CN104023230B

CN104023230B - A kind of non-reference picture quality appraisement method based on gradient relevance

Info

Publication number: CN104023230B
Application number: CN201410284237.7A
Authority: CN
Inventors: 刘利雄; 化毅; 赵清杰; 黄华
Original assignee: Beijing Institute of Technology BIT
Current assignee: Beijing Institute of Technology BIT
Priority date: 2014-06-23
Filing date: 2014-06-23
Publication date: 2016-04-13
Anticipated expiration: 2034-06-23
Also published as: CN104023230A

Abstract

The invention proposes a non-reference image quality evaluation method based on gradient correlation, which belongs to the field of computer image analysis. This method first obtains three sub-properties that are easily affected by distortion in the image gradient, which are the property of the magnitude of the image gradient, the property of the direction change of the image gradient and the property of the change of the magnitude of the image gradient. Divide these three properties into M×M image blocks, find the statistical variance of the above three properties in each image block, and use the average of the statistical variances of all image blocks in the entire image as its image features , and finally the image quality is obtained by combining the support vector machine with the two-step framework in image quality evaluation. The method of the invention has the advantages of small time complexity, high subjective consistency, low image feature dimension and good versatility, and can be applied to small computing devices or applications related to image quality, and has good practical value.

Description

A No-Reference Image Quality Assessment Method Based on Gradient Correlation

技术领域technical field

本发明涉及一种图像质量评价方法，尤其涉及一种基于梯度关联性的无参考图像质量评价方法，属于图像分析技术领域。The invention relates to an image quality evaluation method, in particular to a non-reference image quality evaluation method based on gradient correlation, which belongs to the technical field of image analysis.

背景技术Background technique

视觉乃是人类认知世界最基本最有效的途径之一，图像便是建立在人类视觉基础之上的。图像信息可以准确直接的帮助人们获得该信息所要表达的环境和意思，这是语音、文字等信息无法比拟的。所以，图像分析领域在计算机研究领域中有着十分重要的地位。图像记录了在某一时刻某一环境的状态，在这记录状态的过程中，会出现些许的误差，比如照相机拍照过程中会受到照相人手抖动的影响、X光片图像在成像过程中会受到X射线强弱以及胶片的影响等等。这些影响会直接干扰到成像的清晰度，进而使得图像当中包含的信息减少并引入失真。Vision is one of the most basic and effective ways for humans to perceive the world, and images are based on human vision. Image information can accurately and directly help people obtain the environment and meaning to be expressed by the information, which is incomparable to information such as voice and text. Therefore, the field of image analysis plays a very important role in the field of computer research. The image records the state of a certain environment at a certain moment. In the process of recording the state, there will be some errors. The intensity of X-rays and the influence of film, etc. These effects will directly interfere with the clarity of imaging, thereby reducing the information contained in the image and introducing distortion.

图像质量评价方法一般分为主观评价方法和客观评价方法。一般地，主观评价方法是将个人基于自身对一幅图像的视觉感受来评价该副图像的质量分值，并在多人对于同一幅图像的分值取平均值作为该副图像的最终分值。虽然这种评价方法是最准确的，但是其消耗的人力资源大，耗时时间长，同时也易于受到评价人员之间知识、观点等因素的影响。而客观评价方法则是应用计算机评价图像质量，排除了人员参与的因素，使得评价效率大大的增高。Image quality evaluation methods are generally divided into subjective evaluation methods and objective evaluation methods. In general, the subjective evaluation method is to evaluate the quality score of the sub-image based on the individual's visual perception of an image, and take the average value of the scores of multiple people for the same image as the final score of the sub-image . Although this evaluation method is the most accurate, it consumes a lot of human resources, takes a long time, and is also easily affected by factors such as knowledge and opinions among evaluators. The objective evaluation method is to use a computer to evaluate the image quality, which eliminates the factors of personnel participation and greatly improves the evaluation efficiency.

总体来说，客观质量评价方法即是应用计算机技术来替代主观评价方法之中人的因素，使得客观质量评价方法所评价的结果与主观评价方法的结果相近。最终完成计算机模拟人类感知图像信号的过程。客观质量评价方法在许多方面均有应用：1.作为图像融合、分割等算法的效果的评价算法；2.作为图像处理算法的前置算法，为之后的算法提供初始值等；3.衡量通信信道的好坏；4.嵌入到小型图像采集设备当中作应用等。Generally speaking, the objective quality evaluation method is to use computer technology to replace the human factor in the subjective evaluation method, so that the evaluation results of the objective quality evaluation method are similar to the results of the subjective evaluation method. Finally, the computer simulates the process of human perception of image signals. Objective quality evaluation methods are used in many aspects: 1. As an evaluation algorithm for the effect of algorithms such as image fusion and segmentation; 2. As a pre-algorithm for image processing algorithms, providing initial values for subsequent algorithms, etc.; 3. To measure communication The quality of the channel; 4. Embedded in a small image acquisition device for application, etc.

客观质量评价方法又可以分为三类：全参考图像质量评价方法、部分参考图像质量评价方法和无参考图像质量评价方法。顾名思义，全参考图像质量评价方法不仅需要失真图像，还需要与失真图像所对应的未失真图像。部分参考图像质量评价方法则是对比失真图像与失真图像所对应的未失真图像的部分信息，如提取的特征等等，来评价失真图像的质量。而无参考图像质量评价则是仅仅需要失真图像这一种信息来评价失真图像的质量。在实际应用当中，以无参考图像质量评价最为具有实用意义，因为在实际应用中，与失真图像所对应的原图像是很难获得的。Objective quality assessment methods can be divided into three categories: full reference image quality assessment methods, partial reference image quality assessment methods, and no-reference image quality assessment methods. As the name suggests, the full-reference image quality assessment method requires not only the distorted image, but also the undistorted image corresponding to the distorted image. The partial reference image quality evaluation method is to compare the distorted image with the part information of the undistorted image corresponding to the distorted image, such as extracted features, etc., to evaluate the quality of the distorted image. The non-reference image quality evaluation only needs the information of the distorted image to evaluate the quality of the distorted image. In practical applications, no-reference image quality evaluation has the most practical significance, because in practical applications, the original image corresponding to the distorted image is difficult to obtain.

综上所述，进行对于客观质量评价方法的研究具有广泛的理论意义和重要的应用价值。Moorthy等人在文献《Atwo-stepframeworkforconstructingblindimagequalityindices》中提出了无参考图像质量评价的两步框架，涉及的基础背景技术主要为图像梯度性质和soble算子性质。To sum up, the research on objective quality evaluation methods has extensive theoretical significance and important application value. Moorthy et al. proposed a two-step framework for non-reference image quality evaluation in the document "Atwo-stepframeworkforconstructingblindimagequalityindices". The basic background technologies involved are mainly image gradient properties and soble operator properties.

(一)无参考图像质量评价的两步框架(1) A two-step framework for image quality assessment without reference

Moorthy等人提出无参考图像质量评价的两步框架。该框架中，首先将输入的失真图像进行分类，之后在将该失真图像在每一类失真中预测的分数进行加权求和。Moorthy et al. propose a two-step framework for reference-free image quality assessment. In this framework, the input distorted image is first classified, and then the predicted scores of the distorted image in each type of distortion are weighted and summed.

当给出一个具有n中失真类型的图像训练集的时候，首先需要建立图像特征与失真分类之间的映射，在训练模型是向模型中输入正确的失真分类和图像特征，将失真分类模型训练完成，之后可通过该模型输入图像特征获得图像失真分类。When an image training set with n types of distortion is given, it is first necessary to establish a mapping between image features and distortion classifications. In the training model, the correct distortion classification and image features are input into the model, and the distortion classification model is trained After that, image distortion classification can be obtained by inputting image features through the model.

特别需要指出的是，在这个分类模型中需要建立一个硬分类器，需要的是一个能够说明该图像的失真在每种失真类型当中所占的概率，由此将获得一个n维的向量p，p中的每维数值便代表了输入图像的失真在每种确定的失真类型所占的概率。In particular, it needs to be pointed out that a hard classifier needs to be established in this classification model, and what is needed is a probability that can explain the distortion of the image in each distortion type, and thus an n-dimensional vector p will be obtained, The value of each dimension in p represents the probability of the distortion of the input image in each determined distortion type.

之后，训练针对每种失真类型的回归模型，即是建立图像特征与图像质量之间的映射。在训练时，图像训练集被分为n份，每一份只包含一种确定的失真类型的图像，因此，需要训练n个回归模型来预测当输入图像中存在的失真为某种特定类型时，在该种类型下的图像质量分数。这可以大大加强映射的准确性。Afterwards, a regression model for each distortion type is trained, i.e., a mapping between image features and image quality is established. During training, the image training set is divided into n parts, and each part only contains images of a certain type of distortion. Therefore, n regression models need to be trained to predict when the distortion in the input image is a certain type , the image quality score for this type. This can greatly enhance the accuracy of the mapping.

将带测试图像输入n个回归模型中，将得到n个质量分数，将这些质量分数按照模型分类向量p所对应的顺序变为n维的质量向量q。Inputting test images into n regression models will result in n quality scores, and these quality scores will be converted into n-dimensional quality vector q according to the order corresponding to the model classification vector p.

最后，将这些质量分数使用图像中失真分类向量进行加权求和，从而得到客观预测分数Finally, these quality scores are weighted and summed using the distortion classification vector in the image to obtain an objective prediction score

$Q Q = = {Σ Σ}_{i i = = 11}^{n no} {p p}_{i i} {q q}_{i i} - - - - - - ((11))$

其中，p_i表示向量p的第i维分量，q_i表示向量q的第i维分量，n表示失真的种类数目。Among them, p _i represents the i-th dimension component of vector p, q _i represents the i-th dimension component of vector q, and n represents the number of types of distortion.

(二)图像梯度信息(2) Image gradient information

图像的梯度信息包含了大量图像结构信息，图像的梯度一般表示这图像灰度值有剧烈变化的地方，而这些地方一般是图像边缘部分，同是也是人眼视觉系统的敏感区域。The gradient information of the image contains a large amount of image structure information. The gradient of the image generally indicates the places where the gray value of the image changes drastically, and these places are generally the edge parts of the image, which are also sensitive areas of the human visual system.

对于离散的数字图像，其梯度幅值一般定义为：For discrete digital images, the gradient magnitude is generally defined as:

$Gradient Gradient ((i i,, j j)) = = \sqrt{{((Gradient Gradient__x x ((i i,, j j))))}^{22} + + {((Gradient Gradient__y the y ((i i,, j j))))}^{22}}$

其中Gradient_x(i,j)，Gradient_y(i,j)分别为使用各种近似的离散算子计算的在位置i，j点X和Y两个正交方向的偏导数，例如算子Sobel，Prewitt和Canny算子。Among them, Gradient_x(i,j), Gradient_y(i,j) are the partial derivatives in the two orthogonal directions of X and Y at position i and point j calculated by using various approximate discrete operators, such as operators Sobel, Prewitt and Canny operator.

梯度的方向是梯度幅值变化最快的地方，图像梯度的方向被定义为：The direction of the gradient is where the magnitude of the gradient changes the fastest, and the direction of the image gradient is defined as:

$orientation orientation ((i i,, j j)) = = arc arc tan the tan ((\frac{Gradient Gradient__y the y ((i i,, j j))}{Gradient Gradient__x x ((i i,, j j))}))$

总之，当图像中存在边缘时，一定有较大的梯度值；而图像中较平滑的部分，灰度值变化较小，一般有较小的梯度。图像处理中常把梯度的模简称为梯度，由图像梯度构成的图像称为梯度图像。In short, when there is an edge in the image, there must be a larger gradient value; while in the smoother part of the image, the gray value changes less, and generally has a smaller gradient. In image processing, the modulus of the gradient is often referred to simply as the gradient, and the image composed of image gradients is called the gradient image.

发明内容Contents of the invention

本发明的目的是为了解决当前无参考图像质量评价方法中存在的时间、空间复杂度高、性能低下等问题，通过建立一个拥有高性能、低复杂度、与主观评价结果相一性高的无参考自然图像质量方法。The purpose of the present invention is to solve the problems of high time and space complexity and low performance in the current no-reference image quality evaluation method, by establishing a no-reference image quality evaluation method with high performance, low complexity, and high consistency with subjective evaluation results. See Natural Image Quality Methods.

本发明方法是通过下述技术方案实现的。The method of the present invention is realized through the following technical solutions.

一种基于梯度关联性的无参考图像质量评价方法，包括以下步骤：A non-reference image quality evaluation method based on gradient correlation, comprising the following steps:

步骤一、对输入的失真图像进行特征提取。Step 1. Feature extraction is performed on the input distorted image.

首先，对每一幅图像求取其在图像梯度方面的三种不同的子性质，即梯度幅值性质GM、梯度方向变化性质CO和梯度幅值变化性质CM。其中，三种梯度性质分别由式1、2、3定义：Firstly, three different sub-properties in terms of image gradient are obtained for each image, namely the gradient magnitude property GM, the gradient direction change property CO and the gradient magnitude change property CM. Among them, the three gradient properties are defined by formulas 1, 2 and 3 respectively:

$GM GM ((i i,, j j)) = = \sqrt{{((Gx Gx ((i i,, j j))))}^{22} + + {((Gy Gy ((i i,, j j))))}^{22}} - - - - - - ((11))$

CO(i,j)＝orientation(i,j)-orientation_avg(i,j)(2)CO(i,j)=orientation(i,j)-orientation _avg (i,j)(2)

$CM CM ((i i,, j j)) = = \sqrt{{((Gx Gx ((i i,, j j)) - - {Gx Gx}_{avg avg} ((i i,, j j))))}^{22} + + {((Gy Gy ((i i,, j j)) - - {Gy Gy}_{avg avg} ((i i,, j j))))}^{22}} - - - - - - ((33))$

其中，in,

$orientation orientation ((i i,, j j)) = = arc arc tan the tan ((\frac{Gy Gy ((i i,, j j))}{Gx Gx ((i i,, j j))}))$

${orientation orientation}_{avg avg} ((i i,, j j)) = = arc arc tan the tan ((\frac{{Gy Gy}_{avg avg} ((i i,, j j))}{{Gx Gx}_{avg avg} ((i i,, j j))}))$

${Gx Gx}_{avg avg} = = \frac{{Σ Σ}_{i i}^{M m} {Σ Σ}_{j j}^{N N} Gx Gx ((i i,, j j))}{M m \times \times N N}$

${Gy Gy}_{avg avg} = = \frac{{Σ Σ}_{i i}^{M m} {Σ Σ}_{j j}^{N N} Gy Gy ((i i,, j j))}{M m \times \times N N}$

其中，orientation(i,j)表示图像梯度的方向，Gx和Gy为离散的数字图像在X、Y两个正交方向的导数，M、N是为了描述区域内的变化所设定的窗口的大小。并且在求取梯度Gx和Gy中，使用sobel梯度计算算子，并且将其由一组正交方向扩展到两组。其中包含0度和90度的一组正交方向，以及一组-45度和45度的正交方向。Among them, orientation(i, j) represents the direction of the image gradient, Gx and Gy are the derivatives of the discrete digital image in the two orthogonal directions of X and Y, and M and N are the windows set to describe the changes in the region size. And in calculating the gradients Gx and Gy, use the sobel gradient calculation operator, and extend it from one set of orthogonal directions to two sets. It contains a set of orthogonal directions of 0 degrees and 90 degrees, and a set of orthogonal directions of -45 degrees and 45 degrees.

正交方向0度和90度的sobel算子为：The sobel operators of 0 degrees and 90 degrees in the orthogonal direction are:

$S^{y} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & - 1 & 0 \\ 0 & 2 & 0 & - 2 & 0 \\ 0 & 1 & 0 & - 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]$ 与 $S^{x} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 2 & 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \\ 0 & - 1 & - 2 & - 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]$ $S^{the y} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & - 1 & 0 \\ 0 & 2 & 0 & - 2 & 0 \\ 0 & 1 & 0 & - 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]$ and $S^{x} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 2 & 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \\ 0 & - 1 & - 2 & - 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]$

正交方向-45度和45度的sobel算子为：The sobel operators of -45 degrees and 45 degrees in the orthogonal direction are:

$S^{y} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 1 & 0 & - 1 & 0 & 0 \\ 0 & 2 & 0 & 2 & 0 \\ 0 & 0 & 1 & 0 & - 1 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]$ 与 $S^{x} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 & - 1 \\ 0 & 2 & 0 & - 2 & 0 \\ 1 & 0 & - 1 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]$ $S^{the y} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 1 & 0 & - 1 & 0 & 0 \\ 0 & 2 & 0 & 2 & 0 \\ 0 & 0 & 1 & 0 & - 1 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]$ and $S^{x} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 & - 1 \\ 0 & 2 & 0 & - 2 & 0 \\ 1 & 0 & - 1 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]$

其中上标为y的算子为计算Gy的算子，其中上标为x的算子为计算Gx的算子。The operator with the superscript y is the operator for calculating Gy, and the operator with the superscript x is the operator for calculating Gx.

然后，将上述所求得的6幅图像梯度性质图像进行分块处理，在每一块梯度性质上求取统计方差，在此基础上将该方差以log函数规范后，以计算平均数的形式将每一块梯度所求取的统计方差融合为整体图像梯度性质的方差。具体如下：Then, divide the 6 image gradient properties images obtained above into blocks, and calculate the statistical variance on each block of gradient properties. On this basis, after standardizing the variance with the log function, calculate the average The statistical variance calculated by each block gradient is fused into the variance of the overall image gradient property. details as follows:

将每一种图像梯度性质分割为128×128大小的图像梯度性质图像块，并将每一个图像梯度性质图像块求取其统计方差，即第n块图像块的方差d_n。计算公式如式4：Divide each image gradient property into 128×128 image gradient property image blocks, and calculate the statistical variance of each image gradient property image block, that is, the variance d _n of the nth image block. The calculation formula is as formula 4:

d_n＝∑(h(x)-E(h(x)))²(4)d _n =∑(h(x)-E(h(x))) ² (4)

其中，in,

h(x)＝pdf(θ)h(x)=pdf(θ)

其中，θ为每种性质图像中的参数，pdf为该参数的统计分布，h(x)表示对θ进行量化统计后的统计概率表示，E(h(x))表示h(x)的期望。Among them, θ is the parameter in each property image, pdf is the statistical distribution of the parameter, h(x) represents the statistical probability representation after quantitative statistics of θ, E(h(x)) represents the expectation of h(x) .

接着将每个求出的方差d_n以log函数规范化，并使用平均聚集的方法融合图像梯度性质的图像块所求出的统计方差作为最终的特征f，如式5。Then, each obtained variance d _n is normalized by the log function, and the statistical variance obtained by fusing image blocks with image gradient properties is used as the final feature f by using the average aggregation method, as shown in Equation 5.

$f f = = \frac{11}{N N} \times \times log log (({Π Π}_{n no = = 11}^{N N} {d d}_{n no})),, n no = = 1,2 1,2 . . . . . . . . . . . . . . N N - - - - - - ((55))$

最后，将输入的失真图像进行降采样，变为第二尺度的图像，并重复上述过程，最终得到一组12维特征向量Feature。Finally, the input distorted image is down-sampled to a second-scale image, and the above process is repeated to finally obtain a set of 12-dimensional feature vectors Feature.

Feature＝[f_GM,f_CO,f_CM×2orientation×2scale]Feature＝[f _GM ,f _CO ,f _CM ×2orientation×2scale]

步骤二、进行特征映射。将经步骤一得到的12维特征向量作为最终的图像质量特征，据此建立图像特征和图像分数之间的映射关系。Step 2: Perform feature mapping. The 12-dimensional feature vector obtained in step 1 is used as the final image quality feature, and the mapping relationship between image features and image scores is established accordingly.

首先，将图像库分为训练集和测试集。所述测试集用来建立图像特征与图像质量之间的映射关系，测试集用来测试建立的映射关系的机能。使用支持向量机的方法，将测试集中图像的图像特征进行失真分类模型和对应的每个失真分类的质量评价模型的训练。First, the image library is divided into training and testing sets. The test set is used to establish a mapping relationship between image features and image quality, and the test set is used to test the function of the established mapping relationship. Using the support vector machine method, the image features of the images in the test set are trained for the distortion classification model and the corresponding quality evaluation model for each distortion classification.

然后，基于无参考图像质量评价中两步框架的流程，在测试集中进行分数的预测测试，即使用失真分类模型将受测图像的失真进行分类，再在失真分类当中使用质量评价模型来预测受测图像质量，从而得到受测图像质量分数，进而利用现有算法性能标准对其进行评估。Then, based on the process of the two-step framework in the no-reference image quality evaluation, the score prediction test is carried out in the test set, that is, the distortion classification model is used to classify the distortion of the tested image, and then the quality evaluation model is used in the distortion classification. The image quality is measured, so as to obtain the image quality score under test, and then use the existing algorithm performance standards to evaluate it.

有益效果Beneficial effect

本发明提出的基于图像梯度关联性的无参考图像质量评价方法，与现有的同类别的技术相比，有着主观一致性高、时间及空间复杂度小的特点，可以应用到小型系统当中，或嵌入到图像质量相关的算法和设备中，具有很高的应用价值。The non-reference image quality evaluation method based on image gradient correlation proposed by the present invention has the characteristics of high subjective consistency and small time and space complexity compared with the existing technologies of the same category, and can be applied to small systems. Or embedded in algorithms and devices related to image quality, it has high application value.

附图说明Description of drawings

图1是本发明所述方法的流程图。Figure 1 is a flow chart of the method of the present invention.

图2是本发明具体实施例1中本发明方法与另外几种的全参考、无参考算法进行主观一致性比较的盒形图。Fig. 2 is a box diagram of subjective consistency comparison between the method of the present invention and other several full-reference and no-reference algorithms in the specific embodiment 1 of the present invention.

具体实施方式detailed description

下面结合附图和具体实施例对本发明方法做进一步详细说明。The method of the present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

实施例1：Example 1:

如图1所示，一种基于梯度关联性的无参考图像质量评价方法，包括以下步骤：As shown in Figure 1, a no-reference image quality evaluation method based on gradient correlation includes the following steps:

步骤一、对输入图像进行特征提取。Step 1. Feature extraction is performed on the input image.

首先，对输入图像进行本发明所定义的梯度的三个性质的求取。通过对两个方向的图像梯度求取幅值、方向变化以及幅值变化信息。First, calculate the three properties of the gradient defined by the present invention on the input image. The amplitude, direction change and amplitude change information are obtained by calculating the image gradient in two directions.

然后，将两个方向的三个性质，总共六个梯度性质图像进行分块处理。之后对应每一个分块求取其相应的统计方差。Then, three properties in two directions, a total of six gradient property images are processed into blocks. Then corresponding to each block, its corresponding statistical variance is obtained.

之后，将所得的六种梯度性质图像的分块的统计方差进行合并，使用平均聚集的方式，求取每种性质中所有分块统计方差的平均值，作为最终的六种梯度性质的整体统计方差。Afterwards, the statistical variances of the blocks of the obtained six gradient properties images are combined, and the average aggregation method is used to calculate the average of the statistical variances of all blocks in each property, as the final overall statistics of the six gradient properties variance.

最后，对输入图像进行降采样，将其变为第二尺度的图像，重复上述过程最终将图像特征由6维扩充为12维。Finally, the input image is down-sampled to turn it into a second-scale image, and the above process is repeated to finally expand the image features from 6 dimensions to 12 dimensions.

首先使用支持向量机(SVM)的方法，将测试集中图像的图像特征进行失真分类模型和对应的每个失真分类的质量评价模型的训练。First, using the method of support vector machine (SVM), the image features of the images in the test set are trained for the distortion classification model and the corresponding quality evaluation model for each distortion classification.

然后基于无参考图像质量评价中两步框架的流程在测试集中进行分数的预测测试。即，先使用失真分类模型将受测图像的失真进行分类，再在失真分类当中使用质量评价模型来预测受测图像质量，从而得到受测图像质量分数，进而利用现有的算法性能指标(SROCC)对算法的优劣进行评估。The score prediction test is then performed on the test set based on the procedure of the two-step framework in no-reference image quality assessment. That is, first use the distortion classification model to classify the distortion of the tested image, and then use the quality evaluation model to predict the quality of the tested image in the distortion classification, so as to obtain the quality score of the tested image, and then use the existing algorithm performance index (SROCC ) to evaluate the pros and cons of the algorithm.

本实施例中，使用LIVE图像数据库来测试本发明的效率和性能。为了作出对照，本实施例使用了一些著名的全参考图像质量评价方法和无参考评价质量方法作为本发明的比照。在测试过程中，将数据库分为测试集和训练集，根据五折交叉验证方法，将测试集和训练集分别设置为20％和80％。之后套用本文所依据的两步预测框架来预测测试集的分数，而其中所需要的分类和回归模型由训练集中图像特征训练。最终计算预测分数和实际分数的SROCC指标作为评价本发明的依据。同时为了减少偶然因素的影响，将上述过程重复1000遍，每一次随机的划分训练集和测试集，最后取SROCC，即斯皮尔曼相关系数指数的中值作为最终的算法评价分数(见表1)。SROCC的值更接近于1表示算法与人类感知有更好的相关性。为了更加直观的显示各种算法的优劣关系，还绘制了各种算法的SROCC值的盒形图，如图2所示。In this embodiment, a LIVE image database is used to test the efficiency and performance of the present invention. In order to make a comparison, this embodiment uses some well-known full-reference image quality evaluation methods and no-reference evaluation quality methods as a comparison of the present invention. During the test, the database is divided into a test set and a training set, and according to the five-fold cross-validation method, the test set and the training set are set to 20% and 80% respectively. Then apply the two-step prediction framework based on this paper to predict the score of the test set, and the required classification and regression models are trained by the image features in the training set. Finally, the SROCC index of predicted score and actual score is calculated as the basis for evaluating the present invention. At the same time, in order to reduce the influence of accidental factors, the above process is repeated 1000 times, each time the training set and test set are randomly divided, and finally SROCC, that is, the median value of the Spearman correlation coefficient index, is taken as the final algorithm evaluation score (see Table 1 ). A value of SROCC closer to 1 indicates that the algorithm has a better correlation with human perception. In order to display the advantages and disadvantages of various algorithms more intuitively, a box diagram of the SROCC value of various algorithms is also drawn, as shown in Figure 2.

从表1可以看出，本发明的性能是在五种无参考算法(BIQI，DIIVINE，BLIINDS-II、BRISQUE和本发明)当中总体表现最好的。而且对于各个失真类别的图像都表现出很好的主观一致性，这就可以证明本发明有良好的通用性，并且在性能上有很大的优势。相对于其他三种全参考图像质量评价方法——峰值信噪比(PSNR)、结构相似度算法(SSIM)和视觉信息保真度算法(VIF)，本算法之比VIF算法的性能低。It can be seen from Table 1 that the performance of the present invention is the best overall among the five no-reference algorithms (BIQI, DIIVINE, BLIINDS-II, BRISQUE and the present invention). Moreover, the images of each distortion category show good subjective consistency, which proves that the present invention has good versatility and has great advantages in performance. Compared with other three full-reference image quality assessment methods—Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Algorithm (SSIM) and Visual Information Fidelity Algorithm (VIF), the performance of this algorithm is lower than VIF algorithm.

表1LIVE库中各算法主观一致性指标(SROCC)比较Table 1 Comparison of the subjective consistency index (SROCC) of each algorithm in the LIVE library

JP2KJP2K JPEGJPEG NOISENOISE BLURBLUR FFFF ALLALL PSNRPSNR 0.89900.8990 0.84840.8484 0.98350.9835 0.80760.8076 0.89860.8986 0.82930.8293

SSIMSSIM 0.95100.9510 0.91730.9173 0.96970.9697 0.95130.9513 0.95550.9555 0.89960.8996 VIFVIF 0.95150.9515 0.91040.9104 0.98440.9844 0.97220.9722 0.96310.9631 0.95210.9521 BIQIBIQI 0.85510.8551 0.77670.7767 0.97640.9764 0.92580.9258 0.76950.7695 0.75990.7599 DIIVINEDIIVINE 0.93520.9352 0.89210.8921 0.98280.9828 0.95510.9551 0.90960.9096 0.91740.9174 BLIINDS-IIBLIINDS-II 0.94620.9462 0.93500.9350 0.96340.9634 0.93360.9336 0.89920.8992 0.93290.9329 BRISQUEBRISQUE 0.94450.9445 0.92210.9221 0.98890.9889 0.95780.9578 0.91730.9173 0.94320.9432 Proposedproposed 0.95310.9531 0.94120.9412 0.98580.9858 0.96890.9689 0.90790.9079 0.94760.9476

为了证明本发明在时间复杂度上拥有良好的表现，先将本发明与三种著名的无参考评价算法的时间复杂度做出了比较(DIIVINE，BLIINDS-II，BRISQUE和本发明中提出的方法)。在表2中，列出了这四种算法计算LIVEIQA数据库中982幅图像所需的总体时间和平均每幅图像所需的时间。从中可以知道本发明的时间效率较另外两种无参考算法DIIVINE和BLIINDS-II高出很多，而比BRISQUE略微逊色。In order to prove that the present invention has a good performance in the time complexity, the time complexity of the present invention and three well-known no-reference evaluation algorithms are compared (DIIVINE, BLIINDS-II, BRISQUE and the method proposed in the present invention ). In Table 2, the overall time required by these four algorithms to calculate the 982 images in the LIVEIQA database and the average time required for each image are listed. It can be seen that the time efficiency of the present invention is much higher than that of the other two no-reference algorithms DIIVINE and BLIINDS-II, and slightly inferior to BRISQUE.

表2无参考方法的时间复杂度比较Table 2 Comparison of time complexity of no reference method

982幅图像所用时间Time taken for 982 images 平均每幅图像所用时间average time per image DIIVINEDIIVINE 2.9519*10(4)s2.9519*10(4)s 30.529430.5294 BLIINDS-IIBLIINDS-II 1.3112*10(5)s1.3112*10(5)s 133.5213133.5213 BRISQUEBRISQUE 109.3859s109.3859s 0.111391s0.111391s 本发明this invention 275.7670275.7670 0.280822s0.280822s

Claims

1. A method for evaluating image quality without reference based on gradient correlation, characterized in that it may further comprise the steps:

Step 1, performing feature extraction on the input distorted image;

Firstly, three different sub-properties in terms of image gradient are obtained for each image, namely, the gradient magnitude property GM, the gradient direction change property CO and the gradient magnitude change property CM, among which the three gradient properties are respectively defined as follows:

G G M m ((i i,, j j)) = = \sqrt{{((G G x x ((i i,, j j))))}^{22} + + {((G G y the y ((i i,, j j))))}^{22}} - - - - - - ((11))

CO(i,j)=orientation(i,j)-orientation _avg (i,j)(2)

C C M m ((i i,, j j)) = = \sqrt{{((G G x x ((i i,, j j)) - - {Gx Gx}_{a a v v g g} ((i i,, j j))))}^{22} + + {((G G y the y ((i i,, j j)) - - {Gy Gy}_{a a v v g g} ((i i,, j j))))}^{22}} - - - - - - ((33))

in,

o o r r i i e e n no t t a a t t i i o o n no ((i i,, j j)) = = arctan arctan ((\frac{G G y the y ((i i,, j j))}{G G x x ((i i,, j j))}))

{orientation orientation}_{a a v v g g} ((i i,, j j)) = = arctan arctan ((\frac{{Gy Gy}_{a a v v g g} ((i i,, j j))}{{Gx Gx}_{a a v v g g} ((i i,, j j))}))

{Gx Gx}_{a a v v g g} = = \frac{{Σ Σ}_{i i}^{M m} {Σ Σ}_{j j}^{N N} G G x x ((i i,, j j))}{M m \times \times N N}

{Gy Gy}_{a a v v g g} = = \frac{{Σ Σ}_{i i}^{M m} {Σ Σ}_{j j}^{N N} G G y the y ((i i,, j j))}{M m \times \times N N}

Among them, orientation(i, j) represents the direction of the image gradient, Gx and Gy are the derivatives of the discrete digital image in the two orthogonal directions of X and Y, and M and N are the windows set to describe the changes in the region size; and in calculating the gradients Gx and Gy, use the sobel gradient calculation operator, and extend it from a set of orthogonal directions to two groups; which includes a set of orthogonal directions of 0 degrees and 90 degrees, and a set of Orthogonal directions at -45 degrees and 45 degrees;

The sobel operators of 0 degrees and 90 degrees in the orthogonal direction are:

S^{the y} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & - 1 & 0 \\ 0 & 2 & 0 & - 2 & 0 \\ 0 & 1 & 0 & - 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]

and

S^{x} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 2 & 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \\ 0 & - 1 & - 2 & - 1 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]

The sobel operators of -45 degrees and 45 degrees in the orthogonal direction are:

S^{the y} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 1 & 0 & - 1 & 0 & 0 \\ 0 & 2 & 0 & 2 & 0 \\ 0 & 0 & 1 & 0 & - 1 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]

and

S^{x} = [\begin{matrix} 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 & - 1 \\ 0 & 2 & 0 & - 2 & 0 \\ 1 & 0 & - 1 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{matrix}]

The operator with the superscript y is the operator for calculating Gy, and the operator with the superscript x is the operator for calculating Gx;

Then, divide the above obtained image gradient property image into blocks, calculate the statistical variance on the gradient property of each block, on this basis, standardize the variance with the log function, and calculate the mean value of each block The statistical variance obtained by the gradient is fused into the variance of the overall image gradient property, as follows:

Divide each image gradient property into 128×128 image gradient property image blocks, and calculate the statistical variance of each image gradient property image block, that is, the variance d _n of the nth image block; the calculation formula is as follows 4:

d _n =Σ(h(x)-E(h(x))) ² (4)

in,

h(x)=pdf(θ)

Among them, θ is the parameter in each property image, pdf is the statistical distribution of the parameter, h(x) represents the statistical probability representation after quantitative statistics of θ, E(h(x)) represents the expectation of h(x) ;

Then normalize each calculated variance d _n with the log function, and use the average aggregation method to fuse the image block with the gradient property of the image to obtain the statistical variance as the final feature f, as shown in formula 5;

f = \frac{1}{h} \times l o g (Π_{no = 1}^{h} d_{no}), - - - (5)

n=1,2.......H, where H is the number of image blocks with image gradient properties

Finally, the input distorted image is down-sampled to a second-scale image, and the above process is repeated to finally obtain a set of 12-dimensional feature vectors;

Step 2, perform feature mapping, use the 12-dimensional feature vector obtained in step 1 as the final image quality feature, and establish a mapping relationship between image features and image scores accordingly;

First, the image database is divided into a training set and a test set; the test set is used to establish the mapping relationship between image features and image quality, and the test set is used to test the function of the established mapping relationship; using the method of support vector machine, Carry out the training of the distortion classification model and the corresponding quality evaluation model of each distortion classification with the image features of the images in the test set;

Then, based on the process of the two-step framework in the no-reference image quality evaluation, the score prediction test is carried out in the test set, that is, the distortion classification model is used to classify the distortion of the tested image, and then the quality evaluation model is used in the distortion classification. The measured image quality is obtained to obtain the tested image quality score, which can then be evaluated using existing algorithm performance standards.