation HVS complexity, this method are based on binocular vision characteristic, construct different Binocular constitutional diagram simulates three kinds of binocular bind lines are as follows: binocular fusion, binocular competition, binocular inhibit.In addition, this method is by mentioning The horizontal depth and longitudinal depth perception feature for taking stereo-picture, solve the problems, such as stereo-picture depth quality metric.

In order to achieve the above objectives, the invention provides the following technical scheme:

It is a kind of without asymmetric distortion stereo image quality evaluation method is referred to, based on binocular vision characteristic and depth perception matter Metric realizes asymmetric distortion stereo image quality measurement.For asymmetric distorted image, left and right view is in HVS It is not simply to be added, single width view can be pooled.This method is primarily based on binocular vision characteristic, uses structural similarity mould Type calculates the structural similarity value S of left and right view, and is arranged by threshold value, by combination behavior of the stereo pairs in HVS point For binocular fusion and binocular competition (a kind of special circumstances that binocular inhibition is binocular competition).Then, having for left and right view is calculated Binocular competition is judged to if difference is smaller with information content；If differ greatly, it is judged to binocular inhibition, and choose and contain much information View dominates entire fusion process, because there is subjective experiment to prove that the view to contain much information is easier to attract the attention of human eye. Finally, constructing corresponding one-eyed figure according to different binocular behaviors combines behavior to simulate the binocular of HVS, rather than use fixed only Eye artwork intends three kinds of different binocular behaviors.

Further, the statistical Boot that can characterize image fault is extracted on left and right view and above-mentioned one-eyed figure It seeks peace textural characteristics, as monocular and binocular vision feature, for predicting stereo image quality.

It further, is that the parallax value d (i, j) at pixel (i, j) distributes weights omega according to human eye vision attention mechanism (i, j), to obtain the weighting disparity map D for measuring stereo-picture depth quality_ω.And one is extracted on the weighting disparity map A little simple statistical natures, such as: kurtosis k and degree of skewness s, as horizontal depth feature.

D_ω(i, j)=ω (i, j) * d (i, j) (1)

K=E [(D_ω-E(D_ω))⁴]/E[(D_ω-E(D_ω))²] (2)

S=E [(D_ω-E(D_ω))³]/(E[(D_ω-E(D_ω))²])^3/2 (3)

Further, column count left view I is pressed using auto-correlation function_lWith right parallax compensation figure I_drAnd right view I_r With left parallax compensation figure I_dlThe related coefficient figure of longitudinal information change, then extracts statistical nature, and characterization is led by asymmetric distortion The left and right view of cause longitudinally mismatches influence of the information to depth quality.

I_dl(i, j)=I_l((i,j)+d_l(i,j)) (4)

I_dr(i, j)=I_r((i,j)+d_r(i,j)) (5)

Wherein, d_l(x, y) indicates to look for its maximum similar from right view based on SSIM algorithm using left view as main view Point, and calculate the difference between two o'clock, d_r(x, y) indicates to look for from left view using right view as main view based on SSIM algorithm Its maximum similitude, and calculate the difference between two o'clock.

Finally, high and extreme learning machine (Extreme more faster than traditional learning algorithm speed using study precision Learning Machine, ELM) it is trained, learn the mapping relations model from property field to mass fraction domain.

The beneficial effects of the present invention are: it is provided by the invention a kind of without with reference to the evaluation of asymmetric distortion stereo image quality Method has fully considered binocular vision characteristic, differentiates to combination of the stereo pairs in HVS, and construct difference One-eyed figure combine behavior to simulate different binocular.In addition, extracting horizontal depth and longitudinal depth spy from stereo image pair Sign, to realize the depth quality metric of stereo-picture.

Detailed description of the invention

In order to keep the purpose of the present invention, technical scheme and beneficial effects clearer, the present invention provides following attached drawing and carries out Illustrate:

Stereo image quality evaluation model general flow chart of the Fig. 1 based on binocular vision characteristic and depth perception；

A kind of entire block diagram without with reference to asymmetric distortion stereo image quality evaluation method of Fig. 2.

Specific embodiment

Below in conjunction with attached drawing, embodiments of the present invention are described in detail.

It is provided by the invention it is a kind of be broadly divided into four parts without asymmetric distortion stereo-picture evaluation method is referred to, first Part is, differentiation stereo pairs combination among HVSs for statistical analysis to binocular vision characteristic, and is designed corresponding One-eyed figure；Second part is to extract monocular and binocular vision feature on left and right single width view and corresponding one-eyed figure；Third portion Divide is that horizontal and longitudinal depth characteristic is extracted；Part IV be construct it is pre- to the stereo image quality in mass fraction domain from property field Survey model.

It is provided by the invention it is a kind of without with reference to it is asymmetric distortion stereo-picture evaluation method general flow chart as shown in Figure 1, Specifically includes the following steps:

1. the judgement and simulation of binocular combination behavior

1) it is based on binocular vision characteristic, the structural similarity value S of left and right view is calculated using structural similarity model, and is led to Cross threshold value setting, by combination behavior of the stereo pairs in HVS be divided into binocular fusion and binocular competition (binocular inhibition be binocular A kind of special circumstances of competition).

2) the useful information amount for calculating left and right view is judged to binocular competition if difference is smaller；If differing greatly, sentence Inhibit for binocular, and choose the view to contain much information and dominate entire fusion process, because there is subjective experiment proof to contain much information View is easier to attract the attention of human eye.

3) corresponding one-eyed figure is constructed according to different binocular behaviors and combines behavior to simulate the binocular of HVS, rather than used Fixed one-eyed artwork intends three kinds of different binocular behaviors.

2. monocular and binocular feature extraction

The Statistical Distribution Characteristics that can characterize image fault are extracted on left and right view and above-mentioned one-eyed figure and texture is special Sign, as monocular and binocular vision feature, for predicting stereo image quality.

3. depth characteristic is extracted

It 1) is that the parallax value d (i, j) at pixel (i, j) distributes weights omega (i, j) according to human eye vision attention mechanism, Biggish region is paid close attention to for human eye, biggish weight is distributed, to obtain the weighting for measuring stereo-picture depth quality Disparity map D_ω.And some simple statistical natures are extracted on the weighting disparity map, and such as: kurtosis k and degree of skewness s, as level Depth characteristic.

D_ω(i, j)=ω (i, j) * d (i, j) (1)

K=E [(D_ω-E(D_ω))⁴]/E[(D_ω-E(D_ω))²] (2)

S=E [(D_ω-E(D_ω))³]/(E](D_ω-E(D_ω))²])^3/2 (3)

2) column count left view I is pressed using auto-correlation function_lWith right parallax compensation figure I_drAnd right view I_rIt is poor with left view Compensation

Scheme I_dlThen the related coefficient figure of longitudinal information change extracts statistical nature, characterize as caused by asymmetric distortion It is left

Right view longitudinally mismatches influence of the information to depth quality.

I_dl(i, j)=I_l((i,j)+d_l(i,j)) (4)

I_dr(i, j)=I_r((i,j)+d_r(i,j)) (5)

Wherein, d_l(x, y) indicates to look for its maximum similar from right view based on SSIM algorithm using left view as main view Point, and calculate the difference between two o'clock, d_r(x, y) indicates to look for from left view using right view as main view based on SSIM algorithm Its maximum similitude, and calculate the difference between two o'clock.

4. the building of stereo image quality evaluation model

Limit of utilization learning machine (Extreme Learning Machine, ELM) is trained, and is learnt from property field to matter Measure the mapping relations model of score field.The feature of ELM maximum is for traditional neural network, especially single hidden layer feed forward neural Network (SLFNs), under the premise of guaranteeing to learn precision faster than traditional learning algorithm speed.

Fig. 2 is a kind of entire block diagram without with reference to asymmetric distortion stereo-picture evaluation method proposed by the invention, main It to include training stage and test phase.

1) training stage

Training stage mainly includes the building of feature extraction and stereo image quality evaluation model.Feature is mainly single from left and right The binocular and monocular vision feature for being able to reflect picture material quality distortion, and weighting are extracted on width view and the one-eyed figure of synthesis The quality Perception Features for being able to reflect stereo-picture depth information, building instruction are extracted on disparity map and left and right view related coefficient figure Practice feature set, is then input to training in ELM, obtains final quality prediction model.

2) test phase

Above-mentioned identical feature is extracted on testing image, is then input in trained quality prediction model, thus Obtain the forecast quality score of testing image.