WO2016145571A1 - 基于条件直方图码书的盲图像质量评价方法 - Google Patents

基于条件直方图码书的盲图像质量评价方法 Download PDF

Info

Publication number
WO2016145571A1
WO2016145571A1 PCT/CN2015/074203 CN2015074203W WO2016145571A1 WO 2016145571 A1 WO2016145571 A1 WO 2016145571A1 CN 2015074203 W CN2015074203 W CN 2015074203W WO 2016145571 A1 WO2016145571 A1 WO 2016145571A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
histogram
codebook
dimensional feature
evaluated
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2015/074203
Other languages
English (en)
French (fr)
Inventor
储颖
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen University
Original Assignee
Shenzhen University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen University filed Critical Shenzhen University
Priority to PCT/CN2015/074203 priority Critical patent/WO2016145571A1/zh
Publication of WO2016145571A1 publication Critical patent/WO2016145571A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis

Definitions

  • the invention relates to the field of blind image quality evaluation, in particular to a blind image quality evaluation method based on a conditional histogram codebook.
  • Image quality is the main indicator for evaluating the performance of image video processing systems and algorithms. Therefore, the research on Image Quality Assessment (IQA) method is very important and has broad prospects.
  • IQA Image Quality Assessment
  • the Image Quality Assessment (IQA) method is devoted to the study of how to use objective indicators to evaluate the perceived quality of an image, which is highly consistent with the subjective score of the image.
  • IQA index can be used as the basis for performance comparison and parameter selection of each algorithm; in the field of image coding and communication, IQA indicator can be used to guide the whole image compression, transmission and reception process, and Evaluate different algorithms and system performance; in the field of network transmission quality monitoring, IQA indicators can replace manual operations to achieve automated real-time monitoring; in addition, IQA has potential in many subject areas such as image processing algorithm optimization, biometrics, and medical image compression. Practical value.
  • FR Full-Reference
  • RR Reduced-Reference
  • NR No-Reference
  • VQEG Video Quality Experts Group
  • the NR IQA algorithm can be divided into two types for single distortion and for general distortion.
  • the former mainly includes prediction of JPEG compressed images, JPEG2000 compressed images and blurred image quality; while the latter cannot be known for any information about image distortion (including distortion type), also known as blind image quality assessment (Blind Image Quality Assessment) , BIQA).
  • an image quality objective evaluation algorithm with excellent performance should have the following three characteristics: (1) Accuracy: the difference between objective evaluation value and subjective evaluation value is small; (2) Monotonicity: objective evaluation value with subjective evaluation Increase or decrease in value; (3) Consistency: The performance of the objective evaluation algorithm on the test set is similar to that exhibited on the training set. Commonly used quantitative indicators for measuring the performance of algorithms are:
  • x, y represent the subjective and objective scores of a single picture
  • i represents the picture number
  • N represents the total number of pictures.
  • D represents the difference between ranks, where the absolute value of the difference between the subjective data and the objective data.
  • x, y represent the subjective and objective scores of a single picture, respectively, and N is the number of test pictures.
  • SCCH method Shape Consistency of Conditional Histogram
  • the SCCH method does not need to train the extracted perceptual features and human subjective scoring, and has better stability and practicability than the traditional learning strategy-based blind IQA method.
  • the SCCH method includes two parts of determining the shape consistency feature and the blind image quality evaluation method based on the joint condition histogram of adjacent DNT coefficients, which are respectively discussed below.
  • the traditional blind image quality evaluation method generally extracts the image quality sensitive features in the sense sense from the distribution of the image itself or its linear transform coefficients, but ignores the important physiological and physical characteristics of the human visual system.
  • the goal of early optic neuron processing is to maximize the statistical independence between neural responses in order to be more efficient. Up to the image. For natural images, there is a strong correlation between adjacent image pixels in the spatial domain; in the linear decomposition domain, there is also a high-order statistical dependence between adjacent coefficients. In fact, the DNT transform can significantly reduce the statistical dependence between the extracted components, which can better describe the nonlinear mechanism of the biological vision system.
  • the DNT transform is a nonlinear transform based on linear decomposition of images. Its operation is performed on linear decomposition coefficients, mainly simulating local gain control functions in optic neurophysiology.
  • image DCT coefficients or arbitrary wavelet transform coefficients can be used as inputs to the DNT transform.
  • the steerable pyramid is used to decompose the wavelet, because the wavelet transform allows the number of scales and the number of directions to be decomposed, and it is convenient to obtain a partial description of the image in the scale, direction and spatial position.
  • y be the wavelet coefficient and the DNT coefficient z be defined as follows:
  • p is the DNT factor (p>0), representing the sum of the energy of the wavelet coefficient y on the scale, direction, and spatial neighboring coefficient sets.
  • DNT decomposition For natural images, the statistical independence between adjacent coefficients is enhanced by DNT decomposition, which can be shown by the joint condition histogram between adjacent DNT coefficients.
  • Figure 1 defines a pair of horizontally adjacent DNT coefficients that form a two-dimensional random variable (X, Y) whose sample values traverse row by column from the upper left corner of the DNT subband (x 1 , y 1 ) to the lower right corner (x N , y N ), N is the number of samples.
  • X and Y draw the normalized condition histogram H(Y
  • Figure 2 shows the variation of the joint condition histogram shape of the DNT subband horizontally adjacent DNT coefficients.
  • the degree of consistency of the joint condition histogram shape consistency can be used to characterize the degree of image quality degradation.
  • the adjacent DNT coefficient can be selected as the perceptual feature of the conditional histogram shape consistency to construct the blind image quality evaluation index without training learning.
  • each DNT subband is translated by one pixel along the 0, -45, -90, and -135 degrees directions (as shown in Figure 4), and four spaces are obtained.
  • Adjacent sub-bands are constructed to construct four joint condition histograms, and four shape consistency features are extracted.
  • these vectors are scattered in the feature space, and no obvious similarity is observed.
  • some potential aggregation exists between these feature attributes.
  • the shape consistency feature vector-normal length of the natural reference image in the two public databases LIVE and CSIQ databases is statistically found to be approximately equal to a constant.
  • f 1 , f 2 and f 3 represent the shape-consistent feature sub-vectors of scales 1 , 2 and 3 , respectively:
  • the length vector m (
  • 1 ) is defined.
  • o i represents the subjective score of the ith natural image or the distorted image in the LIVE database
  • m i represents the i-th image-a generic length vector
  • N represents the total number of LIVE database references and distorted images.
  • the calculation result ⁇ d i ⁇ is sorted in ascending order, and the smallest L values are selected to construct the set ⁇ s j ⁇ ; using ⁇ s j ⁇ and the corresponding subjective score interpolation, the image quality score q is constructed as follows:
  • ⁇ is the parameter that controls the interpolation weighted decay rate, and the set value is 32.
  • L is the number of interpolated images closest to the test image in the feature attribute space. It is verified by the LIVE and CSIQ database experiments, and is set to the empirical value of 80, which is 10% of the total number of images in the LIVE database.
  • the existing blind image quality evaluation system includes two core modules: feature extraction and feature attribute dictionary construction.
  • 6 is a block diagram of a blind image quality evaluation system in the prior art scheme
  • FIG. 7 is a schematic diagram of a feature extraction module of the prior art scheme.
  • the existing scheme constructs the feature attribute dictionary by using the conditional histogram shape consistency feature of the database image, and only uses the length attribute of the feature vector when matching, and the image quality is relatively rough, which leads to the general performance of the algorithm.
  • the existing scheme uses the feature attributes of each distorted image in the database as the characters in the dictionary, and the distribution of each character in the feature attribute space depends on the database image characteristics, and the construction method cannot Ensure that there are representative feature attributes within each subjective rating level that may affect the accuracy of the final image quality score.
  • the technical problem to be solved by the present invention is to provide a blind image quality evaluation method based on a conditional histogram codebook for the defects of the prior art.
  • the technical solution adopted by the present invention to solve the technical problem thereof is: a blind image quality evaluation method based on a conditional histogram codebook, comprising the following steps:
  • the step S1 comprises the following steps:
  • the step S3 comprises the following steps:
  • Equation (14) min( ⁇ ) refers to the minimum value, m l,k refers to the kth codeword of the first level in the conditional histogram codebook, and K refers to the code of each subjective rating level.
  • x i refers to the i-th high-dimensional feature vector of the training distortion image, and G l,k refers to the high-dimensional feature vector set of the first level and the k-th codeword of the training distortion image.
  • 2 means the Euclidean distance.
  • the clustering method comprises a spectral clustering method of first dimension reduction and then K-means clustering.
  • the step S4 comprises the following steps:
  • y i is the i-th high-dimensional feature vector of the distortion image to be evaluated
  • m l, k is the k-th codeword of the first level in the conditional histogram code book
  • d i, l, k is the Euclidean distance of the i-th high-dimensional feature vector y i of the distortion image to be evaluated and the k-th codeword m l,k of the first level of the conditional histogram codebook
  • S42 selecting, within each subjective scoring level of the conditional histogram codebook, a codeword closest to the i-th high-dimensional feature vector of the to-be-evaluated distortion image as a high dimension with the to-be-evaluated distortion image.
  • the matching vector code matches the feature vector.
  • the step S5 comprises the following steps:
  • q i is the quality score corresponding to the i-th high-dimensional feature vector of the distortion image to be evaluated
  • L is the number of subjective rating levels
  • o l is the subjective image corresponding to the first subjective rating level.
  • the score, d i,l is the Euclidean distance of the matching codeword whose i-th high-dimensional feature vector matches the first subjective rating level
  • ⁇ is the weighting coefficient of the attenuation speed
  • q refers to the quality evaluation index of the distorted image
  • M refers to the number of high-dimensional feature vectors of the distortion image to be evaluated.
  • the preset database includes one or more of a LIVE database, a CSIQ database, and a TID2008 database.
  • the present invention has the following advantages: implementing the present invention, generating a conditional histogram code book by using a clustering method, so that different decomposition directions and different adjacent relationships exhibit different shape patterns (possibly the same degree of distortion)
  • the conditional histogram curve can be distinguished and classified into corresponding subjective scoring grades, so that the blind image quality evaluation method provided by the present invention is more delicate in describing the statistical independence degree of the joint condition histogram of adjacent DNT coefficients. According to the calculated quality evaluation indicators, it is more accurate.
  • Figure 1 is a joint condition histogram of horizontal DNT coefficients of natural images.
  • 3 is a natural image level 1 pixel adjacent DNT coefficient joint condition histogram and average condition histogram.
  • FIG. 4 is a schematic diagram of a shape consistency sensing feature vector extraction scheme.
  • FIG. 5 is a shape-consistent feature-normal-length vector distribution map of a LIVE database natural reference image and five typical distorted images.
  • FIG. 6 is a block diagram of a blind image quality evaluation system of the prior art.
  • FIG. 7 is a schematic diagram of a feature extraction module of the prior art scheme.
  • FIG. 8 is a system block diagram of a blind image quality evaluation method based on a conditional histogram codebook of the present invention.
  • Fig. 9 is a schematic diagram showing feature extraction of a distorted image in the present invention.
  • 11 is a high dimensional feature vector x diagram of a Gaussian blurred image adjacent DNT coefficient combined with a conditional histogram.
  • Figure 12 is a post-cluster conditional histogram codebook.
  • the blind image quality evaluation method based on a conditional histogram codebook includes the following steps:
  • S1 Feature extraction is performed on the distorted image to obtain a neighboring DNT coefficient joint condition histogram of the distorted image, and the distorted image includes a distortion image for training and a distortion image to be evaluated.
  • the database image in Fig. 8 is a distortion image for training, and the distortion image is a distortion image to be evaluated.
  • step S1 includes the following steps:
  • step S2 connecting the obtained training distortion image and the normalized adjacent DNT coefficient of the distortion image to be evaluated to the end of each curve in the conditional histogram, respectively obtaining the high-dimensional feature vector of the distortion image for training and the distortion image to be evaluated.
  • High dimensional feature vector It can be understood that step S2 adjusts the structure of the shape consistency feature based on the conditional histogram shape consistency feature extraction in the prior art solution, and replaces the calculated value with the conditional histogram pattern mode to better characterize Statistical independence between adjacent DNT coefficients. Taking the H(Y
  • X) curve of the central region in X 5,6,...,13 in Fig. 3 as an example, we connect the H(Y
  • each "series” feature As can be seen from Fig. 10, as the statistical independence between adjacent DNT coefficients of the distorted image is weakened or strengthened, the degree of shape consistency of each "series” feature corresponding to the joint condition histogram changes.
  • the “series” feature For Gaussian white noise distortion, the “series” feature exhibits periodic signal characteristics because each H(Y
  • step S3 Clustering high-dimensional feature vectors of all the training distortion images in the preset database to obtain all codewords in the cluster-based conditional histogram codebook and the conditional histogram codebook.
  • the preset database includes one or more of a LIVE database, a CSIQ database, and a TID2008 database. It can be understood that for each single distortion type, as the image subjective score DMOS value changes, the shape of each "series" feature will change accordingly. Therefore, considering the joint condition of adjacent DNT coefficients for the training distortion image The high-dimensional feature vector obtained by "series" in the histogram Row clustering. Specifically, step S3 includes the following steps:
  • S31 Selecting a preset database as a codebook generation library.
  • one of the LIVE database, the CSIQ database, and the TID2008 database may be selected, or another database may be selected as the codebook generation library.
  • the LIVE database is used. For example.
  • S32 The subjective evaluation of the distorted image in the codebook generation library is divided into L subjective scoring grades, and the clustering method is used to cluster the high-dimensional feature vectors of the training distorted image respectively in each subjective scoring level, and each of the presets is preset.
  • the clustering center of the subjective scoring level is K to obtain a cluster-based conditional histogram codebook for training distorted images. It can be understood that each cluster center is a codeword.
  • the conditional histogram codebook is composed of K ⁇ L codewords.
  • the existing shape consistent perception feature extraction scheme strategy is used in this patent to extract features from adjacent sub-bands of 1 pixel space.
  • wavelet and DNT transform the image is decomposed into 3 scales and 4 directions.
  • the "series" features of different scales, the same decomposition and translation direction have strong shape consistency (see Figure 11), in order to facilitate clustering.
  • the clustering method used is a spectral clustering method of first dimension reduction and then K-means clustering.
  • Equation (14) min( ⁇ ) refers to the minimum value, m l,k refers to the kth codeword of the first level in the conditional histogram codebook, and K refers to the number of codewords for each subjective rating level.
  • x i refers to the i-th high-dimensional feature vector of the distortion image for training
  • G l,k refers to the high-dimensional feature vector set of the first level and the k-th code word of the training distortion image
  • 2 refers to the Euclidean distance.
  • each cluster center in the conditional histogram code book is gradually Weakened.
  • the signal shape shows significant irregularities.
  • the codewords of subjective rating levels 1 and 2 exhibit better periodic changes in the signal or periodic variations in the envelope. This shows that by clustering, each codeword nicely gathers the shape uniform distribution characteristics and modes of the high-dimensional feature vector of the distortion image of its neighborhood training.
  • step S4 After the conditional histogram codebook is determined, in each subjective scoring level of the conditional histogram codebook, a matching code matching the high-dimensional feature vector of the to-be-evaluated distortion image is found in all the codewords of the conditional histogram codebook. word.
  • the high-dimensional feature vector of the distortion image to be evaluated is extracted, and the matching codeword matching the matched codeword is found in all the codewords of the conditional histogram codebook determined in step S3, the matching The codeword is the codeword closest to the high dimensional feature vector of the distorted image to be evaluated.
  • step S4 includes the following steps:
  • Equation (15): y i is the i-th high-dimensional feature vector of the distortion image to be evaluated, and m l,k is the k-th codeword of the first level in the conditional histogram, d i,l,k is The Ei's distance of the i-th high-dimensional feature vector y i of the distortion image to be evaluated and the k-th codeword m l,k of the first level of the conditional histogram code book; M is the number of high-dimensional feature vectors of the distortion image to be evaluated .
  • S42 Select, within each subjective scoring level of the conditional histogram codebook, a codeword closest to the i-th high-dimensional feature vector of the to-be-evaluated distortion image as a matching match with the high-dimensional feature vector of the to-be-evaluated distortion image.
  • step S5 Calculate a quality evaluation index of the distortion image to be evaluated according to the matching codeword. Specifically, step S5 includes the following steps:
  • S52 Perform interpolation in the subjective scoring level of the conditional histogram codebook according to the high-dimensional feature vector of the distortion image to be evaluated and the Euclidean distance of the matching codeword, and calculate each high-dimensional feature vector in the to-be-evaluated distortion image.
  • Quality score
  • Equation (16): q i is the quality score corresponding to the i-th high-dimensional feature vector of the distortion image to be evaluated, L is the number of subjective rating levels, and o l is the subjective score of the image corresponding to the first subjective rating level.
  • d i,l is the Euclidean distance of the matching codeword whose i-th high-dimensional feature vector matches the first subjective rating level, and ⁇ is the weighting coefficient of the attenuation speed.
  • Equation (17) q refers to the quality evaluation index of the distorted image, and M refers to the number of high-dimensional feature vectors of the distorted image to be evaluated.
  • conditional histogram codebook is generated by using the clustering method, so that the conditional histogram curves with different shape directions (possibly the same degree of distortion) can be distinguished by different decomposition directions and different adjacent relationships, and can be distinguished.
  • the corresponding subjective scoring level is adopted, so that the blind image quality evaluation method provided by the present invention is more delicate in describing the statistical independence degree of the joint condition histogram of adjacent DNT coefficients, so the image quality scoring formula is constructed in the same manner.
  • the quality evaluation indicators are more accurate.
  • this patent selects JPEG2000 compression, JPEG compression, Gaussian white noise and Gaussian blur distortion image in LIVE database as the training distortion image, and uses the spectral clustering method to generate the Code Book of Conditional Histogram. And named the blind image quality evaluation method as CBCH algorithm. It should be emphasized that the CBCH algorithm does not require training after quality evaluation of the distortion image to be evaluated after the conditional histogram codebook is determined, and the image quality score can be calculated immediately, which has very good practicability. Tables 2 through 7 show the results of the CCCCH algorithm comparing the LCC and SROCC performances of the existing schemes on the LIVE, CSIQ, and TID2008 databases, respectively.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

本发明公开了一种基于条件直方图码书的盲图像质量评价方法，该方法包括如下步骤：对失真图像进行特征提取，以获得失真图像的相邻DNT系数联合条件直方图，失真图像包括训练用失真图像和待评估失真图像；将相邻DNT系数联合条件直方图中各曲线首尾相连，以获得失真图像的高维特征矢量；对预设数据库中所有训练用失真图像的高维特征矢量进行聚类，以获得基于聚类的条件直方图码书和条件直方图码书中所有码字；在条件直方图码书所有码字中寻找与待评估失真图像的高维特征矢量相匹配的匹配码字；根据匹配码字，计算待评估失真图像的质量评价指标。本发明对相邻DNT系数联合条件直方图统计独立性程度的描述方式更为细腻，其算出的图像质量评价指标更加准确。

Description

基于条件直方图码书的盲图像质量评价方法 技术领域
本发明涉及盲图像质量评价领域，尤其涉及一种基于条件直方图码书的盲图像质量评价方法。
背景技术
图像质量是评价图像视频处理系统及算法性能的主要指标，因此对图像质量评价(Image Quality Assessment，IQA)方法的研究非常重要且具有广泛的前景。图像质量评价(Image Quality Assessment，IQA)方法致力于研究如何用客观指标评价图像的感知质量，使其与图像主观得分高度一致。在图像去噪、图像恢复、图像增强等领域，IQA指标可以作为各算法性能比较、参数选择的依据；在图像编码与通信领域，IQA指标可以用来指导整个图像压缩、传输、接收过程，并评估不同算法及系统性能；在网络传输质量监控领域，IQA指标可以取代人工操作，实现自动化实时监控；此外，IQA在图像处理算法优化、生物特征识别、医学图像压缩等众多学科领域都具有潜在的实用价值。
绝大部分应用场合中，需要对失真图像进行质量评估。根据其原始图像信息是否完备，可将IQA算法分为三类：全参考型(Full-Reference，FR)、部分参考型(Reduced-Reference，RR)和无参考型(No-Reference，NR)IQA。其中，NR IQA是最具挑战性的质量评价任务。鉴于其在数字通信及传输质量监控等领域不可替代的重要作用，VQEG(Video Quality Experts Group)组织已将NR IQA标准化工作列为其未来重要发展方向之一。
根据评价图像的失真类型范围不同，NR IQA算法又可分为针对单一失真和针对通用失真两种类型。其中，前者主要包括对JPEG压缩图像、JPEG2000压缩图像以及模糊图像质量的预测；而后者由于无法获知有关图像失真(包括失真类型)的任何信息，也被称为盲图像质量评价(Blind Image Quality Assessment，BIQA)。
一般来说，性能优良的图像质量客观评价算法应该具备以下三个特性：(1)准确性：客观评价值与主观评价值之间差异较小；(2)单调性：客观评价值随主观评价值增减而增减；(3)一致性：客观评价算法在测试集上展现的性能与其在训练集上表现出的性能相似。常用的衡量算法性能的定量指标有：
1)线性相关系数(Linear Correlation Coefficient，LCC)：
Figure PCTCN2015074203-appb-000001
其中，x，y分别代表单张图片的主观评分和客观得分，i代表图片编号，N代表图片总数，
Figure PCTCN2015074203-appb-000002
分别代表全部图片的主观评分平均值和客观得分平均值。
2)秩相关系数(Spearman Rank Order Correlation Coefficient，SROCC)：
Figure PCTCN2015074203-appb-000003
其中，D代表等级间差异，此处为主观数据与客观数据的差值绝对值。
3)均方根误差(Root Mean Squared Error，RMSE)：
Figure PCTCN2015074203-appb-000004
其中，x，y分别代表单张图片的主观评分和客观得分，N为测试图片数量。
针对传统盲IQA系统训练学习过程复杂、评价指标性能稳定性差等问题，现有技术提出一种采用形状一致性特征(Shape Consistency of Conditional Histogram，以下简称SCCH方法)、且无需训练的盲图像质量评价方法(详见储颖，牟轩沁，洪伟，“采用形状一致性特征的盲图像质量评价方法，”西安交通大学学报，第48卷，第8期，pp.12-17，2014年8月)。SCCH方法通过提取图像DNT域(Divisive Normalization Transform，DNT)形状一致性特征，以及构造特征属性-主观评分字典，确立了具有明确映射关系的图像质量评分公式。同时，SCCH方法无需对所提取感知特征以及人类主观评分进行训练，比起传统基于学习策略的盲IQA方法，具有更好的稳定性和实用性。SCCH方法包括确定基于相邻DNT系数联合条件直方图的形状一致性特征和盲图像质量评价方法两部分，以下分别进行论述。
1.基于相邻DNT系数联合条件直方图的形状一致性特征
传统盲图像质量评价方法普遍从图像本身或其线性变换系数的分布中抽取感知意义下的图像质量敏感特征，但忽略了人类视觉系统的重要生理物理学特性。根据视神经生理学研究，早期视神经元处理的目的是尽量增加神经响应间的统计独立性，以便更高效地表 达图像。对自然图像，空间域中相邻图像像素之间存在极强的相关性；在线性分解域中，相邻系数之间也存在高阶统计依赖性。事实上，DNT变换能显著降低所提取分量间的统计依赖性，可更好地描述生物视觉系统非线性机制。
1.1相邻DNT系数联合条件直方图
DNT变换是一种建立在图像线性分解基础上的非线性变换，其操作针对线性分解系数进行，主要模拟视神经生理学中的局部增益控制功能。理论上，图像DCT系数或任意小波变换系数均可作为DNT变换的输入。此处选用可操纵金字塔分解小波，因为该小波变换允许设置分解的尺度数和方向数，方便获得图像在尺度、方向和空间位置上的局部描述。设y代表小波系数，DNT系数z定义如下：
z＝y/p     (4)
式中：p为DNT因子(p＞0)，代表小波系数y在尺度、方向和空间相邻系数集上的能量总和。
计算DNT因子的方法有很多，此处采用高斯尺度混合模型。
对自然图像，经DNT分解后相邻系数间统计独立性得到增强，可通过相邻DNT系数间联合条件直方图展示。图1定义出一对水平相邻DNT系数组成二维随机变量(X，Y)，其样本值从DNT子带左上角(x1，y1)逐行逐列遍历至右下角(xN，yN)，N为样本个数。为进一步观察X与Y间统计依赖程度，绘制归一化条件直方图H(Y|X)，设离散化级数等于17，统计(X，Y)的归一化联合直方图H(X，Y)如图1所示。
如图1所示，无论X取何值，各条H(Y|X)曲线的形状非常接近，整组曲线在很大程度上保持形状一致性。这说明，Y的分布不依赖于X的取值，X和Y是近似统计独立的。
考察自然图像垂直相邻、对角相邻以及尺度相邻关系下DNT系数联合条件直方图，均可观察到各条曲线形状保持一致。接下来，需考察相邻DNT系数联合条件直方图形状一致性在失真图像上的变化情况。
1.2失真图像联合条件直方图形状一致性
以JPEG2000、JPEG、高斯白噪声和高斯模糊典型失真为例，当失真程度加剧时，图2展示了DNT子带水平相邻DNT系数联合条件直方图形状的变化情况。
由图2a、2b、2d可见，随着失真程度加剧，图2e、2f、2h中各条件直方图曲线形状一致性变差，说明相邻DNT系数间统计独立程度下降。这可以解释为：JPEG2000压缩、JPEG压缩和高斯模糊3种图像退化过程中，图像结构的破坏导致了DNT分量间统计独立性的丧失，并且失真越严重，这种独立性的改变越大。图2c中，对高斯白噪声失真，曲线形状一致性变化趋势却相反。随着失真程度加剧，各联合条件直方图形状趋向一致，表明相邻DNT系数间的统计独立性加强。这一现象也符合逻辑，因为高斯白噪声信号与图像信号完全独立，在图像中加入高斯白噪声信号，意味着引入了更多统计独立成分。
综上，无论相邻DNT系数间统计独立性加强或减弱，其联合条件直方图形状一致性改变程度均可表征图像质量退化程度。可选取相邻DNT系数联合条件直方图形状一致性作为感知特征，以构造无需训练学习的盲图像质量评价指标。
1.3联合条件直方图形状一致性特征
为定量描述相邻DNT系数联合条件直方图间形状一致性，定义平均条件直方图
Figure PCTCN2015074203-appb-000005
如下：
Figure PCTCN2015074203-appb-000006
式中，考虑到在联合直方图两侧边缘区域，归一化条件直方图曲线形状规律性不强，不适宜参与整组曲线形状一致性计算。因此为降低边缘区域影响，仅选取X＝5，6，...，13共9条的中心区域曲线参与运算，如图3所示。
由图3可见，若整组联合条件直方图形状较一致，则平均条件直方图与各条件直方图曲线重合度应较高，各曲线与平均条件直方图间距离应较小。每一根条件直方图曲线与平均条件直方图曲线的距离可用KL(Kullback-Leibler)距离描述，定义如下：
Figure PCTCN2015074203-appb-000007
从而整组联合条件直方图曲线的形状一致性特征定义为：
Figure PCTCN2015074203-appb-000008
f越小，联合条件直方图曲线组形状一致性越好；f越大，整组曲线的形状一致性越差。
2.盲图像质量评价方法
对整幅图像，小波及DNT分解中存在3种基本相邻关系：空间相邻、尺度相邻和方向相邻。其中，空间相邻关系对盲图像质量评价模型的性能起关键主导作用。基于此，仅从空间相邻子带中提取特征。对分解尺度为3、分解方向为4的DNT变换，将每一DNT子带分别沿0、-45、-90和-135度方向平移一个像素(如图4所示)，可得到4个空间相邻子带，从而构造4幅联合条件直方图，提取出4个形状一致性特征。因此，对于每一幅自然图像，均可提取出维数为48(4空间×3尺度×4方向)的形状一致性感知特征矢量f＝(f1，f2，...，f48)。表面上看，这些矢量在特征空间中分散分布，观察不到明显相似性。事实上，作为自然图像感知特征矢量，某种潜在的聚集性是存在于这些特征属性之间的。例如，对两大公开数据库LIVE和CSIQ数据库中自然参考图像的形状一致性特征矢量一范长度进行统计，发现近似等于常数。
Figure PCTCN2015074203-appb-000009
定义f1、f2和f3分别代表尺度1、2和3的形状一致性特征子矢量，则：
f＝(f1，f2，f3)     (9)
fi＝(f1，f2，...，f16)i，i＝1，2，3     (10)
令||f1||1、||f2||1和||f3||1分别代表各子矢量的一范长度，比较结果如表1所示。
表1 LIVE和CSIQ数据库中自然图像形状一致性特征矢量一范长度均值和标准差比较
Figure PCTCN2015074203-appb-000010
为了更好地观察自然及失真图像形状一致性特征矢量长度属性的变化规律，定义长度矢量m＝(||f1||1，||f2||1，||f3||1)，并绘制LIVE数据库中所有图像的一范长度矢量分布图，如图5。
由图5可见，LIVE数据库中自然参考图像一范长度矢量的位置很接近，可以认为近似 聚集于一点；各失真图像的一范长度矢量随失真程度加剧，沿同方向近似成束状延伸；特别是JPEG2000、JPEG和高斯白噪声失真，可以观察到良好的线性和单调性。构造特征属性-主观得分字典如下：
D＝(mi，oi)，i＝1，2，...，N     (11)
式中：oi代表LIVE数据库中第i幅自然图像或失真图像的主观得分；mi代表第i幅图像一范长度矢量；N代表LIVE数据库参考和失真图像总数。
对每一幅待评价失真图像，分别计算其m与字典中所有mi的欧氏距离di：
di＝||m-mi||2，i＝1，2，...N     (12)
按照升序对计算结果{di}排序，并选择最小的L个值构造集合{sj}；利用{sj}及对应的主观得分插值，构造图像质量评分q如下：
Figure PCTCN2015074203-appb-000011
式中：λ是控制插值加权衰减速度的参数，设置值取32(实验表明，当λ在[1，64]区间变化时，方法评估值的波动范围仅为万分之几，几乎不受λ取值影响)；L为特征属性空间中距离测试图像最近的插值图像数目，经LIVE和CSIQ数据库实验验证，设定为经验值80，即LIVE数据库中图像总数的10％。
综上，现有盲图像质量评价系统包括两个核心模块：特征提取和特征属性字典构造。图6是现有技术方案中盲图像质量评价系统框图，图7是现有技术方案特征提取模块示意图。在LIVE和CSIQ两大公开数据库上的实验结果表明，SCCH方法性能良好(线性相关系数均大于0.82)、无需训练学习、且评分公式形式简单。
现有方案利用数据库图像的条件直方图形状一致性特征构造特征属性字典，匹配时仅使用特征矢量的长度属性，对图像质量的刻画方式比较粗糙，导致算法性能一般。此外，构造特征属性字典时，现有方案将数据库中每一幅失真图像的特征属性分别作为字典中的字元，各字元在特征属性空间中的分布取决于数据库图像特性，该构造方法无法确保每一主观评分等级内均存在有代表性的特征属性，从而可能影响最终图像质量评分的准确性。
发明内容
本发明要解决的技术问题在于，针对现有技术的缺陷，提供一种基于条件直方图码书的盲图像质量评价方法。
本发明解决其技术问题所采用的技术方案是：一种基于条件直方图码书的盲图像质量评价方法，包括如下步骤：
S1：对失真图像进行特征提取，以获得所述失真图像的相邻DNT系数联合条件直方图，所述失真图像包括训练用失真图像和待评估失真图像；
S2：将所述相邻DNT系数联合条件直方图中各曲线首尾相连，以获得所述失真图像的高维特征矢量；
S3：对预设数据库中所有所述训练用失真图像的高维特征矢量进行聚类，以获得基于聚类的条件直方图码书和所述条件直方图码书中所有码字；
S4：在所述条件直方图码书所有码字中寻找与所述待评估失真图像的高维特征矢量相匹配的匹配码字；
S5：根据所述匹配码字，计算所述待评估失真图像的质量评价指标。
优选地，所述步骤S1包括如下步骤：
S11：对所述失真图像进行小波分解，以获得多尺度多方向的小波分解子带；
S12：对所述多尺度多方向小波分解子带进行DNT变换，以获得联合直方图；
S13：对所述联合直方图中各联合条件直方图进行归一化处理，以获取归一化的相邻DNT系数联合条件直方图。
优选地，所述步骤S3包括如下步骤：
S31：选择预设数据库作为码书生成库；
S32：将所述码书生成库中的失真图像主观评价分均分成L个主观评分等级，在各主观评分等级内采用聚类方法对所述训练用失真图像的高维特征矢量分别聚类，预设每一主观评分等级的聚类中心为K个，以获取所述训练用失真图像的基于聚类的条件直方图码书；
S33：在所述条件直方图码书中获取所有所述聚类中心，所述聚类中心为所述训练用失真图像的码字，所述码字根据如下条件确定：
Figure PCTCN2015074203-appb-000012
式(14)中：min(·)是指求最小值，ml，k是指所述条件直方图码书中第l等级的第k个 码字，K是指每一主观评分等级的码字数目，xi是指所述训练用失真图像的第i个高维特征矢量，Gl，k是指所述训练用失真图像的第l等级、第k个码字的高维特征矢量集合，||·||2是指欧氏距离。
优选地，所述聚类方法包括先降维再K-means聚类的谱聚类方法。
优选地，所述步骤S4包括如下步骤：
S41：计算所述待评估失真图像的高维特征矢量与所述条件直方图码书所有码字的欧式距离：
di，l，k＝||xi-ml，k||2，i＝1，2，...M，l＝1，2，...L，k＝1，2，...K     (15)
式(15)中：yi是指所述待评估失真图像的第i个高维特征矢量，ml，k是指所述条件直方图码书中第l等级的第k个码字，di，l，k是指所述待评估失真图像的第i个高维特征矢量yi与所述条件直方图码书第l等级的第k个码字ml，k的欧氏距离；M是指所述待评估失真图像高维特征矢量数目；
S42：在所述条件直方图码书每一主观评分等级内，分别选取与所述待评估失真图像的第i个高维特征矢量最接近的码字作为与所述待评估失真图像的高维特征矢量相匹配的匹配码字。
优选地，所述步骤S5包括如下步骤：
S51：计算所述待评估失真图像的高维特征矢量与所述匹配码字的欧式距离：
di，l＝min(di，l，k)，k＝1，2，...，K，其中，di，l是指所述待评估失真图像的第i个高维特征矢量与第l主观评分等级相匹配的匹配码字的欧氏距离；
S52：根据所述待评估失真图像的高维特征矢量与所述匹配码字的欧式距离计算所述待评估失真图像中每一高维特征矢量的质量评分：
Figure PCTCN2015074203-appb-000013
式(16)中：qi是指所述待评估失真图像的第i个高维特征矢量对应的质量评分，L是指主观评分等级数目，ol是指第l主观评分等级对应的图像主观评分，di，l是指第i个高维特 征矢量与第l主观评分等级匹配的匹配码字的欧氏距离，λ是指衰减速度的权重系数；
S53：计算所述高维特征矢量的质量评分的平均值作为所述待评估失真图像的质量评价指标，所述待评估失真图像的质量评价指标：
Figure PCTCN2015074203-appb-000014
式(17)中：q是指失真图像的质量评价指标，M是指所述待评估失真图像高维特征矢量数目。
优选地，所述预设数据库包括LIVE数据库、CSIQ数据库和TID2008数据库中的一种或多种。
本发明与现有技术相比具有如下优点：实施本发明，采用聚类方式生成条件直方图码书，以使不同分解方向、不同相邻关系所呈现出具有不同形状模式(可能是相同失真程度)的条件直方图曲线能够得以区分，并归入相应的主观评分等级，使得采用本发明所提供的盲图像质量评价方法在对相邻DNT系数联合条件直方图统计独立性程度的描述方式更细腻，根据其计算出的质量评价指标更加准确。
附图说明
下面将结合附图及实施例对本发明作进一步说明，附图中：
图1是自然图像水平相邻DNT系数联合条件直方图。
图2是不同失真类型图像及对应的水平相邻DNT系数联合条件直方图。
图3是自然图像水平1像素相邻DNT系数联合条件直方图和平均条件直方图。
图4是形状一致性感知特征矢量提取方案示意图。
图5是LIVE数据库自然参考图像和5种典型失真图像的形状一致性特征一范长度矢量分布图。
图6是现有技术方案盲图像质量评价系统框图。
图7是现有技术方案特征提取模块示意图。
图8是本发明基于条件直方图码书的盲图像质量评价方法的系统框图。
图9是本发明中对失真图像进行特征提取的示意图。
图10是失真图像及对应的水平1像素相邻DNT系数联合条件直方图串联特征图。
图11是高斯模糊图像相邻DNT系数联合条件直方图串联后的高维特征矢量x图。
图12是聚类后条件直方图码书。
具体实施方式
为了对本发明的技术特征、目的和效果有更加清楚的理解，现对照附图详细说明本发明的具体实施方式。
图8示出本发明基于条件直方图码书的盲图像质量评价方法的系统框图，如图8所示，该基于条件直方图码书的盲图像质量评价方法包括如下步骤：
S1：对失真图像进行特征提取，以获得失真图像的相邻DNT系数联合条件直方图，失真图像包括训练用失真图像和待评估失真图像。图8中的数据库图像为训练用失真图像，失真图像为待评估失真图像。
如图9所示，步骤S1包括如下步骤：
S11：对失真图像进行小波分解，以获得多尺度多方向的小波分解子带；
S12：对多尺度多方向小波分解子带进行DNT变换，以获得联合直方图；
S13：对联合直方图中各联合条件直方图进行归一化处理，以获得归一化的相邻DNT系数联合条件直方图。
S2：将获得的训练用失真图像和待评估失真图像的归一化的相邻DNT系数联合条件直方图中各曲线首尾相连，分别获得训练用失真图像的高维特征矢量和待评估失真图像的高维特征矢量。可以理解地，步骤S2在现有技术方案基于条件直方图形状一致性特征提取的基础上，对形状一致性特征的构造方式进行调整，用条件直方图图案模式代替计算数值，以更好地刻画相邻DNT系数之间的统计独立性。以图3中X＝5，6，...，13共9条位于中心区域的H(Y|X)曲线为例，我们将各条H(Y|X)曲线头尾相连，形成了图10所示的相邻DNT系数联合条件直方图“串联”特征图。
由图10可见，随着失真图像相邻DNT系数间统计独立性减弱或加强，各“串联”特征对应联合条件直方图的形状一致性程度发生了改变。对于高斯白噪声失真，“串联”特征呈现出周期信号特点，因为各H(Y|X)曲线高度相似；对于JPEG2000压缩、JPEG压缩和高斯模糊三种失真，“串联”特征的内部形状一致性则急剧下降。
S3：对预设数据库中所有所述训练用失真图像的高维特征矢量进行聚类，以获得基于聚类的条件直方图码书和条件直方图码书中所有码字。具体地，预设数据库包括LIVE数据库、CSIQ数据库和TID2008数据库中的一种或多种。可以理解地，对每一种单一失真类型，随着图像主观评分DMOS值的改变，各“串联”特征的形状将随之发生变化，因此，考虑对训练用失真图像的相邻DNT系数联合条件直方图中“串联”得到的高维特征矢量进 行聚类。具体地，步骤S3包括如下步骤：
S31：选择预设数据库作为码书生成库，可以理解地，可以选择LIVE数据库、CSIQ数据库和TID2008数据库中的一种，也可以选择其他数据库作为码书生成库，本实施例中，以LIVE数据库为例。
S32：将码书生成库中的失真图像主观评价分均分成L个主观评分等级，在各主观评分等级内采用聚类方法对训练用失真图像的高维特征矢量分别聚类，预设每一主观评分等级的聚类中心为K个，以获取训练用失真图像的基于聚类的条件直方图码书。可以理解地，每一聚类中心为一个码字，本实施例中，条件直方图码书由K×L个码字构成。
具体地，由于空间相邻关系对盲图像质量评价模型的性能起关键主导作用，本专利中沿用现有的形状一致性感知特征提取方案策略，从1像素空间相邻子带中提取特征。在小波及DNT变换中，图像被分解为3个尺度、4个方向。考虑到不同方向、不同相邻关系的条件直方图模式不尽相同，而不同尺度、同一分解和平移方向的“串联”特征则具有较强的形状一致性(参见图11)，为便于聚类，我们将相同方向、相同相邻关系的3个“串联”特征再次“串联”，得到459维(17维×9条×3尺度)高维特征矢量x。因此，对每一幅图像，可提取高维特征矢量数目为16(4方向×4相邻关系)。由于高维特征矢量x的维度高，故本实施例中，所采用的聚类方法为先降维再K-means聚类的谱聚类方法。
S33：在条件直方图码书中获取所有聚类中心，聚类中心为训练用失真图像的码字，码字根据如下条件确定：
Figure PCTCN2015074203-appb-000015
式(14)中：min(·)是指求最小值，ml，k是指条件直方图码书中第l等级的第k个码字，K是指每一主观评分等级的码字数目，xi是指训练用失真图像的第i个高维特征矢量，Gl，k是指训练用失真图像的第l等级、第k个码字的高维特征矢量集合，||·||2是指欧氏距离。
以应用最为广泛的LIVE数据库为例，我们把四种典型失真(JPEG2000压缩、JPEG压缩、高斯白噪声和高斯模糊)的训练用失真图像按各自主评价分均分至L个主观评分等级，在每一主观评分等级内计算各训练用失真图像的高维特征矢量，并将它们作为条件直方图码书的高维特征矢量，采用谱聚类方式生成每一主观评分等级的K个聚类中心(即 码字)。可以理解地，最优聚类中心数目和最佳均分等级数目均通过实验确定，如图12展示了通过谱聚类得到的条件直方图码书的聚类中心数目K＝8，主观评分等级数目L＝5。
由图12可见，随着失真图像质量逐渐下降(主观评分等级区间由1至5递增)，条件直方图码书中每一级聚类中心(码字)自身呈现出的形状一致性程度在逐渐减弱。尤其是主观评分等级4和5的某些码字，信号形状表现出明显的无规律性。相比之下，主观评分等级1和2的码字则呈现出较好的信号周期性变化或包络周期性变化规律。这说明：通过聚类，各码字很好地聚集了其邻域训练用失真图像的高维特征矢量的形状一致性分布特点及模式。
S4：条件直方图码书确定后，在条件直方图码书的每一主观评分等级内，在条件直方图码书所有码字中寻找与待评估失真图像的高维特征矢量相匹配的匹配码字。对于待评估失真图像，根据步骤S1和步骤S2，提取待评估失真图像的高维特征矢量，在步骤S3中确定的条件直方图码书所有码字中寻找与其相匹配的匹配码字，该匹配码字为与待评估失真图像的高维特征矢量最接近的码字。
具体地，步骤S4包括如下步骤：
S41：计算待评估失真图像的高维特征矢量与条件直方图码书所有码字的欧式距离：
di，l，k＝||yi-ml，k||2，i＝1，2，...，M，l＝1，2,...，L，k＝1，2，...，K     (15)
式(15)中：yi是指待评估失真图像的第i个高维特征矢量，ml，k是指条件直方图中第l等级的第k个码字，di，l，k是指待评估失真图像的第i个高维特征矢量yi与条件直方图码书第l等级的第k个码字ml，k的欧氏距离；M是待评估失真图像高维特征矢量数目。
S42：在条件直方图码书每一主观评分等级内，分别选取与待评估失真图像的第i个高维特征矢量最接近的码字作为与待评估失真图像的高维特征矢量相匹配的匹配码字。具体地，通过di，l＝min(di，l，k)，k＝1，2，...，K计算与待评估失真图像的高维特征矢量相匹配的匹配码字，其中，di，l是指待评估失真图像的第i个高维特征矢量与第l主观评分等级相匹配的匹配码字的欧氏距离。
S5：根据匹配码字，计算待评估失真图像的质量评价指标。具体地，步骤S5包括如下步骤：
S51：计算待评估失真图像的高维特征矢量与匹配码字的欧式距离：
di，l＝min(di，l，k)，k＝1，2，...，K，其中，di，l是指待评估失真图像的第i个高维特征矢量与第l主观评分等级相匹配的匹配码字的欧氏距离。
S52：在条件直方图码书的每一主观评分等级内，根据待评估失真图像的高维特征矢量与匹配码字的欧式距离做插值运算，计算待评估失真图像中每一高维特征矢量的质量评分：
Figure PCTCN2015074203-appb-000016
式(16)中：qi是指待评估失真图像的第i个高维特征矢量对应的质量评分，L是指主观评分等级数目，ol是指第l主观评分等级对应的图像主观评分，di，l是指第i个高维特征矢量与第l主观评分等级匹配的匹配码字的欧氏距离，λ是指衰减速度的权重系数。
S53：计算高维特征矢量的质量评分的平均值作为待评估失真图像的质量评价指标，待评估失真图像的质量评价指标：
Figure PCTCN2015074203-appb-000017
式(17)中：q是指失真图像的质量评价指标，M是指待评估失真图像高维特征矢量数目。
实施本发明，采用聚类方式生成条件直方图码书，以使不同分解方向、不同相邻关系所呈现出具有不同形状模式(可能是相同失真程度)的条件直方图曲线能够得以区分，并归入相应的主观评分等级，使得采用本发明所提供的盲图像质量评价方法在对相邻DNT系数联合条件直方图统计独立性程度的描述方式更细腻，故采用相同方式构造图像质量评分公式时得到的质量评价指标更加准确。
经实验验证与比较，本专利选取LIVE数据库中JPEG2000压缩、JPEG压缩、高斯白噪声和高斯模糊失真图像作为训练用失真图像，采用谱聚类方法生成条件直方图码书(Code Book of Conditional Histogram)，并命名该盲图像质量评价方法为CBCH算法。需要强调的是，CBCH算法在条件直方图码书确定之后，对待评估失真图像进行质量评估时不需要训练，可以立刻计算出图像质量评分，具有非常好的实用性。表2至表7分别列出了CBCH算法在LIVE、CSIQ和TID2008数据库上与现有方案的LCC和SROCC性能比较结果。
表2 本专利方案与现有方案在LIVE数据库上的LCC性能比较
Figure PCTCN2015074203-appb-000018
表3 本专利方案与现有方案在LIVE数据库上的SROCC性能比较
Figure PCTCN2015074203-appb-000019
表4 本专利方案与现有方案在CSIQ数据库上的LCC性能比较
Figure PCTCN2015074203-appb-000020
表5 本专利方案与现有方案在CSIQ数据库上的SROCC性能比较
Figure PCTCN2015074203-appb-000021
表6 本专利方案与现有方案在TID2008数据库上的LCC性能比较
Figure PCTCN2015074203-appb-000022
表7 本专利方案与现有方案在TID2008数据库上的SROCC性能比较
Figure PCTCN2015074203-appb-000023
由表2至表7可见，CBCH算法在三大数据库上的性能远远超越了现有技术方案算法，表现非常出色。这说明，采用本发明对条件直方图的高维特征矢量聚类生成条件直方图码书的方法能够显著改善图像质量评价性能，通过对相邻DNT系数联合条件直方图图案模式聚类并构造条件直方图码书的盲图像质量评价方法是可行和有效的。
本发明是通过一个具体实施例进行说明的，本领域技术人员应当明白，在不脱离本发明范围的情况下，还可以对本发明进行各种变换和等同替代。另外，针对特定情形或具体情况，可以对本发明做各种修改，而不脱离本发明的范围。因此，本发明不局限于所公开的具体实施例，而应当包括落入本发明权利要求范围内的全部实施方式。

Claims (7)

  1. 一种基于条件直方图码书的盲图像质量评价方法，其特征在于，包括如下步骤：
    S1：对失真图像进行特征提取，以获得所述失真图像的相邻DNT系数联合条件直方图，所述失真图像包括训练用失真图像和待评估失真图像；
    S2：将所述相邻DNT系数联合条件直方图中各曲线首尾相连，以获得所述失真图像的高维特征矢量；
    S3：对预设数据库中所有所述训练用失真图像的高维特征矢量进行聚类，以获得基于聚类的条件直方图码书和所述条件直方图码书中所有码字；
    S4：在所述条件直方图码书所有码字中寻找与所述待评估失真图像的高维特征矢量相匹配的匹配码字；
    S5：根据所述匹配码字，计算所述待评估失真图像的质量评价指标。
  2. 根据权利要求1所述的基于条件直方图码书的盲图像质量评价方法，其特征在于，所述步骤S1包括如下步骤：
    S11：对所述失真图像进行小波分解，以获得多尺度多方向的小波分解子带；
    S12：对所述多尺度多方向小波分解子带进行DNT变换，以获得联合直方图；
    S13：对所述联合直方图中各联合条件直方图进行归一化处理，以获得归一化的相邻DNT系数联合条件直方图。
  3. 根据权利要求2所述的基于条件直方图码书的盲图像质量评价方法，其特征在于，所述步骤S3包括如下步骤：
    S31：选择预设数据库作为码书生成库；
    S32：将所述码书生成库中的失真图像主观评价分均分成L个主观评分等级，在各主观评分等级内采用聚类方法对所述训练用失真图像的高维特征矢量分别聚类，预设每一主观评分等级的聚类中心为K个，以获取所述训练用失真图像的基于聚类的条件直方图码书；
    S33：在所述条件直方图码书中获取所有所述聚类中心，所述聚类中心为所述训练用失真图像的码字，所述码字根据如下条件确定：
    Figure PCTCN2015074203-appb-100001
    式(14)中：min(·)是指求最小值，ml，k是指所述条件直方图码书中第l等级的第k个码字，K是指每一主观评分等级的码字数目，xi是指所述训练用失真图像的第i个高维特征矢量，Gl，k是指所述训练用失真图像的第l等级、第k个码字的高维特征矢量集合，||·||2是指欧氏距离。
  4. 根据权利要求3所述的基于条件直方图码书的盲图像质量评价方法，其特征在于，所述聚类方法包括先降维再K-means聚类的谱聚类方法。
  5. 根据权利要求3所述的基于条件直方图码书的盲图像质量评价方法，其特征在于，所述步骤S4包括如下步骤：
    S41：计算所述待评估失真图像的高维特征矢量与所述条件直方图码书所有码字的欧式距离：
    Figure PCTCN2015074203-appb-100002
    式(15)中：yi是指所述待评估失真图像的第i个高维特征矢量，ml，k是指所述条件直方图码书中第l等级的第k个码字，di，l，k是指所述待评估失真图像的第i个高维特征矢量yi与所述条件直方图码书第l等级的第k个码字ml，k的欧氏距离；M是指所述待评估失真图像高维特征矢量数目；
    S42：在所述条件直方图码书每一主观评分等级内，分别选取与所述待评估失真图像的第i个高维特征矢量最接近的码字作为与所述待评估失真图像的高维特征矢量相匹配的匹配码字。
  6. 根据权利要求1所述的基于条件直方图码书的盲图像质量评价方法，其特征在于，所述步骤S5包括如下步骤：
    S51：计算所述待评估失真图像的高维特征矢量与所述匹配码字的欧式距离：di，l＝min(di，l，k)，k＝1，2，...，K，其中，di，l是指所述待评估失真图像的第i个高维特征矢量与第l主观评分等级相匹配的匹配码字的欧氏距离；
    S52：根据所述待评估失真图像的高维特征矢量与所述匹配码字的欧式距离计算所述待评估失真图像中每一高维特征矢量的质量评分：
    Figure PCTCN2015074203-appb-100003
    式(16)中：qi是指所述待评估失真图像的第i个高维特征矢量对应的质量评分，L是指主观评分等级数目，ol是指第l主观评分等级对应的图像主观评分，di，l是指第i个高维特征矢量与第l主观评分等级匹配的匹配码字的欧氏距离，λ是指衰减速度的权重系数；
    S53：计算所述高维特征矢量的质量评分的平均值作为所述待评估失真图像的质量评价指标，所述待评估失真图像的质量评价指标：
    Figure PCTCN2015074203-appb-100004
    式(17)中：q是指失真图像的质量评价指标，M是指所述待评估失真图像高维特征矢量数目。
  7. 根据权利要求1-6任一项所述的基于条件直方图码书的盲图像质量评价方法，其特征在于，所述预设数据库包括LIVE数据库、CSIQ数据库和TID2008数据库中的一种或多种。
PCT/CN2015/074203 2015-03-13 2015-03-13 基于条件直方图码书的盲图像质量评价方法 Ceased WO2016145571A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2015/074203 WO2016145571A1 (zh) 2015-03-13 2015-03-13 基于条件直方图码书的盲图像质量评价方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2015/074203 WO2016145571A1 (zh) 2015-03-13 2015-03-13 基于条件直方图码书的盲图像质量评价方法

Publications (1)

Publication Number Publication Date
WO2016145571A1 true WO2016145571A1 (zh) 2016-09-22

Family

ID=56918151

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/074203 Ceased WO2016145571A1 (zh) 2015-03-13 2015-03-13 基于条件直方图码书的盲图像质量评价方法

Country Status (1)

Country Link
WO (1) WO2016145571A1 (zh)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107743225A (zh) * 2017-10-16 2018-02-27 杭州电子科技大学 一种利用多层深度表征进行无参考图像质量预测的方法
CN107945154A (zh) * 2017-11-10 2018-04-20 西安电子科技大学 基于四元数离散余弦变换的彩色图像质量评价方法
CN108335289A (zh) * 2018-01-18 2018-07-27 天津大学 一种全参考融合的图像客观质量评价方法
CN109919921A (zh) * 2019-02-25 2019-06-21 天津大学 基于生成对抗网络的环境影响程度建模方法
CN110503632A (zh) * 2019-07-26 2019-11-26 南昌大学 一种盲图像质量评价算法中svr参数优化方法
CN111127298A (zh) * 2019-06-12 2020-05-08 上海大学 一种全景图像盲质量评估方法
CN114820541A (zh) * 2022-05-07 2022-07-29 武汉象点科技有限公司 基于重构网络的缺陷检测方法
CN115272663A (zh) * 2022-08-12 2022-11-01 杭州类脑科技有限公司 医学图像的病理分布生成方法、系统及计算机可读存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2009065151A2 (en) * 2007-11-15 2009-05-22 Certifi Media Inc. Method for image quality assessment using quality vectors
CN103366378A (zh) * 2013-07-26 2013-10-23 深圳大学 基于条件直方图形状一致性的无参考型图像质量评价方法
CN104751456A (zh) * 2015-03-13 2015-07-01 深圳大学 基于条件直方图码书的盲图像质量评价方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2009065151A2 (en) * 2007-11-15 2009-05-22 Certifi Media Inc. Method for image quality assessment using quality vectors
CN103366378A (zh) * 2013-07-26 2013-10-23 深圳大学 基于条件直方图形状一致性的无参考型图像质量评价方法
CN104751456A (zh) * 2015-03-13 2015-07-01 深圳大学 基于条件直方图码书的盲图像质量评价方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
SHI, YONGCHANG.: "No-Reference Image Quality Assessment of Two-dimension Code Based on Codebook", SHIP ELECTRONIC ENGINEERING, vol. 34, no. 244, 31 October 2014 (2014-10-31), pages 89, XP032080324 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107743225A (zh) * 2017-10-16 2018-02-27 杭州电子科技大学 一种利用多层深度表征进行无参考图像质量预测的方法
CN107945154A (zh) * 2017-11-10 2018-04-20 西安电子科技大学 基于四元数离散余弦变换的彩色图像质量评价方法
CN108335289A (zh) * 2018-01-18 2018-07-27 天津大学 一种全参考融合的图像客观质量评价方法
CN109919921A (zh) * 2019-02-25 2019-06-21 天津大学 基于生成对抗网络的环境影响程度建模方法
CN109919921B (zh) * 2019-02-25 2023-10-20 天津大学 基于生成对抗网络的环境影响程度建模方法
CN111127298A (zh) * 2019-06-12 2020-05-08 上海大学 一种全景图像盲质量评估方法
CN111127298B (zh) * 2019-06-12 2023-05-16 上海大学 一种全景图像盲质量评估方法
CN110503632A (zh) * 2019-07-26 2019-11-26 南昌大学 一种盲图像质量评价算法中svr参数优化方法
CN110503632B (zh) * 2019-07-26 2022-08-09 南昌大学 一种盲图像质量评价算法中svr参数优化方法
CN114820541A (zh) * 2022-05-07 2022-07-29 武汉象点科技有限公司 基于重构网络的缺陷检测方法
CN115272663A (zh) * 2022-08-12 2022-11-01 杭州类脑科技有限公司 医学图像的病理分布生成方法、系统及计算机可读存储介质

Similar Documents

Publication Publication Date Title
CN104751456B (zh) 基于条件直方图码书的盲图像质量评价方法
WO2016145571A1 (zh) 基于条件直方图码书的盲图像质量评价方法
Liu et al. Blind image quality assessment by relative gradient statistics and adaboosting neural network
CN103475898B (zh) 一种基于信息熵特征的无参考图像质量评价方法
CN109389591B (zh) 基于彩色描述子的彩色图像质量评价方法
CN104376565B (zh) 基于离散余弦变换和稀疏表示的无参考图像质量评价方法
CN110110596B (zh) 高光谱图像特征提取、分类模型构建及分类方法
CN108765414B (zh) 基于小波分解和自然场景统计的无参考立体图像质量评价方法
CN105006001B (zh) 一种基于非线性结构相似度偏差的有参图像的质量评估方法
CN107481236A (zh) 一种屏幕图像的质量评价方法
CN110415207A (zh) 一种基于图像失真类型的图像质量评估的方法
Vishwakarma et al. An efficient hybrid DWT-fuzzy filter in DCT domain based illumination normalization for face recognition
CN111325720B (zh) 一种多视觉特征集成的无参考超分辨图像质量评价方法
CN106651829B (zh) 一种基于能量和纹理分析的无参考图像客观质量评价方法
CN109829905A (zh) 一种面部美化感知质量的无参考评价方法
Jiang et al. Supervised dictionary learning for blind image quality assessment using quality-constraint sparse coding
CN108389189A (zh) 基于字典学习的立体图像质量评价方法
CN109816646A (zh) 一种基于退化决策逻辑的无参考图像质量评价方法
CN109788275A (zh) 自然性、结构和双目不对称无参考立体图像质量评价方法
Li et al. Blind image quality assessment based on joint log-contrast statistics
CN117115600A (zh) 无参考图像质量评价的方法、装置及电子设备
CN109754390B (zh) 一种基于混合视觉特征的无参考图像质量评价方法
CN108629771B (zh) 一种具有尺度鲁棒性的图像质量盲评价方法
Yang et al. Blind image quality measurement via data-driven transform-based feature enhancement
Cai et al. Towards a blind image quality evaluator using multi-scale second-order statistics

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15884955

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 15884955

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 15884955

Country of ref document: EP

Kind code of ref document: A1