WO2020062901A1 - Method and system for analyzing image quality of super-resolution image - Google Patents
Method and system for analyzing image quality of super-resolution image Download PDFInfo
- Publication number
- WO2020062901A1 WO2020062901A1 PCT/CN2019/088478 CN2019088478W WO2020062901A1 WO 2020062901 A1 WO2020062901 A1 WO 2020062901A1 CN 2019088478 W CN2019088478 W CN 2019088478W WO 2020062901 A1 WO2020062901 A1 WO 2020062901A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- super
- image
- component
- source reference
- resolution image
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/40—Analysis of texture
- G06T7/49—Analysis of texture based on structural texture description, e.g. using primitives or placement rules
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30168—Image quality inspection
Definitions
- the invention relates to the technical field of digital image processing and computer vision, and in particular, to an image quality analysis method and system for super-resolution images.
- Image super-resolution is an important research topic in the field of image processing. By performing super-resolution processing on the image, the resolution of the image can be increased and the image becomes clearer. It has many important applications in image reconstruction, such as:
- Super-resolution reconstruction technology can be used for image compression. Low-resolution image information is usually stored or transmitted. When there are different needs, super-resolution reconstruction technology is used to obtain images and videos of different resolutions.
- image quality analysis There have been many methods of image quality analysis, such as PSNR algorithm (Peak Signal to Noise Ratio), MOS algorithm (Mean opinion scores, subjective quality score), and SSIM algorithm (Structural and SIMilarity, image similarity) .
- PSNR algorithm Peak Signal to Noise Ratio
- MOS algorithm Mean opinion scores, subjective quality score
- SSIM algorithm Structuretural and SIMilarity, image similarity
- image quality analysis uses many other features and information. For example, the information fidelity criterion and visual information fidelity, the mutual information between the reference image and the distorted image is calculated as the visual quality.
- an object of the present invention is to provide an image quality analysis method and system for super-resolution images, which overcomes the shortcomings of the prior art that lack quality analysis methods for super-resolution images.
- a first embodiment provided by the present invention is a method for quality analysis of a super-resolution image, which includes the following steps:
- the three pooling scores are fused to obtain the image quality analysis value.
- the step of decomposing the texture components of the source reference image and the super-resolution image includes:
- the texture components of the super-resolution image and the source reference image are divided into multiple image blocks, and the variance values of the same-sized image blocks in the two images of the super-resolution image and the source reference image are calculated, and the maximum of the two variance values is obtained. value;
- the texture component is calculated by combining the inner product value, the maximum value of the variance value, and a preset first adjustment factor.
- the step of calculating the statistical feature vector of each pixel in the super-resolution image and the source reference image separately includes:
- the step of decomposing the structural components of the source reference image and the super-resolution image includes:
- the step of decomposing the source reference image and the super-resolution image into high-frequency components includes:
- the step of decomposing the source reference image and the super-resolution image into high-frequency components includes:
- the high-frequency energy vectors of the source reference image and the super-resolution image are calculated, and the product of the two norms of the high-frequency energy vector and the high-frequency energy vector is used as a high-frequency component.
- the calculation formula of the high-frequency energy is:
- N (i) is the adjacent pixel of pixel i
- N N is the number of adjacent pixels
- s is the structural component
- s ⁇ is the s through a Gaussian convolution kernel with a variance of ⁇ Product obtained
- s ⁇ represents the low frequency part of s.
- the step of pooling the decomposed texture component, structure component, and high-frequency component separately to obtain three pooling scores corresponding to the three components further includes:
- a weighted average is performed on the texture component, the structure component, and the high-frequency component.
- the step of fusing the three pooling scores to obtain an image quality analysis value includes:
- the image quality analysis value is calculated by the following formula:
- p is an image quality analysis value
- ⁇ and ⁇ are adjusted shadow coefficients
- p t , p s , and p h are weighted average values of the texture component, structure component, and high-frequency component, respectively.
- a second embodiment provided by the present invention is a quality analysis system for super-resolution images, which includes:
- An image information acquisition module configured to acquire a source reference image and a super-resolution image having the same content and number of pixels as the source reference image;
- An image component decomposition module configured to decompose the texture component, the structural component, and the high-frequency component of the source reference image and the super-resolution image respectively;
- a pooling processing module configured to separately pool the decomposed texture component, structure component, and high-frequency component to obtain three pooling scores corresponding to the three components;
- a score fusion module is used to fuse the three pooling scores to obtain an image quality analysis value.
- the present invention provides an image quality analysis method and system for a super-resolution image, by acquiring a source reference image and acquiring a super-resolution image with the same content and number of pixels as the source reference image; Texture components, structural components, and high-frequency components of the source reference image and super-resolution image; and pooling the decomposed texture components, structural components, and high-frequency components, respectively, to obtain three corresponding to the three components. Pooling scores; the three pooling scores are combined to obtain image quality analysis values.
- the present invention is closer to the human visual evaluation mechanism than the existing image quality analysis methods, and the objective experimental data is superior to the existing image quality analysis methods. It provides important application value for the analysis of super-resolution image quality after reconstruction of low-resolution images in the surveillance field, satellite images and medical images.
- FIG. 1 is a flowchart of steps of an image quality analysis method for a super-resolution image provided by the present invention
- FIG. 2 is a schematic structural diagram of an image quality analysis system of the super-resolution image provided by the present invention.
- the first embodiment provided by the present invention is a method for quality analysis of a super-resolution image. As shown in FIG. 1, the method includes the following steps:
- Step S1 Acquire a source reference image and acquire a super-resolution image with the same content and number of pixels as the source reference image.
- the source reference image and the super-resolution image have the same content and the number of pixels.
- the super-resolution image is compared with the source reference image to obtain the quality analysis value of the super-resolution image. Therefore, in this step, the super-resolution image and its source reference image to be compared are first obtained.
- Step S2 Decompose the texture component, the structure component, and the high-frequency component of the source reference image and the super-resolution image, respectively.
- the texture components, structural components, and high-frequency components of the source reference image and the super-resolution image are decomposed, and the three components of the two images are compared in turn. According to the comparison results of the three components, the quality analysis results are obtained.
- Step S3 Perform pooling on the resolved texture component, structure component, and high-frequency component respectively to obtain three pooling scores corresponding to the three components.
- Step S4 Fusion the three pooling scores to obtain an image quality analysis value.
- the step of fusing the three pooling scores in this step to obtain an image quality analysis value includes: using the image quality analysis value described below to calculate the image quality analysis value:
- p is an image quality analysis value
- ⁇ and ⁇ are adjusted shadow coefficients
- p t , p s , and p h are weighted average values of the texture component, structure component, and high-frequency component, respectively.
- the three components are pooled, and the pooling scores are combined to obtain the quality analysis value of the super-resolution image.
- the above step S2 includes decomposing the source reference image and the super-resolution picture into texture components and structural components, respectively, and extracting high-frequency components from the source reference image and the super-resolution picture, respectively, and comparing the image quality of each component.
- the step of decomposing the texture components of the source reference image and the super-resolution image includes:
- the texture components of the super-resolution image and the source reference image are divided into multiple image blocks, and the variance values of the same-sized image blocks in the two images of the super-resolution image and the source reference image are calculated, and the maximum of the two variance values is obtained. value;
- the texture component is calculated by combining the inner product value, the maximum value of the variance value, and a preset first adjustment factor.
- Texture components use statistical features, such as SIFT histogram statistical features to describe texture features. Given a source reference image r and a super-resolution image u with the same content and the same number of pixels, let the texture component be represented as Mt. For the i-th pixel in the images u and r:
- t r (i) and tu (i) are image blocks centered on the i-th pixel in the texture components of the super-resolution image r and the source reference image u, respectively, and the two image blocks are the same size.
- C t is a normal amount used to adjust the range of K t .
- var ( ⁇ ) is used to calculate the variance of its parameters, and max ( ⁇ , ⁇ ) is used to return the maximum value of its parameters.
- the texture components M t can be calculated by formulas (1) and (2), and it is easy to prove that the range of M t is 0 to 1.
- the step of calculating the statistical feature vector of each pixel in the super-resolution image and the source reference image includes: using a direction gradient histogram algorithm, a scale-invariant feature matching algorithm, or a local binary pattern algorithm to obtain a super-resolution image and A statistical feature vector for each pixel in the source reference image.
- the statistical features are calculated to calculate the texture components.
- the method for actually calculating the statistical features may also be a HOG algorithm.
- the HOG algorithm may extract a 3780-dimensional vector as a statistical feature vector in calculating the texture components.
- SIFT algorithm SIFT algorithm can extract 128-dimensional vector as the statistical feature vector in computing texture components.
- the LBP algorithm uses the statistical histogram of the LBP feature spectrum as the statistical feature vector. There are also statistical features based on histograms. It is conceivable that the HOG algorithm, the SIFT algorithm, the LBP algorithm, and the statistical feature vectors extracted based on histogram statistics and the like are arbitrarily combined, and the combined vector can also be used as a statistical feature vector in calculating the texture components.
- the step of decomposing the structural components of the source reference image and the super-resolution image includes:
- the decomposition using the gradient feature with the structural component is detailed as follows:
- the structural component be represented as M s .
- the main direction of each image block in the image needs to be calculated first.
- the main direction of the image block is obtained by the semi-definite matrix J:
- i is a central pixel of an image block
- g x and g y are gradient vectors of a horizontal axis and a vertical axis in the image block, respectively.
- J has two eigenvalues and two eigenvectors, and the eigenvector corresponding to the smaller eigenvalue is the main direction of the gradient of the image block.
- n r (i) and n u (i) are normalized feature vectors, which are the main directions of the i-th pixel in the structural component of the super-resolution image r and the source reference image u, respectively.
- the range of M s is 0 to 1.
- K s is defined in formula (4) as follows:
- g mr (i) and g mu (i) represent the normalized gradient magnitudes of the i-th pixel in the structural component of the super-resolution image r and the source reference image u, respectively.
- C s is a normal quantity used to adjust the range of K s .
- the main component gradient of the image block is calculated to calculate the structural component.
- the actual calculated structural component can also be extracted from the texture component by the relative total variation method, and the structural component can also be obtained using an iterable rolling guidance filter.
- canny operator, sobel operator, etc. may also be used, and the size of the image block may be variable.
- steps of decomposing the source reference image and the super-resolution image into high-frequency components respectively include:
- the step of decomposing the source reference image and the super-resolution image into high-frequency components includes:
- the high-frequency energy vectors of the source reference image and the super-resolution image are calculated, and the product of the two norms of the high-frequency energy vector and the high-frequency energy vector is used as a high-frequency component.
- the following uses the characteristics of high-frequency components in different frequency bands as examples to further analyze the decomposition method.
- the high-frequency component be represented as M h , and the high-frequency component is obtained by formula (7).
- the high-frequency energy of the super-resolution image r and the source reference image u must be compared.
- N (i) is the neighboring pixels of pixel i
- N N is the number of neighboring pixels
- s is the structural component.
- s ⁇ is obtained by convolving s with a Gaussian convolution kernel with variance ⁇ , and s ⁇ represents the low frequency part of s.
- h r and h u are the high-frequency energy calculated by formula (6) of the super-resolution image r and the source reference image u , respectively.
- C h is a normal quantity to avoid instability caused by too small denominator.
- a Gaussian convolution kernel is used to calculate high-frequency components.
- high-frequency components can also be calculated using different frequency band characteristics. The following formula for calculating high-frequency energy:
- h (i) represents the high-frequency energy of the i-th pixel, which is a vector
- k can be taken as (2, 3, 4 7), and different Gaussian convolution kernels are subtracted to obtain frequency components in different frequency bands.
- the combined vector is used as the high-frequency energy h.
- the high-frequency energy h is a vector.
- the formula for calculating high-frequency components is as follows:
- high frequency energy is 2 times the vector image and the super-resolution image of the reference source the inner product
- the ratio of the number 2 norm of each vector as the high frequency energy high-frequency component of the super-resolution image quality analysis, C h is a Normal number, used to avoid instability caused by too small denominator.
- the high-frequency energy vectors can also be multiplied by their respective norms as the characteristics of the high-frequency energy, and other arbitrary norms can be taken.
- the step of pooling the decomposed texture component, structure component, and high-frequency component respectively to obtain three pooling scores corresponding to the three components further includes:
- a weighted average is performed on the texture component, the structure component, and the high-frequency component.
- the steps for processing the three components include:
- the pooled texture component, structure component and high-frequency component need weighted average.
- the weighted average formula is as follows:
- N is the number of image pixels
- q ⁇ ⁇ t, s, h ⁇ respectively represents the index of the texture component, the structural component and the high-frequency component
- p q represents the pooling score of the three components
- w q is each component The weight of each pixel, w q is obtained from the results calculated by formulas (9) to (11).
- t r (i) and t u (i) are calculated in the same way as in formula (2)
- g mr (i) and g mu (i) are calculated in the same way as in formula (5)
- h r (i) and h u (i ) Is calculated in the same way as formula (6).
- the denominators in equations (9) to (11) are set for normalization.
- ⁇ > 0 and ⁇ > 0 are for adjusting the influence degree of different components in the fusion process.
- p t , p s , p h are calculated by formula (8), respectively.
- the value of ⁇ can be obtained by comparing the average intensity of the structural component and the texture component, as shown in formula (13):
- s and t are the intensity of the image structural component and the texture component, respectively.
- mean ( ⁇ ) is the mean function and log ( ⁇ ) is the logarithmic function.
- the value of ⁇ is greater than 1.
- the Weber-Fechner rule is used to obtain ⁇ , and ⁇ can also be obtained using an exhaustive method.
- the method can be improved or replaced as follows:
- ⁇ is the Lagrangian multiplier coefficient
- u is the texture component
- BV is the Banach space
- ⁇ is a region of the picture
- f is a picture.
- other optimization models such as Meyer, Vese-Osher can also be used.
- the calculation of the adaptive variable K t in the present invention may actually be the following three formulas:
- mean (.) Represents the mean function.
- the formula for calculating ⁇ in the present invention may actually be the following two formulas:
- sum (.) Represents the summation function.
- Calculating ⁇ can actually be a combination of a sum function and a variance function.
- ⁇ is calculated to be 3.9709 using an exhaustive method.
- the present invention uses a structural texture decomposition method to implement image visual quality analysis on super-resolution pictures. It provides a more objective analysis method for the quality analysis after the source reference image is restored to the super-resolution image, and avoids the blindness of subjectively evaluating the quality of the super-resolution image. It provides important application value for the analysis of super-resolution image quality after reconstruction of low-resolution images in the surveillance field, satellite images and medical images. Under certain conditions, the super-resolution reconstruction technology can overcome the limitation of the inherent resolution of the image system and improve the resolution of the processed image. Therefore, it has very important applications in the fields of video, remote sensing, medicine, and security monitoring.
- the second embodiment provided by the present invention is a quality analysis system for super-resolution images, as shown in FIG. 2, including:
- the image information acquisition module 310 is configured to acquire a source reference image and a super-resolution image having the same content and number of pixels as the source reference image; its function is as described in step S1.
- the image component decomposition module 320 is configured to decompose the texture component, the structure component, and the high-frequency component of the source reference image and the super-resolution image, respectively, and its function is as described in step S2.
- the pooling processing module 330 is configured to pool the resolved texture component, structure component, and high-frequency component respectively to obtain three pooling scores corresponding to the three components; its function is as described in step S3.
- the score fusion module 340 is configured to fuse the three pooling scores to obtain an image quality analysis value. Its function is as described in step S4.
- texture components, structural components, and high-frequency components obtained in different ways can be arbitrarily combined.
- the present invention provides a method and system for image quality analysis of a super-resolution image, by acquiring a source reference image and a super-resolution image having the same content and number of pixels as the source reference image; and decomposing the source separately
- the texture components, structural components, and high-frequency components of the reference image and the super-resolution image; and the pooled texture components, structural components, and high-frequency components are separately pooled to obtain three pooling corresponding to the three components Score; the three pooling scores are combined to obtain an image quality analysis value.
- the present invention is closer to the human visual evaluation mechanism than the existing image quality analysis methods, and the objective experimental data is superior to the existing image quality analysis methods.
- the invention can be applied to other up-sampled images and many existing image databases.
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Quality & Reliability (AREA)
- Image Processing (AREA)
Abstract
Provided by the present invention are a method and system for analyzing the image quality of a super-resolution image, the method comprising: acquiring a source reference image and a super-resolution image having the same content and number of pixels as the source reference image; decomposing texture components, structural components and high frequency components of the source reference image and the super-resolution image respectively; pooling the decomposed texture components, structural components and high frequency components to obtain three pooling fractions corresponding to the three components; and fusing the three pooling fractions to obtain an image quality analysis value. By means of the present invention, the analysis of a super-resolution image is more impartial and accurate than existing image quality analysis methods, thereby providing important application value for the analysis of the quality of super-resolution images obtained after reconstruction of low-resolution images in the field of monitoring, satellite imaging and medical imaging.
Description
本发明涉及数字图像处理及计算机视觉技术领域,尤其涉及的是一种超分辨率图像的图像质量分析方法及***。The invention relates to the technical field of digital image processing and computer vision, and in particular, to an image quality analysis method and system for super-resolution images.
图像超分辨率是图像处理领域中的一个重要研究课题,通过对图像进行超分辨率处理,可以提高图像的分辨率,使得图像更为清晰,其在图像重建中具有较多重要应用,例如:Image super-resolution is an important research topic in the field of image processing. By performing super-resolution processing on the image, the resolution of the image can be increased and the image becomes clearer. It has many important applications in image reconstruction, such as:
(1)在数字电视(DTV)向高清晰度电视(HDTV)过度阶段,仅有部分电视节目会以HDTV的形式播出,不少节目采用的是DTV的形式。因此,可以利用超分辨率重建技术将DTV信号转化为与HDTV接收机相匹配的信号,提高电视节目的兼容性;(1) During the transition from digital television (DTV) to high-definition television (HDTV), only some TV programs will be broadcast in the form of HDTV, and many programs are in the form of DTV. Therefore, the super-resolution reconstruction technology can be used to convert the DTV signal into a signal that matches the HDTV receiver to improve the compatibility of TV programs;
(2)在采集军事与气象遥感图像时,由于受到成像条件与成像***分辨率的限制,不可能获得清晰度很高的图像,而通过利用超分辨率重建技术,在不改变卫星图像探测***的前提下,可实现高于***分辨率的图像观测;(2) When collecting military and meteorological remote sensing images, it is impossible to obtain high-definition images due to the limitations of imaging conditions and the resolution of the imaging system. By using super-resolution reconstruction technology, the satellite image detection system is not changed. Under the premise of the above, it can realize image observation higher than the system resolution;
(3)在医学成像***中(如CT、MRI和超声波仪器等),可以用超分辨率重建技术来提高图像质量,对病变目标进行仔细地检测;(3) In medical imaging systems (such as CT, MRI, and ultrasound instruments, etc.), super-resolution reconstruction technology can be used to improve image quality and carefully detect diseased targets;
(4)在银行、证劵等部门的安全监控***中,当有异常情况发生后,可对监控录像进行超分辨率重建,提高图像要害部分的分辨率,从而为事件的处理提供重要的线索;(4) In the security monitoring system of banks, securities and other departments, when an abnormal situation occurs, super-resolution reconstruction of the monitoring video can be performed to improve the resolution of the key part of the image, thereby providing important clues for the processing of events. ;
(5)可以将超分辨率重建技术用于图像压缩。平时存储或传输低分辨率的图像信息,当有不同需要时,再利用超分辨率重构技术获得不同分辨率的图像和视频。(5) Super-resolution reconstruction technology can be used for image compression. Low-resolution image information is usually stored or transmitted. When there are different needs, super-resolution reconstruction technology is used to obtain images and videos of different resolutions.
随着深度学习的不断发展,图像超分辨率处理技术也在不断前进,已经出现了很多研究成果。然而,对于超分辨图像质量分析仍然面临着巨大挑战。由于人的视觉感知存在差异,对图像质量的主观评价标准参差不齐,亟需一种客观准确的超分辨率图像质量分析方法,建立图像质量的客观评判标准,以此来推动图像 超分辨率技术的不断发展。With the continuous development of deep learning, image super-resolution processing technology is also continuously progressing, and many research results have appeared. However, there are still huge challenges for super-resolution image quality analysis. Due to differences in human visual perception, subjective evaluation standards for image quality are uneven, and an objective and accurate method of super-resolution image quality analysis is urgently needed to establish objective evaluation standards for image quality in order to promote image super-resolution Continuous development of technology.
目前已经出现了许多图像质量分析的方法,如PSNR算法(Peak Signal to Noise Ratio,峰值信噪比),MOS算法(Mean Opinion Scores,主观质量评分)以及SSIM算法(Structural SIMilarity,图像相似度)等。为了分析高光谱图像的质量,还提出了一个多元SSIM,将多光谱像素作为一个多元向量。除了在SSIM中使用的平均值、方差和协方差外,图像质量分析利用到许多其他特征和信息。比如信息保真度准则及视觉信息保真度,将参考图像与失真图像之间的相互信息作为视觉质量计算。There have been many methods of image quality analysis, such as PSNR algorithm (Peak Signal to Noise Ratio), MOS algorithm (Mean opinion scores, subjective quality score), and SSIM algorithm (Structural and SIMilarity, image similarity) . In order to analyze the quality of hyperspectral images, a multivariate SSIM is also proposed, which uses multispectral pixels as a multivariate vector. In addition to the mean, variance, and covariance used in SSIM, image quality analysis uses many other features and information. For example, the information fidelity criterion and visual information fidelity, the mutual information between the reference image and the distorted image is calculated as the visual quality.
因为图像结构特征在视觉感观中占主导地位,许多现有的全参考图像质量方法强调图像结构失真的重要性。然而对于超分辨率图像来说,图像细节同样重要。另外,对于一些超分辨率图像里面存在的虚假纹理,现有的图像质量分析方法没有进行考虑。因此现有的这些图像质量分析方法并不适合于超分辨率图像的图像质量分析。Because image structure features dominate the visual perception, many existing full-reference image quality methods emphasize the importance of image structure distortion. However, for super-resolution images, image details are equally important. In addition, the existing image quality analysis methods have not considered the false textures existing in some super-resolution images. Therefore, these existing image quality analysis methods are not suitable for image quality analysis of super-resolution images.
因此,现有技术有待于进一步的改进。Therefore, the prior art needs to be further improved.
发明内容Summary of the Invention
鉴于上述现有技术中的不足之处,本发明的目的在于提供一种超分辨率图像的图像质量分析方法及***,克服现有技术中缺乏针对超分辨率图像的质量分析方法的缺陷。In view of the above-mentioned shortcomings in the prior art, an object of the present invention is to provide an image quality analysis method and system for super-resolution images, which overcomes the shortcomings of the prior art that lack quality analysis methods for super-resolution images.
本发明提供的第一实施例为一种超分辨率图像的质量分析方法,其中,包括步骤:A first embodiment provided by the present invention is a method for quality analysis of a super-resolution image, which includes the following steps:
获取源参考图像和获取与所述源参考图像内容和像素个数均相同的超分辨率图像;Acquiring a source reference image and acquiring a super-resolution image having the same content and number of pixels as the source reference image;
分别分解出所述源参考图像和超分辨率图像的纹理分量,结构分量和高频分量;Decomposing the texture component, structure component and high-frequency component of the source reference image and super-resolution image respectively;
对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数;Performing pooling on the decomposed texture component, structure component, and high-frequency component respectively to obtain three pooling scores corresponding to the three components;
将三个池化分数相融合得到图像质量分析值。The three pooling scores are fused to obtain the image quality analysis value.
可选的,分解出所述源参考图像和超分辨图像纹理分量的步骤包括:Optionally, the step of decomposing the texture components of the source reference image and the super-resolution image includes:
分别计算超分辨率图像和源参考图像中每个像素的统计特征向量,将计算出 的统计特征向量分别与各自的2范数相比,将相比之后,计算得到的比值做内积,得到内积值;Calculate the statistical feature vectors of each pixel in the super-resolution image and the source reference image, compare the calculated statistical feature vectors with their respective 2 norms, and compare the calculated ratios as inner products after comparison. Inner product value
将超分辨率图像和源参考图像的纹理分量分成多个图像块,计算超分辨率图像和源参考图像两个图像中相同大小图像块的方差值,并获取两个方差值中的最大值;The texture components of the super-resolution image and the source reference image are divided into multiple image blocks, and the variance values of the same-sized image blocks in the two images of the super-resolution image and the source reference image are calculated, and the maximum of the two variance values is obtained. value;
将所述内积值、所述方差值中的最大值和预设第一调整因子相结合计算纹理分量。The texture component is calculated by combining the inner product value, the maximum value of the variance value, and a preset first adjustment factor.
可选的,所述分别计算超分辨率图像和源参考图像中每个像素的统计特征向量步骤包括:Optionally, the step of calculating the statistical feature vector of each pixel in the super-resolution image and the source reference image separately includes:
使用方向梯度直方图算法、尺度不变特征匹配算法或者局部二值模式算法获取超分辨率图像和源参考图像中每个像素的统计特征向量。Use the directional gradient histogram algorithm, scale-invariant feature matching algorithm, or local binary mode algorithm to obtain the statistical feature vector of each pixel in the super-resolution image and the source reference image.
可选的,分解出所述源参考图像和超分辨图像结构分量的步骤包括:Optionally, the step of decomposing the structural components of the source reference image and the super-resolution image includes:
将超分辨率图像和源参考图像分成多个图像块;Split the super-resolution image and the source reference image into multiple image blocks;
计算每个图像块的主方向和所述主方向的内积并取绝对值,得到内积值;Calculating an inner product of the main direction of each image block and the main direction and taking an absolute value to obtain an inner product value;
计算超分辨率图像和源参考图像在结构分量中每个像素的归一化的梯度幅度;Calculate the normalized gradient magnitude of each pixel of the super-resolution image and the source reference image in the structural component;
将所述内积值、所述梯度幅度值和预设第二调整因子相结合,得到所述结构分量。Combining the inner product value, the gradient amplitude value, and a preset second adjustment factor to obtain the structural component.
可选的,将所述源参考图像和超分辨图像分解成高频分量的步骤包括:Optionally, the step of decomposing the source reference image and the super-resolution image into high-frequency components includes:
计算源参考图像和超分辨图像的高频能量向量做内积,将内积的值与所述高频能量向量的2范数之比作为高频分量。Calculate the inner product of the high-frequency energy vectors of the source reference image and the super-resolution image, and use the ratio of the value of the inner product to the 2 norm of the high-frequency energy vector as the high-frequency component.
可选的,将所述源参考图像和超分辨图像分解成高频分量的步骤包括:Optionally, the step of decomposing the source reference image and the super-resolution image into high-frequency components includes:
计算源参考图像和超分辨图像的高频能量向量,将所述高频能量向量与所述高频能量向量的2范数之积作为高频分量。The high-frequency energy vectors of the source reference image and the super-resolution image are calculated, and the product of the two norms of the high-frequency energy vector and the high-frequency energy vector is used as a high-frequency component.
可选的,所述高频能量的计算公式为:Optionally, the calculation formula of the high-frequency energy is:
其中,j是像素的位置索引,N(i)是像素i的相邻像素,N
N是相邻像素的数目,s是结构分量,s
σ是s通过一个方差为σ的高斯卷积核卷积得到,s
σ代表了s的低频部分。
Where j is the position index of the pixel, N (i) is the adjacent pixel of pixel i, N N is the number of adjacent pixels, s is the structural component, and s σ is the s through a Gaussian convolution kernel with a variance of σ Product obtained, s σ represents the low frequency part of s.
可选的,所述对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数的步骤还包括:Optionally, the step of pooling the decomposed texture component, structure component, and high-frequency component separately to obtain three pooling scores corresponding to the three components further includes:
对所述纹理分量、结构分量和高频分量进行加权平均。A weighted average is performed on the texture component, the structure component, and the high-frequency component.
可选的,所述将三个池化分数相融合得到图像质量分析值的步骤包括:Optionally, the step of fusing the three pooling scores to obtain an image quality analysis value includes:
所述图像质量分析值如下公式计算得到:The image quality analysis value is calculated by the following formula:
其中,p为图像质量分析值,α>0且β>0,α和β为调整影系数;p
t,p
s,p
h分别为所述纹理分量、结构分量和高频分量的加权平均值。
Among them, p is an image quality analysis value, α> 0 and β> 0, α and β are adjusted shadow coefficients; p t , p s , and p h are weighted average values of the texture component, structure component, and high-frequency component, respectively. .
本发明提供的第二实施例为一种超分辨率图像的质量分析***,其中,包括:A second embodiment provided by the present invention is a quality analysis system for super-resolution images, which includes:
图像信息获取模块,用于获取源参考图像和获取与所述源参考图像的内容和像素个数均相同的超分辨率图像;An image information acquisition module, configured to acquire a source reference image and a super-resolution image having the same content and number of pixels as the source reference image;
图像分量分解模块,用于分别分解出所述源参考图像和超分辨率图像的纹理分量,结构分量和高频分量;An image component decomposition module, configured to decompose the texture component, the structural component, and the high-frequency component of the source reference image and the super-resolution image respectively;
池化处理模块,用于对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数;A pooling processing module, configured to separately pool the decomposed texture component, structure component, and high-frequency component to obtain three pooling scores corresponding to the three components;
分值融合模块,用于将三个池化分数相融合得到图像质量分析值。A score fusion module is used to fuse the three pooling scores to obtain an image quality analysis value.
有益效果,本发明提供了一种超分辨率图像的图像质量分析方法及***,通过获取源参考图像和获取与所述源参考图像内容和像素个数均相同的超分辨率图像;分别分解出所述源参考图像和超分辨率图像的纹理分量,结构分量和高频分量;对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数;将三个池化分数相融合得到图像质量分析值。在超分辨率图像上的测试,本发明比现有的图像质量分析方法更贴近人类视觉评判机制,且客观实验数据优胜于现有的图像质量分析方法。为监控领域、卫星图像以及医学影像中对低分辨图像进行重建后得到的超分辨率图像质量进行分析提供了重要的应用价值。Advantageously, the present invention provides an image quality analysis method and system for a super-resolution image, by acquiring a source reference image and acquiring a super-resolution image with the same content and number of pixels as the source reference image; Texture components, structural components, and high-frequency components of the source reference image and super-resolution image; and pooling the decomposed texture components, structural components, and high-frequency components, respectively, to obtain three corresponding to the three components. Pooling scores; the three pooling scores are combined to obtain image quality analysis values. For testing on super-resolution images, the present invention is closer to the human visual evaluation mechanism than the existing image quality analysis methods, and the objective experimental data is superior to the existing image quality analysis methods. It provides important application value for the analysis of super-resolution image quality after reconstruction of low-resolution images in the surveillance field, satellite images and medical images.
图1是本发明所提供的一种超分辨率图像的图像质量分析方法的步骤流程图;FIG. 1 is a flowchart of steps of an image quality analysis method for a super-resolution image provided by the present invention; FIG.
图2是本发明提供的所述超分辨率图像的图像质量分析***的原理结构示意图。FIG. 2 is a schematic structural diagram of an image quality analysis system of the super-resolution image provided by the present invention.
为使本发明的目的、技术方案及优点更加清楚、明确,以下参照附图并举实施例对本发明进一步详细说明。应当理解,此处所描述的具体实施例仅仅用于解释本发明,并不用于限定本发明。In order to make the objectives, technical solutions, and advantages of the present invention more clear and specific, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
本发明提供的第一实施例为一种超分辨率图像的质量分析方法,如图1所示,包括步骤:The first embodiment provided by the present invention is a method for quality analysis of a super-resolution image. As shown in FIG. 1, the method includes the following steps:
步骤S1、获取源参考图像和获取与所述源参考图像内容和像素个数均相同的超分辨率图像。Step S1: Acquire a source reference image and acquire a super-resolution image with the same content and number of pixels as the source reference image.
源参考图像和超分辨率图像之间具有相同的内容和像素个数,将超分辨率图像与源参考图像相对比,得到该超分辨率图像的质量分析值。因此本步骤中首先得到需要进行比对的超分辨率图像及其源参考图像。The source reference image and the super-resolution image have the same content and the number of pixels. The super-resolution image is compared with the source reference image to obtain the quality analysis value of the super-resolution image. Therefore, in this step, the super-resolution image and its source reference image to be compared are first obtained.
步骤S2、分别分解出所述源参考图像和超分辨率图像的纹理分量,结构分量和高频分量。Step S2: Decompose the texture component, the structure component, and the high-frequency component of the source reference image and the super-resolution image, respectively.
分别分解出源参考图像和超分辨率图像的纹理分量、结构分量和高频分量,依次对两个图像的这三个分量进行比对,根据三个分量的比对结果,得到质量分析结果。The texture components, structural components, and high-frequency components of the source reference image and the super-resolution image are decomposed, and the three components of the two images are compared in turn. According to the comparison results of the three components, the quality analysis results are obtained.
步骤S3、对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数。Step S3: Perform pooling on the resolved texture component, structure component, and high-frequency component respectively to obtain three pooling scores corresponding to the three components.
步骤S4、将三个池化分数相融合得到图像质量分析值。Step S4: Fusion the three pooling scores to obtain an image quality analysis value.
具体的,本步骤中所述将三个池化分数相融合得到图像质量分析值的步骤包括:利用如下所述图像质量分析值如下公式计算得到图像质量分析值:Specifically, the step of fusing the three pooling scores in this step to obtain an image quality analysis value includes: using the image quality analysis value described below to calculate the image quality analysis value:
其中,p为图像质量分析值,α>0且β>0,α和β为调整影系数;p
t,p
s,p
h分别为所述纹理分量、结构分量和高频分量的加权平均值。
Among them, p is an image quality analysis value, α> 0 and β> 0, α and β are adjusted shadow coefficients; p t , p s , and p h are weighted average values of the texture component, structure component, and high-frequency component, respectively. .
当分别从源参考图像和超分辨率图像中分解出纹理分量、结构分量和高频分量后,对三个分量进行池化处理,再将池化分数相融合得到超分辨率图像的质量 分析值。After the texture components, structural components, and high-frequency components are decomposed from the source reference image and the super-resolution image, the three components are pooled, and the pooling scores are combined to obtain the quality analysis value of the super-resolution image. .
上述步骤S2中包括分别将源参考图像和超分辨率图片分解成纹理分量和结构分量,并分别从源参考图像和超分辨率图片中提取出高频分量,分别对各个分量进行图像质量比较。The above step S2 includes decomposing the source reference image and the super-resolution picture into texture components and structural components, respectively, and extracting high-frequency components from the source reference image and the super-resolution picture, respectively, and comparing the image quality of each component.
具体的,分解出所述源参考图像和超分辨图像纹理分量的步骤包括:Specifically, the step of decomposing the texture components of the source reference image and the super-resolution image includes:
分别计算超分辨率图像和源参考图像中每个像素的统计特征向量,将计算出的统计特征向量分别与各自的2范数相比,将相比之后,计算得到的比值做内积,得到内积值;Calculate the statistical feature vectors of each pixel in the super-resolution image and the source reference image, compare the calculated statistical feature vectors with their respective 2 norms, and compare the calculated ratios as inner products after comparison. Inner product value
将超分辨率图像和源参考图像的纹理分量分成多个图像块,计算超分辨率图像和源参考图像两个图像中相同大小图像块的方差值,并获取两个方差值中的最大值;The texture components of the super-resolution image and the source reference image are divided into multiple image blocks, and the variance values of the same-sized image blocks in the two images of the super-resolution image and the source reference image are calculated, and the maximum of the two variance values is obtained. value;
将所述内积值、所述方差值中的最大值和预设第一调整因子相结合计算纹理分量。The texture component is calculated by combining the inner product value, the maximum value of the variance value, and a preset first adjustment factor.
上述分解出图像纹理分量的方法可以通过如下实施例实现The above method for decomposing the image texture components can be implemented by the following embodiments
纹理分量使用统计特征,比如使用SIFT直方图统计特征来描述纹理特征。给定源参考图像r和与其相同内容和相同像素个数的超分辨率图像u,设纹理分量表示为Mt,对图像u和r中的第i个像素有:Texture components use statistical features, such as SIFT histogram statistical features to describe texture features. Given a source reference image r and a super-resolution image u with the same content and the same number of pixels, let the texture component be represented as Mt. For the i-th pixel in the images u and r:
其中||·||
2代表2范数,<·,·>代表内积运算,f
r(i)和f
u(i)分别是超分辨率图像r和源参考图像u在第i个像素上的统计特征向量,K
t是适应性变量由公式(2)得到:
Where || · || 2 represents the 2 norm, and <·, ·> represents the inner product operation, f r (i) and f u (i) are the super-resolution image r and the source reference image u at the i-th pixel, respectively. The statistical feature vector, K t is the adaptive variable obtained by formula (2):
其中t
r(i)和t
u(i)分别是超分辨率图像r和源参考图像u的纹理分量中以第i个像素为中心的图像块,这两个图像块是相同大小。C
t是一个正常量用来调整K
t的范围。var(·)用来表示计算其参数的方差,max(·,·)表示返回其参数的最大值。通过公式(1)和(2)可以计算得到纹理分量M
t,容易证明M
t的范围是0到1。
Where t r (i) and tu (i) are image blocks centered on the i-th pixel in the texture components of the super-resolution image r and the source reference image u, respectively, and the two image blocks are the same size. C t is a normal amount used to adjust the range of K t . var (·) is used to calculate the variance of its parameters, and max (·, ·) is used to return the maximum value of its parameters. The texture components M t can be calculated by formulas (1) and (2), and it is easy to prove that the range of M t is 0 to 1.
此外,所述分别计算超分辨率图像和源参考图像中每个像素的统计特征向量 步骤包括:使用方向梯度直方图算法、尺度不变特征匹配算法或者局部二值模式算法获取超分辨率图像和源参考图像中每个像素的统计特征向量。In addition, the step of calculating the statistical feature vector of each pixel in the super-resolution image and the source reference image includes: using a direction gradient histogram algorithm, a scale-invariant feature matching algorithm, or a local binary pattern algorithm to obtain a super-resolution image and A statistical feature vector for each pixel in the source reference image.
本发明所述方法中计算统计特征来计算纹理分量,实际计算统计特征的方法还可以是HOG算法,HOG算法可以提取3780维的向量,作为计算纹理分量中的统计特征向量。SIFT算法,SIFT算法可以提取128维的向量,作为计算纹理分量中的统计特征向量。LBP算法,使用LBP特征谱的统计直方图作为统计特征向量。还有基于直方图的统计特征等。可以想到的是,所述的HOG算法,SIFT算法,LBP算法,基于直方图统计等方法提取的统计特征向量任意组合,组合向量也可以作为计算纹理分量中的统计特征向量。In the method of the present invention, the statistical features are calculated to calculate the texture components. The method for actually calculating the statistical features may also be a HOG algorithm. The HOG algorithm may extract a 3780-dimensional vector as a statistical feature vector in calculating the texture components. SIFT algorithm, SIFT algorithm can extract 128-dimensional vector as the statistical feature vector in computing texture components. The LBP algorithm uses the statistical histogram of the LBP feature spectrum as the statistical feature vector. There are also statistical features based on histograms. It is conceivable that the HOG algorithm, the SIFT algorithm, the LBP algorithm, and the statistical feature vectors extracted based on histogram statistics and the like are arbitrarily combined, and the combined vector can also be used as a statistical feature vector in calculating the texture components.
进一步的,分解出所述源参考图像和超分辨图像结构分量的步骤包括:Further, the step of decomposing the structural components of the source reference image and the super-resolution image includes:
将超分辨率图像和源参考图像分成多个图像块;Split the super-resolution image and the source reference image into multiple image blocks;
计算每个图像块的主方向和所述主方向的内积并取绝对值,得到内积值;Calculating an inner product of the main direction of each image block and the main direction and taking an absolute value to obtain an inner product value;
计算超分辨率图像和源参考图像在结构分量中每个像素的归一化的梯度幅度;Calculate the normalized gradient magnitude of each pixel of the super-resolution image and the source reference image in the structural component;
将所述内积值、所述梯度幅度值和预设第二调整因子相结合,得到所述结构分量。Combining the inner product value, the gradient amplitude value, and a preset second adjustment factor to obtain the structural component.
具体的,本实施例中以结构分量使用梯度特征进行的分解,其详细内容如下:Specifically, in this embodiment, the decomposition using the gradient feature with the structural component is detailed as follows:
设结构分量表示为M
s,计算图像结构分量中主方向梯度,需要先计算图像中每个图像块的主方向,图像块的主方向由半正定矩阵J得到:
Let the structural component be represented as M s . To calculate the main direction gradient in the image structural component, the main direction of each image block in the image needs to be calculated first. The main direction of the image block is obtained by the semi-definite matrix J:
其中,i是图像块的中心像素,g
x和g
y分别是该图像块中横轴和纵轴的梯度向量。J有两个特征值和两个特征向量,较小的特征值对应的特征向量即是该图像块梯度主方向。
Among them, i is a central pixel of an image block, and g x and g y are gradient vectors of a horizontal axis and a vertical axis in the image block, respectively. J has two eigenvalues and two eigenvectors, and the eigenvector corresponding to the smaller eigenvalue is the main direction of the gradient of the image block.
其中|·|代表返回绝对值。n
r(i)和n
u(i)是归一化的特征向量,它们分别是超分辨率图像r和源参考图像u在结构分量中的第i个像素的主方向。同样M
s的范围也是0到1,公式(4)中K
s定义如下:
Where | · | represents the absolute value returned. n r (i) and n u (i) are normalized feature vectors, which are the main directions of the i-th pixel in the structural component of the super-resolution image r and the source reference image u, respectively. Similarly, the range of M s is 0 to 1. K s is defined in formula (4) as follows:
其中,g
mr(i)和g
mu(i)分别表示超分辨率图像r和源参考图像u在结构分量中的第i个像素的归一化的梯度幅度。C
s是一个正常量用来调整K
s的范围。
Among them, g mr (i) and g mu (i) represent the normalized gradient magnitudes of the i-th pixel in the structural component of the super-resolution image r and the source reference image u, respectively. C s is a normal quantity used to adjust the range of K s .
上述步骤中计算图像块的主方向梯度来计算结构分量,实际计算结构分量还可以通过相对总变差方法从纹理分量中提取得到,使用可迭代的滚动指导滤波器也可以获得结构分量。实际在计算结构分量中主方向梯度的时候,还可以使用canny算子,sobel算子等,并且所述的图像块的大小可以是可变的。In the above steps, the main component gradient of the image block is calculated to calculate the structural component. The actual calculated structural component can also be extracted from the texture component by the relative total variation method, and the structural component can also be obtained using an iterable rolling guidance filter. Actually, when calculating the main direction gradient in the structural component, canny operator, sobel operator, etc. may also be used, and the size of the image block may be variable.
进一步的,分别将所述源参考图像和超分辨图像分解成高频分量的步骤包括:Further, the steps of decomposing the source reference image and the super-resolution image into high-frequency components respectively include:
计算源参考图像和超分辨图像的高频能量向量做内积,将内积的值与所述高频能量向量的2范数之比作为高频分量。Calculate the inner product of the high-frequency energy vectors of the source reference image and the super-resolution image, and use the ratio of the value of the inner product to the 2 norm of the high-frequency energy vector as the high-frequency component.
进一步的,将所述源参考图像和超分辨图像分解成高频分量的步骤包括:Further, the step of decomposing the source reference image and the super-resolution image into high-frequency components includes:
计算源参考图像和超分辨图像的高频能量向量,将所述高频能量向量与所述高频能量向量的2范数之积作为高频分量。The high-frequency energy vectors of the source reference image and the super-resolution image are calculated, and the product of the two norms of the high-frequency energy vector and the high-frequency energy vector is used as a high-frequency component.
下面,以高频分量使用不同频段的特征为实施例,对其分解方法做进一步的解析。The following uses the characteristics of high-frequency components in different frequency bands as examples to further analyze the decomposition method.
设高频分量表示为M
h,高频分量由公式(7)获得。计算高频分量之前要先比较超分辨率图像r和源参考图像u的高频能量,高频能量计算由公式(6)得到。
Let the high-frequency component be represented as M h , and the high-frequency component is obtained by formula (7). Before calculating the high-frequency components, the high-frequency energy of the super-resolution image r and the source reference image u must be compared.
其中j是位置索引,N(i)是像素i的相邻像素,N
N是相邻像素的数目,s是结构分量。s
σ是s通过一个方差为σ的高斯卷积核卷积得到,s
σ代表了s的低频部分。
Where j is the position index, N (i) is the neighboring pixels of pixel i, N N is the number of neighboring pixels, and s is the structural component. s σ is obtained by convolving s with a Gaussian convolution kernel with variance σ, and s σ represents the low frequency part of s.
其中,h
r和h
u分别是超分辨率图像r和源参考图像u由公式(6)计算得到的高频能量。C
h是一个正常量,用以避免分母过小造成的不稳定性。
Among them, h r and h u are the high-frequency energy calculated by formula (6) of the super-resolution image r and the source reference image u , respectively. C h is a normal quantity to avoid instability caused by too small denominator.
上述方法中使用高斯卷积核来计算高频分量,实际计算高频分量还可以使用 不同频段特征的方法。如下高频能量计算公式:In the above method, a Gaussian convolution kernel is used to calculate high-frequency components. In practice, high-frequency components can also be calculated using different frequency band characteristics. The following formula for calculating high-frequency energy:
其中h(i)表示第i个像素的高频能量,它是一个向量,k可取(2,3,4...),不同的高斯卷积核卷积相减得到不同频段的频率分量,其组合而成的向量作为高频能量h。Where h (i) represents the high-frequency energy of the i-th pixel, which is a vector, k can be taken as (2, 3, 4 ...), and different Gaussian convolution kernels are subtracted to obtain frequency components in different frequency bands. The combined vector is used as the high-frequency energy h.
高频能量h是一个向量,则计算高频分量的公式如下:The high-frequency energy h is a vector. The formula for calculating high-frequency components is as follows:
其中2倍的超分辨率图像和源参考图像的高频能量向量做内积,与各自的高频能量向量的2范数之比作为超分辨率图像质量分析的高频分量,C
h是一个正常数,用来避免分母过小产生的不稳定性。实际上高频能量向量还可以取各自的范数后再相乘作为高频能量的特征,可以取其他任意范数情况。
Wherein high frequency energy is 2 times the vector image and the super-resolution image of the reference source the inner product, the ratio of the number 2 norm of each vector as the high frequency energy high-frequency component of the super-resolution image quality analysis, C h is a Normal number, used to avoid instability caused by too small denominator. In fact, the high-frequency energy vectors can also be multiplied by their respective norms as the characteristics of the high-frequency energy, and other arbitrary norms can be taken.
通过上述公式(1)至(7),分解出源参考图像和超分辨率图像的纹理分量、结构分量和高频分量。下面对分解出的三个分量对进行处理。Through the above formulas (1) to (7), the texture components, structural components, and high-frequency components of the source reference image and the super-resolution image are decomposed. The three component pairs that are resolved are processed below.
所述对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数的步骤还包括:The step of pooling the decomposed texture component, structure component, and high-frequency component respectively to obtain three pooling scores corresponding to the three components further includes:
对所述纹理分量、结构分量和高频分量进行加权平均。A weighted average is performed on the texture component, the structure component, and the high-frequency component.
具体的,对三个分量进行处理的步骤包括:Specifically, the steps for processing the three components include:
需要将这三个分量分别进行池化,池化后得到的分数最终融合得到对超分辨率图像r和源参考图像u的分析分数。These three components need to be pooled separately, and the scores obtained after pooling are finally fused to obtain the analysis scores for the super-resolution image r and the source reference image u.
池化纹理分量、结构分量和高频分量需要加权平均,加权平均公式如下:The pooled texture component, structure component and high-frequency component need weighted average. The weighted average formula is as follows:
其中,N是图像像素的数目,q∈{t,s,h}分别表示纹理分量、结构分量和高频分量的索引,p
q代表三个分量的池化分数,w
q是每个分量里每个像素的权重,w
q由公式(9)至(11)计算后的结果得到。
Among them, N is the number of image pixels, q ∈ {t, s, h} respectively represents the index of the texture component, the structural component and the high-frequency component, p q represents the pooling score of the three components, and w q is each component The weight of each pixel, w q is obtained from the results calculated by formulas (9) to (11).
其中,t
r(i)和t
u(i)与公式(2)相同计算,g
mr(i)和g
mu(i)与公式(5)相同计算,h
r(i)和h
u(i)与公式(6)相同计算。公式(9)至(11)中分母的设置是为了归一化。
Among them, t r (i) and t u (i) are calculated in the same way as in formula (2), g mr (i) and g mu (i) are calculated in the same way as in formula (5), h r (i) and h u (i ) Is calculated in the same way as formula (6). The denominators in equations (9) to (11) are set for normalization.
最后融合三个分量池化后的分数得到p,p由公式(10)计算得到:Finally, the pooled scores of the three components are used to obtain p, which is calculated by formula (10):
其中,α>0且β>0,α和β是为了调整不同分量在融合过程中的影响程度。p
t,p
s,p
h分别由公式(8)计算得到。通过比较结构分量和纹理分量的平均强度可以获得β的值,如公式(13):
Among them, α> 0 and β> 0, α and β are for adjusting the influence degree of different components in the fusion process. p t , p s , p h are calculated by formula (8), respectively. The value of β can be obtained by comparing the average intensity of the structural component and the texture component, as shown in formula (13):
其中,s和t分别是图像结构分量和纹理分量的强度。mean(·)为求均值函数,log(·)为取对数函数。β的值大于1,在获得β时使用了Weber-Fechner法则,也可以使用穷举法获得β。Among them, s and t are the intensity of the image structural component and the texture component, respectively. mean (·) is the mean function and log (·) is the logarithmic function. The value of β is greater than 1. The Weber-Fechner rule is used to obtain β, and β can also be obtained using an exhaustive method.
至此,我们已经得到了超分辨图像u相对于源参考图像r的图像质量分析的分数p。So far, we have obtained the score p of the image quality analysis of the super-resolution image u relative to the source reference image r.
具体的,在上述方法的基础上,对上述方法还可以有一下改进或者替换方法:Specifically, on the basis of the above method, the method can be improved or replaced as follows:
本发明在使用结构纹理分解时,使用了TV-L
1模型优化公式:
When the present invention uses structural texture decomposition, the TV-L 1 model optimization formula is used:
其中λ是拉格朗日乘子系数,u是纹理分量,BV是巴拿赫空间,Ω是图片的一个区域,f是一张图片。实际上还可以使用Meyer,Vese-Osher等其他优化模型。Where λ is the Lagrangian multiplier coefficient, u is the texture component, BV is the Banach space, Ω is a region of the picture, and f is a picture. In fact, other optimization models such as Meyer, Vese-Osher can also be used.
本发明中计算适应性变量K
t是实际上还可以是如下三种公式:
The calculation of the adaptive variable K t in the present invention may actually be the following three formulas:
其中,mean(.)表示求均值函数。Among them, mean (.) Represents the mean function.
其中,|.|表示取绝对值。Among them, |. | Means taking the absolute value.
其中,||.||表示取2范数。实际上还可以是其他范数情况。Among them, ||. || means taking 2 norms. In fact, it can be other norm cases.
本发明中计算β的公式实际上还可以是如下两种公式:The formula for calculating β in the present invention may actually be the following two formulas:
其中var(.)代表求方差函数。Where var (.) Represents the variance function.
其中sum(.)代表求和函数。计算β实际上还可以是求和函数和求方差函数的结合。较佳的,β使用穷举法计算的值为3.9709。Where sum (.) Represents the summation function. Calculating β can actually be a combination of a sum function and a variance function. Preferably, β is calculated to be 3.9709 using an exhaustive method.
本发明使用结构纹理分解方法,实现了在超分辨率图片上进行图像视觉质量分析。对将源参考图像恢复到超分辨率图像后的质量分析提供了更加客观的分析方法,避免了主观上评价超分辨率图像质量的盲目性。为监控领域、卫星图像以及医学影像中对低分辨图像进行重建后得到的超分辨率图像质量进行分析提供了重要的应用价值。由于超分辨率重建技术在一定条件下,可以克服图像***内在分辨率的限制,提高被处理图像的分辨率,因而在视频、遥感、医学和安全监控等领域具都有十分重要的应用。The present invention uses a structural texture decomposition method to implement image visual quality analysis on super-resolution pictures. It provides a more objective analysis method for the quality analysis after the source reference image is restored to the super-resolution image, and avoids the blindness of subjectively evaluating the quality of the super-resolution image. It provides important application value for the analysis of super-resolution image quality after reconstruction of low-resolution images in the surveillance field, satellite images and medical images. Under certain conditions, the super-resolution reconstruction technology can overcome the limitation of the inherent resolution of the image system and improve the resolution of the processed image. Therefore, it has very important applications in the fields of video, remote sensing, medicine, and security monitoring.
本发明提供的第二实施例为一种超分辨率图像的质量分析***,如图2所示,包括:The second embodiment provided by the present invention is a quality analysis system for super-resolution images, as shown in FIG. 2, including:
图像信息获取模块310,用于获取源参考图像和获取与所述源参考图像的内容和像素个数均相同的超分辨率图像;其功能如步骤S1所述。The image information acquisition module 310 is configured to acquire a source reference image and a super-resolution image having the same content and number of pixels as the source reference image; its function is as described in step S1.
图像分量分解模块320,用于分别分解出所述源参考图像和超分辨率图像的纹理分量,结构分量和高频分量;其功能如步骤S2所述。The image component decomposition module 320 is configured to decompose the texture component, the structure component, and the high-frequency component of the source reference image and the super-resolution image, respectively, and its function is as described in step S2.
池化处理模块330,用于对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数;其功能如步骤S3所述。The pooling processing module 330 is configured to pool the resolved texture component, structure component, and high-frequency component respectively to obtain three pooling scores corresponding to the three components; its function is as described in step S3.
分值融合模块340,用于将三个池化分数相融合得到图像质量分析值。其功能如步骤S4所述。The score fusion module 340 is configured to fuse the three pooling scores to obtain an image quality analysis value. Its function is as described in step S4.
本发明所公开的方法及***中不同方式获得纹理分量、结构分量、高频分量可任意组合。In the method and system disclosed in the present invention, texture components, structural components, and high-frequency components obtained in different ways can be arbitrarily combined.
本发明提供了一种超分辨率图像的图像质量分析方法及***,通过获取源参考图像和获取与所述源参考图像内容和像素个数均相同的超分辨率图像;分别分解出所述源参考图像和超分辨率图像的纹理分量,结构分量和高频分量;对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数;将三个池化分数相融合得到图像质量分析值。在超分辨率图像上的测试,本发明比现有的图像质量分析方法更贴近人类视觉评判机制,且客观实验数据优胜于现有的图像质量分析方法。本发明可以应用于其他上采样图像以及现有的许多图像数据库。The present invention provides a method and system for image quality analysis of a super-resolution image, by acquiring a source reference image and a super-resolution image having the same content and number of pixels as the source reference image; and decomposing the source separately The texture components, structural components, and high-frequency components of the reference image and the super-resolution image; and the pooled texture components, structural components, and high-frequency components are separately pooled to obtain three pooling corresponding to the three components Score; the three pooling scores are combined to obtain an image quality analysis value. For testing on super-resolution images, the present invention is closer to the human visual evaluation mechanism than the existing image quality analysis methods, and the objective experimental data is superior to the existing image quality analysis methods. The invention can be applied to other up-sampled images and many existing image databases.
可以理解的是,对本领域普通技术人员来说,可以根据本发明的技术方案及其发明构思加以等同替换或改变,而所有这些改变或替换都应属于本发明所附的权利要求的保护范围。It can be understood that for a person of ordinary skill in the art, equivalent replacements or changes can be made according to the technical solution of the present invention and its inventive concept, and all these changes or replacements should fall within the protection scope of the claims attached to the present invention.
Claims (10)
- 一种超分辨率图像的质量分析方法,其特征在于,包括:A method for quality analysis of super-resolution images, comprising:获取源参考图像和获取与所述源参考图像内容和像素个数均相同的超分辨率图像;Acquiring a source reference image and acquiring a super-resolution image having the same content and number of pixels as the source reference image;分别分解出所述源参考图像和超分辨率图像的纹理分量,结构分量和高频分量;Decomposing the texture component, structure component and high-frequency component of the source reference image and super-resolution image respectively;对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数;Performing pooling on the decomposed texture component, structure component, and high-frequency component respectively to obtain three pooling scores corresponding to the three components;将三个池化分数相融合得到图像质量分析值。The three pooling scores are fused to obtain the image quality analysis value.
- 根据权利要求1所述的超分辨率图像的质量分析方法,其特征在于,分解出所述源参考图像和超分辨图像纹理分量的步骤包括:The method for quality analysis of a super-resolution image according to claim 1, wherein the step of decomposing the source reference image and the super-resolution image texture component comprises:分别计算超分辨率图像和源参考图像中每个像素的统计特征向量,将计算出的统计特征向量分别与各自的2范数相比,将相比之后,计算得到的比值做内积,得到内积值;Calculate the statistical feature vectors of each pixel in the super-resolution image and the source reference image, compare the calculated statistical feature vectors with their respective 2 norms, and compare the calculated ratios as inner products after comparison. Inner product value将超分辨率图像和源参考图像的纹理分量分成多个图像块,计算超分辨率图像和源参考图像两个图像中相同大小图像块的方差值,并获取两个方差值中的最大值;The texture components of the super-resolution image and the source reference image are divided into multiple image blocks, and the variance values of the same-sized image blocks in the two images of the super-resolution image and the source reference image are calculated, and the maximum of the two variance values is obtained. value;将所述内积值、所述方差值中的最大值和预设第一调整因子相结合计算纹理分量。The texture component is calculated by combining the inner product value, the maximum value of the variance value, and a preset first adjustment factor.
- 根据权利要求2所述的超分辨率图像的质量分析方法,其特征在于,所述分别计算超分辨率图像和源参考图像中每个像素的统计特征向量步骤包括:The method for quality analysis of a super-resolution image according to claim 2, wherein the step of calculating a statistical feature vector of each pixel in the super-resolution image and the source reference image comprises:使用方向梯度直方图算法、尺度不变特征匹配算法或者局部二值模式算法获取超分辨率图像和源参考图像中每个像素的统计特征向量。Use the directional gradient histogram algorithm, scale-invariant feature matching algorithm, or local binary mode algorithm to obtain the statistical feature vector of each pixel in the super-resolution image and the source reference image.
- 根据权利要求1所述的超分辨率图像的质量分析方法,其特征在于,分解出所述源参考图像和超分辨图像结构分量的步骤包括:The method for quality analysis of a super-resolution image according to claim 1, wherein the step of decomposing the source reference image and the structural component of the super-resolution image comprises:将超分辨率图像和源参考图像分成多个图像块;Split the super-resolution image and the source reference image into multiple image blocks;计算每个图像块的主方向和所述主方向的内积并取绝对值,得到内积值;Calculating an inner product of the main direction of each image block and the main direction and taking an absolute value to obtain an inner product value;计算超分辨率图像和源参考图像在结构分量中每个像素的归一化的梯度幅度;Calculate the normalized gradient magnitude of each pixel of the super-resolution image and the source reference image in the structural component;将所述内积值、所述梯度幅度值和预设第二调整因子相结合,得到所述结构 分量。Combining the inner product value, the gradient amplitude value, and a preset second adjustment factor to obtain the structural component.
- 根据权利要求1所述的超分辨率图像的质量分析方法,其特征在于,将所述源参考图像和超分辨图像分解成高频分量的步骤包括:The method of claim 1, wherein the step of decomposing the source reference image and the super-resolution image into high-frequency components comprises:计算源参考图像和超分辨图像的高频能量向量做内积,将内积的值与所述高频能量向量的2范数之比作为高频分量。Calculate the inner product of the high-frequency energy vectors of the source reference image and the super-resolution image, and use the ratio of the value of the inner product to the 2 norm of the high-frequency energy vector as the high-frequency component.
- 根据权利要求1所述的超分辨率图像的质量分析方法,其特征在于,将所述源参考图像和超分辨图像分解成高频分量的步骤包括:The method of claim 1, wherein the step of decomposing the source reference image and the super-resolution image into high-frequency components comprises:计算源参考图像和超分辨图像的高频能量向量,将所述高频能量向量与所述高频能量向量的2范数之积作为高频分量。The high-frequency energy vectors of the source reference image and the super-resolution image are calculated, and the product of the two norms of the high-frequency energy vector and the high-frequency energy vector is used as a high-frequency component.
- 根据权利要求5或6所述的超分辨率图像的质量分析方法,其特征在于,所述高频能量的计算公式为:The method for quality analysis of a super-resolution image according to claim 5 or 6, wherein the calculation formula of the high-frequency energy is:其中,j是像素的位置索引,N(i)是像素i的相邻像素,N N是相邻像素的数目,s是结构分量,s σ是s通过一个方差为σ的高斯卷积核卷积得到,s σ代表了s的低频部分。 Where j is the position index of the pixel, N (i) is the adjacent pixel of pixel i, N N is the number of adjacent pixels, s is the structural component, and s σ is the s through a Gaussian convolution kernel with a variance of σ Product obtained, s σ represents the low frequency part of s.
- 根据权利要求1所述的超分辨率图像的质量分析方法,其特征在于,所述对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数的步骤还包括:The method for quality analysis of a super-resolution image according to claim 1, characterized in that said decomposing said texture component, structure component and high-frequency component are respectively pooled to obtain three components The three pooling steps also include:对所述纹理分量、结构分量和高频分量进行加权平均。A weighted average is performed on the texture component, the structure component, and the high-frequency component.
- 根据权利要求8所述的超分辨率图像的质量分析方法,其特征在于,The method for quality analysis of a super-resolution image according to claim 8, wherein:所述将三个池化分数相融合得到图像质量分析值的步骤包括:The step of fusing the three pooling scores to obtain an image quality analysis value includes:所述图像质量分析值如下公式计算得到:The image quality analysis value is calculated by the following formula:其中,p为图像质量分析值,α>0且β>0,α和β为调整系数;p t,p s,p h分别为所述纹理分量、结构分量和高频分量的加权平均值。 Among them, p is an image quality analysis value, α> 0 and β> 0, α and β are adjustment coefficients; p t , p s , and p h are weighted average values of the texture component, structure component, and high-frequency component, respectively.
- 一种超分辨率图像的质量分析***,其特征在于,包括:A quality analysis system for a super-resolution image, comprising:图像信息获取模块,用于获取源参考图像和获取与所述源参考图像的内容和像素个数均相同的超分辨率图像;An image information acquisition module, configured to acquire a source reference image and a super-resolution image having the same content and number of pixels as the source reference image;图像分量分解模块,用于分别分解出所述源参考图像和超分辨率图像的纹理分量,结构分量和高频分量;An image component decomposition module, configured to decompose the texture component, the structural component, and the high-frequency component of the source reference image and the super-resolution image respectively;池化处理模块,用于对分解出的所述纹理分量、结构分量和高频分量分别进行池化,得到与三个分量相对应的三个池化分数;A pooling processing module, configured to separately pool the decomposed texture component, structure component, and high-frequency component to obtain three pooling scores corresponding to the three components;分值融合模块,用于将三个池化分数相融合得到图像质量分析值。A score fusion module is used to fuse the three pooling scores to obtain an image quality analysis value.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811140284.9A CN109410177B (en) | 2018-09-28 | 2018-09-28 | Image quality analysis method and system for super-resolution image |
CN201811140284.9 | 2018-09-28 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2020062901A1 true WO2020062901A1 (en) | 2020-04-02 |
Family
ID=65466480
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2019/088478 WO2020062901A1 (en) | 2018-09-28 | 2019-05-27 | Method and system for analyzing image quality of super-resolution image |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN109410177B (en) |
WO (1) | WO2020062901A1 (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111598779A (en) * | 2020-05-14 | 2020-08-28 | Oppo广东移动通信有限公司 | Image super-resolution processing method and device, electronic device and storage medium |
CN113705632A (en) * | 2021-08-11 | 2021-11-26 | 扬州大学 | Rectal cancer MRI image classification method and device, electronic equipment and storage medium |
CN114119593A (en) * | 2021-12-03 | 2022-03-01 | 深圳大学 | Super-resolution image quality evaluation method based on shallow layer and depth structure texture features |
CN116740777A (en) * | 2022-09-28 | 2023-09-12 | 荣耀终端有限公司 | Training method of face quality detection model and related equipment thereof |
CN118115495A (en) * | 2024-04-23 | 2024-05-31 | 武汉大学 | Super-resolution image quality evaluation method and system combined with ordering learning |
Families Citing this family (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109410177B (en) * | 2018-09-28 | 2022-04-01 | 深圳大学 | Image quality analysis method and system for super-resolution image |
CN110555823B (en) * | 2019-04-11 | 2023-04-07 | 江南大学 | Image fusion quality evaluation method based on TVL structure texture decomposition |
CN110211193B (en) * | 2019-05-17 | 2020-08-04 | 山东财经大学 | Three-dimensional CT (computed tomography) interlayer image interpolation restoration and super-resolution processing method and device |
CN110415242B (en) * | 2019-08-02 | 2020-05-19 | 中国人民解放军军事科学院国防科技创新研究院 | Super-resolution magnification evaluation method based on reference image |
CN113298764B (en) * | 2021-05-11 | 2022-06-14 | 合肥富煌君达高科信息技术有限公司 | High-speed camera imaging quality analysis method based on image noise analysis |
CN113436077A (en) * | 2021-07-28 | 2021-09-24 | 山东大学 | Edge-preserving hyperspectral image super-resolution reconstruction method and system |
EP4396767A1 (en) * | 2021-09-30 | 2024-07-10 | Peking University | Systems and methods for image processing |
CN116091747B (en) * | 2023-04-10 | 2023-06-16 | 山东省地质矿产勘查开发局第五地质大队(山东省第五地质矿产勘查院) | Verification method and system for mapping result |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104021523A (en) * | 2014-04-30 | 2014-09-03 | 浙江师范大学 | Novel method for image super-resolution amplification based on edge classification |
CN106709945A (en) * | 2017-01-09 | 2017-05-24 | 方玉明 | Super-resolution image quality evaluation method |
CN109410177A (en) * | 2018-09-28 | 2019-03-01 | 深圳大学 | A kind of image quality analysis method and system of super-resolution image |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2011250013A (en) * | 2010-05-25 | 2011-12-08 | Nippon Telegr & Teleph Corp <Ntt> | Image quality evaluation method, image quality evaluation apparatus, and program |
CN102436655B (en) * | 2011-09-02 | 2014-01-15 | 清华大学 | Super-resolution reconstruction image quality evaluation method based on SVD (singular value decomposition) |
-
2018
- 2018-09-28 CN CN201811140284.9A patent/CN109410177B/en active Active
-
2019
- 2019-05-27 WO PCT/CN2019/088478 patent/WO2020062901A1/en active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104021523A (en) * | 2014-04-30 | 2014-09-03 | 浙江师范大学 | Novel method for image super-resolution amplification based on edge classification |
CN106709945A (en) * | 2017-01-09 | 2017-05-24 | 方玉明 | Super-resolution image quality evaluation method |
CN109410177A (en) * | 2018-09-28 | 2019-03-01 | 深圳大学 | A kind of image quality analysis method and system of super-resolution image |
Non-Patent Citations (2)
Title |
---|
SU, HENG ET AL.: "Survey of Super-resolution Image Reconstruction Methods", ACTA AUTOMATICA SINICA, vol. 39, no. 8, 31 August 2013 (2013-08-31), pages 1202 - 1213, XP055698880, ISSN: 0254-4156 * |
ZHOU WANG: "Image Quality Assessment: From Error Visibility to Structural Similarity", IEEE TRANSACTIONS ON IMAGE PROCESSING, vol. 13, no. 4, 30 April 2004 (2004-04-30), pages 600 - 612, XP011110418, DOI: 10.1109/TIP.2003.819861 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111598779A (en) * | 2020-05-14 | 2020-08-28 | Oppo广东移动通信有限公司 | Image super-resolution processing method and device, electronic device and storage medium |
CN111598779B (en) * | 2020-05-14 | 2023-07-14 | Oppo广东移动通信有限公司 | Image super-resolution processing method and device, electronic equipment and storage medium |
CN113705632A (en) * | 2021-08-11 | 2021-11-26 | 扬州大学 | Rectal cancer MRI image classification method and device, electronic equipment and storage medium |
CN114119593A (en) * | 2021-12-03 | 2022-03-01 | 深圳大学 | Super-resolution image quality evaluation method based on shallow layer and depth structure texture features |
CN114119593B (en) * | 2021-12-03 | 2024-04-02 | 深圳大学 | Super-resolution image quality evaluation method based on texture features of shallow and deep structures |
CN116740777A (en) * | 2022-09-28 | 2023-09-12 | 荣耀终端有限公司 | Training method of face quality detection model and related equipment thereof |
CN118115495A (en) * | 2024-04-23 | 2024-05-31 | 武汉大学 | Super-resolution image quality evaluation method and system combined with ordering learning |
Also Published As
Publication number | Publication date |
---|---|
CN109410177B (en) | 2022-04-01 |
CN109410177A (en) | 2019-03-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2020062901A1 (en) | Method and system for analyzing image quality of super-resolution image | |
Madhusudana et al. | Subjective and objective quality assessment of stitched images for virtual reality | |
Jiang et al. | Single image super-resolution quality assessment: a real-world dataset, subjective studies, and an objective metric | |
Yang | Multimodal medical image fusion through a new DWT based technique | |
CN108389189B (en) | Three-dimensional image quality evaluation method based on dictionary learning | |
CN112184672A (en) | No-reference image quality evaluation method and system | |
CN112001904A (en) | Comprehensive evaluation module and evaluation method for remote sensing image quality definition | |
Moorthy et al. | Visual perception and quality assessment | |
Nizami et al. | Natural scene statistics model independent no-reference image quality assessment using patch based discrete cosine transform | |
Rajevenceltha et al. | A novel approach for image focus measure | |
Wang et al. | Quality assessment for color correction-based stitched images via bi-directional matching | |
Kim et al. | A multidimensional histogram equalization by fitting an isotropic Gaussian mixture to a uniform distribution | |
CN108629771B (en) | A kind of blind evaluation method of picture quality with scale robustness | |
CN108682005B (en) | Semi-reference 3D synthetic image quality evaluation method based on covariance matrix characteristics | |
CN116664462B (en) | Infrared and visible light image fusion method based on MS-DSC and I_CBAM | |
CN108460756A (en) | Based on statistical nature without reference ir image quality evaluating method | |
Petrovic | Multilevel image fusion | |
Lu et al. | A no-reference image quality assessment approach based on steerable pyramid decomposition using natural scene statistics | |
Aishwarya et al. | Contrast Enhanced Multi-Modal Image Fusion using Top hat transform and Sparse Dictionary | |
Storozhilova et al. | Two approaches for noise filtering in 3D medical CT-images | |
Shao et al. | Infrared and visible image fusion based on spatial convolution sparse representation | |
NA et al. | Detection of camouflaged people | |
EP2884427B1 (en) | Method and system for describing an image | |
Ben Amor et al. | Improved performance of quality metrics using saliency map and CSF filter for standard coding H264/AVC | |
Maleki et al. | Spatial quality assessment of pansharpened images based on gray level co-occurrence matrix |
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: 19864152 Country of ref document: EP Kind code of ref document: A1 |
|
DPE2 | Request for preliminary examination filed before expiration of 19th month from priority date (pct application filed from 20040101) | ||
NENP | Non-entry into the national phase |
Ref country code: DE |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 08.07.2021) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 19864152 Country of ref document: EP Kind code of ref document: A1 |