CN110910347A - Image segmentation-based tone mapping image no-reference quality evaluation method - Google Patents
Image segmentation-based tone mapping image no-reference quality evaluation method Download PDFInfo
- Publication number
- CN110910347A CN110910347A CN201910993399.0A CN201910993399A CN110910347A CN 110910347 A CN110910347 A CN 110910347A CN 201910993399 A CN201910993399 A CN 201910993399A CN 110910347 A CN110910347 A CN 110910347A
- Authority
- CN
- China
- Prior art keywords
- image
- feature
- tone
- area
- region
- 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.)
- Granted
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/40—Analysis of texture
- G06T7/41—Analysis of texture based on statistical description of texture
- G06T7/45—Analysis of texture based on statistical description of texture using co-occurrence matrix computation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/90—Determination of colour characteristics
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Quality & Reliability (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a tone mapping image quality evaluation method based on image segmentation, aiming at the characteristic that main distortion types of different areas of a tone mapping image are not the same, the tone mapping image is divided into a complex area and a flat area, texture detail features are extracted from the complex area, chrominance features are extracted from the flat area, and the texture detail features and the chrominance features are also extracted from a global area. Aiming at the characteristic that details of a highlight area and a low-dark area of an image are distorted too much, the image is divided into the highlight area, the low-dark area and other areas, information entropy characteristics are extracted in different areas respectively to represent the distortion degree of the image, then the brightness distribution uniformity degree of the image is judged by taking a threshold value of the highlight area and the low-dark area as a characteristic, characteristic values with good effects when the threshold value is evaluated in different areas are reserved, characteristic values with poor effects are removed, and characteristic redundancy is reduced; the relevance between the obtained objective evaluation result and the subjective perception of human eyes is effectively improved.
Description
Technical Field
The invention relates to an image quality evaluation technology, in particular to a tone mapping image non-reference quality evaluation method based on image segmentation.
Background
The Dynamic Range of a High-Dynamic Range (HDR) image is very High, and compared with a common image, the HDR image can provide more image details and restore a real scene; in recent years, high-dynamic images and imaging technologies thereof have been widely applied in the fields of movie special effects, professional photography, virtual reality, image rendering and the like. However, since professional equipment for acquisition and display is expensive, the application of high-motion images is difficult to popularize in society. In order to enable HDR images to be displayed on conventional Dynamic Range (SDR) display devices, they can be mapped onto SDR, typically by a Tone-mapping (TM) algorithm; in the process, a corresponding image quality degradation may be introduced; an effective quality evaluation tool is essential for obtaining high-quality TM images.
According to the degree of dependence on a reference image, the objective quality evaluation of the image can be divided into 3 types of full reference, half reference and no reference, wherein the method for evaluating the quality without reference has the strongest practicability and has larger research difficulty; representative examples of the full reference methods include 1) Tone-mapped image quality index (TMQI): the method comprises the steps of obtaining structural fidelity on the basis of a Structural Similarity Index (SSIM), then establishing a non-reference naturalness evaluation model, and finally weighting and summing the structural fidelity and the naturalness of an image to obtain a quality score of a TM image; the representative method of the no-reference method is 2) the index quality assessment of tone-mapped images (BTMQI), which is researched from three aspects of information entropy, naturalness and structure, and the maximum characteristic of the HDR image compared with the LDR image is that more detail information is reserved in a highlight area and a low-brightness area, so that the brightness of the TM image is respectively enlarged and reduced by 4 scales, and the information amount reserved in the highlight area, the low-brightness area and the global area of the TM image is judged by using the information entropy of each scale.
However, the currently used tone mapping image quality evaluation method lacks deep research on local areas of an image, especially flat and complex areas of the image and high-brightness and low-dark areas of the image, and lacks an effective measurement method for feature importance to screen and optimize feature values.
Disclosure of Invention
The invention aims to provide a tone mapping image quality evaluation method based on image segmentation, which can effectively improve the correlation between objective evaluation results and subjective perception quality of human eyes.
The technical scheme adopted by the invention for solving the technical problems is as follows: a no-reference tone mapping image quality evaluation method based on image segmentation comprises the following steps:
① selecting 747 ESPL-LIVE image libraries to obtain any tone mapping image, and converting the tone mapping image into gray level image;
② texture-dividing the tone-mapped image, and recording the divided complex region image as GCompFlat area image denoted as Gflat;
③, performing brightness segmentation on the gray level image by using a maximum entropy threshold segmentation method, and marking the segmented middle brightness region as N, the low dark region as D and the high bright region as B;
④ tensor resolution is performed on the tone mapped image, and the first descendant of the core tensor is recorded as G1For the complex area image GCompThe tensor resolution is carried out and the first filial generation of the core tensor is marked as C1;
⑤ pairs G1Calculating to obtain gray level gradient co-occurrence matrix, and calculating G on the gray level gradient co-occurrence matrix1Corresponding 15 common features, respectively small gradient dominance f1Great gradient advantage f2Distribution of gray scaleUniformity f3Gradient distribution nonuniformity f4Energy f5Correlation f6Entropy of gray scale f7Gradient entropy f8Mixed entropy f9Difference moment f10Inverse differential moment f11Mean value of gray level f12Mean value of gradient f13Standard deviation of gray scale f14Gradient standard deviation f15The 15 features are used as feature vectorsIt is shown that,
⑥ calculate 9 features of the tone-mapped image related to the color moments, the first moment of the RGB three channels of the tone-mapped image, the second moment of the RGB three channels of the tone-mapped image, f37-f39Respectively representing the three moments of the RGB three channels of the tone-mapped image and using these 9 featuresIt is shown that,wherein f is31-f33;
For flat area image GflatThese 9 features are also computed and the feature vector is usedIt is shown that,wherein f is40-f42Respectively representing a flat area image GflatFirst moment of the three channels of RGB of (1), f43-f45Respectively representing a flat area image GflatSecond moment of the three channels of RGB of (1), f46-f48Respectively representing a flat area image GflatThe third moment of the RGB three channels of (1);
⑦ the gray level median G of the gray level imageMIDThreshold T for dividing highlight and intermediate luminance regionsBAnd threshold T for low dark and medium brightness regionsDAs a characteristic, is described as Respectively calculating the information entropies of the three areas B, D and N, and recording the information entropies as characteristics
⑧ will be characterized byAndmerge into a feature of 30 dimensions in length And toQuantifying the importance degree of the characteristic value by using a random forest model to obtain a characteristic importance vector [ w ] with the length of 30 dimensions1,..wi,..w30]Wherein w isiRepresents fiAt f1-f30Degree of importance in dimensional features, followed by a comparison of w1And w16,w2And w17,...,w15And w30Size; when w is1>w16When, let k 11, otherwise k1When w is 162>w17When, let k22, otherwise k2When w is 1715>w30When, let k1515, otherwise k1530; finally, let FW be1=FeatG+C(k1),...FWx=FeatG+C(kx),...FW15=FeatG+C(k15) Wherein, FeatG+C(kx) Is represented in30-dimensional feature of (2)xIs characterized in that x is more than or equal to 15 and more than or equal to 1; for G1And in C1The feature with higher importance is screened out from the same feature of the texture mentioned in the section, and the screened texture feature is recorded as
Will be characterized byAndmerge into one feature of 18 dimensions in length And toBy performing feature screening in the same manner as described above, for the same chromaticity feature in the tone-mapped image and the flat-region image, the feature having higher importance is screened out, and the texture feature after screening is recorded as Wherein, FSjJ is more than or equal to 9 and more than or equal to 1 for the screened chromaticity characteristics;
Randomly selecting 600 images from 747 tone-mapped images according to steps ① to stepsIn the same manner, to obtain a feature vector for each tone-mapped imageAs input to the training set, denoted as ITRAIN,M is more than or equal to 600 and more than or equal to 1, and the corresponding subjective score MOS is also used as the input of the training set and is recorded as OTRAIN,OTRAIN=[MOS1,MOS2,...,MOSm]Constructing a model;
the tone-mapped image to be evaluated is converted into a grayscale image and then steps ② through stepsCalculated feature vectorThe output of the model, denoted as MOS, is obtained as the input to the modelTAnd define MOSTA higher value of (d) represents a better quality of the image to be evaluated.
The specific texture segmentation method in step ② is as follows:
② _1a, performing edge extraction on the gray level image by using a canny operator, and recording the image after the edge extraction as a;
② _1b, expanding a to make the image form a connected region as much as possible, recording the expanded image as b,wherein S is a disc with the radius of 1 pixel, and Z is the displacement generated when the expansion element S is translated;
② _1c, edge-filling b with a line segment of 10 pixels in length, and recording the filled image as c;
② _1d, filling the c by using a hole filling algorithm, and marking the filled image as d;
② _1e, removing a region with the area less than 1500 pixels or the length and the width less than 10 in the image d by using a denoising algorithm, and recording the denoised image as e;
② _1f, traversing the pixel points in e, recording the position set of the point with the pixel value of 255, taking the point with the same position in the tone mapping image to form a complex area image, and forming the rest points in the tone mapping image into a flat area image.
The specific brightness dividing method in step ③ is as follows:
③ _1a, calculating the grayscale median of grayscale image by image grayscale histogram, and recording as GMIDMake the gray value greater than GMIDIs denoted as GBIs less than GMIDIs denoted as GDAnd the probability for each gray value is found and is denoted as pi,i=1...255;
③ _1b, at GBCalculating the maximum entropy threshold of gray scale according to the maximum entropy division method, and recording the obtained threshold as TBMake the gray value greater than TBThe area (D) is marked as a highlight area B, and the gray value is smaller than TBIs denoted as a first intermediate luminance region N1;
③ _1c, at GDCalculating the maximum entropy threshold of gray level in the region, and recording the obtained threshold as TDMake the gray value greater than TDIs designated as a second intermediate luminance region N2Make the gray value less than TDIs marked as a low dark region D, and N is1And N2Adding the areas into an area N; in summary, the image is divided into a high brightness region B, a low dark region D, and a middle brightness region N.
The specific tensor decomposition method in step ④ is as follows:
④ _1a, first image ITMExpressed as third order tensorl1And l2Width and height of the image, respectively; then, the third-order tensor is expanded according to the mode 1, the mode 2 and the mode 3 respectively to obtain corresponding matrixes
④ _1b, performing SVD on the matrix to obtain an orthogonal matrix U(1),U(2),U(3)Respectively having a size of l1×l1,l2×l23 × 3; the specific calculation mode is [ U ](i),S(i),V(i)]=SVD(X(i)) (i ═ 1,2,3), where SVD (·) represents the singular value decomposition of the function;
Since the sub-tensors of the core tensor G satisfy the ordering, i.e. G1>G2>G3Here, G will be1Is defined as child one, G2Is defined as the child two, G3Defining as the third generation; according to the order of the core tensor, G1Carries more information and energy, so it is ready for G1And calculating the texture features.
The specific manner of the feature importance measure in step ⑧ is as follows:
⑧ _1a, the Gini index of the node m in each decision tree is calculated as
Where K denotes a feature having K classes, pkRepresenting the proportion of the class k in the node m;
⑧ _1b, recalculating feature XjThe significance of the node m, i.e., the amount of change in Gini index before and after the node m branches, isWherein, GIlAnd GIrRespectively representing Gini indexes of two new nodes after branching;
⑧ _1c, then calculating the importance of the features in a decision tree, the decision tree i will have the feature XjIs denoted as M, then XjThe importance of the ith tree is:
⑧ _1d, finally calculating the feature importance of all decision trees, assuming n trees in random forestAnd finally, performing normalization processing on all the obtained importance scores:
the calculation method of the contrast C in the step ⑨ is as follows:
where g (x, y) represents the pixel value of point (x, y), BRI represents the mean value of the luminance of the image, and M and N represent the length and width of the image, respectively.
Compared with the prior art, the invention has the advantages that: the method considers the difference of main distortion types of tone mapping images in a complex area and a flat area, divides the images into the complex area and the flat area, and extracts different characteristic values in different areas, so that the subsequent quality characteristic extraction is more targeted.
The method provides image characteristics such as gray gradient co-occurrence matrix, color moment and the like which are different from the traditional image quality evaluation, so that the extracted characteristics can reflect the quality degradation degree of the tone mapping image more accurately.
The method fully considers the redundancy among the features, so the random forest model is used for sequencing the importance of the features, and the feature with better effect is selected from the global and local texture and chromaticity features, so that the feature with poor effect is removed, and the feature redundancy is reduced.
The method comprehensively considers the difference between the tone mapping image and the traditional image, and innovations and improvements are carried out from the three angles of region segmentation, quality feature extraction and feature dimension reduction, so that the correlation between the objective evaluation result obtained by the method and the subjective perception of human eyes is effectively improved.
In order to verify the effectiveness of the method for evaluating the quality of the tone mapping image, three evaluation indexes are selected as measures of the quality of the method, wherein the evaluation indexes are respectively Pearson Linear Correlation Coefficient (PLCC), Spearman sequential correlation coefficient (SROCC) and Root Mean Square Error (RMSE) which respectively represent the correlation between the predicted fraction and the actual fraction. PLCC and SROCC have values between (0,1), and the closer to 1 the better, the smaller the RMSE the better.
Taking 147 residual color tone mapping images to be evaluated under ESPL-LIVEFeature vector ofWhen the input is taken as the input of the test set, 747 is more than or equal to l and more than or equal to 601, the MOS value obtained by predicting each image is taken as the output and is recorded as PMOSlThe PMOS values of 147 pictures are taken as a set RF _ output, which is [ PMOS ═ PMOS1,PMOS2,...,PMOSl]Then 147 tone-mapped images to be evaluatedThe MOS value of (1) is recorded as the set output _ test, which is [ MOS ] value1,MOS2,...,MOSl](ii) a The correlation coefficients PLCC, SROCC and RMSE are found using the IQA function, where f is IQA (RF _ output, output _ test). Wherein IQA (: indicates a fitting function, and f indicates PLCC, SROCC and RMSE correlation coefficients.
As can be seen from the data listed in Table 1, the objective quality evaluation predicted value of the tone mapping image calculated by the method of the present invention has good correlation with the average subjective score difference, wherein the PLCC correlation coefficient reaches 0.8252, the SROCC correlation coefficient reaches 0.7795, and the RMSE reaches 5.7833.
TABLE 1 Performance indicators for the correlation between the objective quality evaluation prediction value and the average subjective score difference for tone mapped images in a test image set calculated according to the method of the present invention
Type of index | PLCC | SROCC | RMSE |
End result | 0.8252 | 0.7795 | 5.7833 |
Drawings
FIG. 1 is a block diagram of an overall implementation of the method of the present invention;
FIG. 2 is a gray scale gradient co-occurrence matrix feature table proposed by the method of the present invention.
Detailed Description
The invention is described in further detail below with reference to the embodiments of the drawings;
example (b): a no-reference tone mapping image quality evaluation method based on image segmentation comprises the following steps:
① selecting 747 ESPL-LIVE image libraries to obtain any tone mapping image marked as ITMTone mapping the image ITMConversion into a grayscale image, denoted as IG-TM;
② pairs tone-mapped image ITMPerforming texture segmentation, and recording the segmented complex region image as GCompFlat area image denoted as Gflat(ii) a The method specifically comprises the following steps:
② _1a, performing edge extraction on the gray level image by using a canny operator, and recording the image after the edge extraction as a;
② _1b, expanding a to make the image form a connected region as much as possible, recording the expanded image as b,wherein S is a disc with the radius of 1 pixel, and Z is the displacement generated when the expansion element S is translated;
② _1c, edge-filling b with a line segment of 10 pixels in length, and recording the filled image as c;
② _1d, filling the c by using a hole filling algorithm, and marking the filled image as d;
② _1e, removing a region with the area less than 1500 pixels or the length and the width less than 10 in the image d by using a denoising algorithm, and recording the denoised image as e;
② _1f, traversing the pixel points in e, recording the position set of the point with the pixel value of 255, taking the point with the same position in the tone mapping image to form a complex area image, and forming the rest points in the tone mapping image into a flat area image;
③ pairs of grayscale images IG-TMPerforming brightness segmentation by using a maximum entropy threshold segmentation method, and recording a segmented middle brightness region as N, a low dark region as D and a high bright region as B; the method specifically comprises the following steps:
③ _1a, calculating the grayscale median of grayscale image by image grayscale histogram, and recording as GMIDMake the gray value greater than GMIDIs denoted as GBIs less than GMIDIs denoted as GDAnd the probability for each gray value is found and is denoted as pi,i=1...255;
③ _1b, at GBRegion GBThe regions were determined according to the literature [ Kapur N, Sahoo P, Wong A. "A new method for gray-level picture thresholding using entropy of the histogram,"ComputerVision,Graphics,and Image Processing,vol.29,pp.273-285,1985]Calculating the maximum entropy threshold of gray scale by the maximum entropy division method, and recording the obtained threshold as TBMake the gray value greater than TBThe area (D) is marked as a highlight area B, and the gray value is smaller than TBIs denoted as a first intermediate luminance region N1;
③ _1c, at GDCalculating the maximum entropy threshold of gray level in the region, and recording the obtained threshold as TDMake the gray value greater than TDIs designated as a second intermediate luminance region N2Make the gray value less than TDIs marked as a low dark region D, and N is1And N2Adding the areas into an area N; in conclusion, the image is divided into a high brightness area B, a low dark area D and a middle brightness area N;
④ pairs tone-mapped image ITMCarrying out tensor decomposition, and recording the first filial generation of the core tensor as G1For the complex area image GCompThe tensor resolution is carried out and the first filial generation of the core tensor is marked as C1(ii) a The specific way of tensor decomposition is:
④ _1a, first image ITMExpressed as third order tensorl1And l2Width and height of the image, respectively; then, the third-order tensor is expanded according to the mode 1, the mode 2 and the mode 3 respectively to obtain corresponding matrixes
④ _1b, performing SVD on the matrix to obtain an orthogonal matrix U(1),U(2),U(3)Respectively having a size of l1×l1,l2×l23 × 3; the specific calculation mode is [ U ](i),S(i),V(i)]=SVD(X(i)) (i ═ 1,2,3), where SVD (·) represents the singular value decomposition of the function;
Since the sub-tensors of the core tensor G satisfy the ordering, i.e. G1>G2>G3Here, G will be1Is defined as child one, G2Is defined as the child two, G3Defining as the third generation; according to the order of the core tensor, G1Carries more information and energy, so it is ready for G1Calculating the texture features;
⑤ pairs G1Calculating to obtain gray level gradient co-occurrence matrix, and calculating G on the gray level gradient co-occurrence matrix1Corresponding 15 common features, respectively small gradient dominance f1Great gradient advantage f2Non-uniformity of gray distribution f3Gradient distribution nonuniformity f4Energy f5Correlation f6Entropy of gray scale f7Gradient entropy f8Mixed entropy f9Difference moment f10Inverse differential moment f11Mean value of gray level f12Mean value of gradient f13Standard deviation of gray scale f14Gradient standard deviation f15The specific calculation method is described in the literature [ Hongtong light, gray level-gradient co-occurrence matrix texture analysis method [ J]Journal of Automation, 1984(01):22-25.]The 15 features are used as feature vectorsIt is shown that,
⑥ calculating tone mapped image ITM9 features relating to the color moments, respectively tone-mapped image ITMRGB first moment, tone-mapped image I of three channelsTMSecond moment of the three channels of RGB of (1), f37-f39Respectively representing tone-mapped images ITMAnd use these 9 features as the third moment of the RGB three channelsIt is shown that,wherein f is31-f33;
For flat area image GflatThese 9 features are also computed and the feature vector is usedIt is shown that,wherein f is40-f42Respectively representing a flat area image GflatFirst moment of the three channels of RGB of (1), f43-f45Respectively representing a flat area image GflatSecond moment of the three channels of RGB of (1), f46-f48Respectively representing a flat area image GflatThe third moment of the RGB three channels of (1);
⑦ grayscale image IG-TMGray median value G ofMIDThreshold T for dividing highlight and intermediate luminance regionsBAnd threshold T for low dark and medium brightness regionsDAs a characteristic, is described as Are respectively provided withCalculating the information entropy of the three areas B, D and N, and recording as the characteristics
⑧ will be characterized byAndmerge into a feature of 30 dimensions in length And toUsing a random forest model to measure the feature importance to obtain a feature importance vector [ w ] with the length of 30 dimensions1,..wi,..w30]Wherein w isiRepresents fiAt f1-f30Degree of importance in dimensional features, followed by a comparison of w1And w16,w2And w17,...,w15And w30Size; when w is1>w16When, let k 11, otherwise k1When w is 162>w17When, let k22, otherwise k2When w is 1715>w30When, let k1515, otherwise k1530; finally, let FW be1=FeatG+C(k1),...FWx=FeatG+C(kx),...FW15=FeatG+C(k15) Wherein, FeatG+C(kx) Is represented in30-dimensional feature of (2)xIs characterized in that x is more than or equal to 15 and more than or equal to 1; for G1And in C1The feature with higher importance is screened out from the same feature of the texture mentioned in the section, and the screened texture feature is recorded as
Will be characterized byAndmerge into one feature of 18 dimensions in length And toImage I is tone-mapped by feature screening in the same manner as described aboveTMAnd in flat area image GflatThe feature with higher importance is screened out from the same chromaticity feature in the image, and the screened texture feature is recorded as the texture feature Wherein, FSjJ is more than or equal to 9 and more than or equal to 1 for the screened chromaticity characteristics; the specific way of the feature importance measure is as follows:
⑧ _1a, the Gini index of the node m in each decision tree is calculated as
Where K denotes a feature having K classes, pkRepresenting the proportion of the class k in the node m;
⑧ _1b, recalculating feature XjThe significance of the node m, i.e., the amount of change in Gini index before and after the node m branches, isWherein, GIlAnd GIrRespectively representing Gini indexes of two new nodes after branching;
⑧ _1c, then calculating the importance of the features in a decision tree, the decision tree i will have the feature XjIs denoted as M, then XjThe importance of the ith tree is:
⑧ _1d, finally calculating the feature importance of all decision trees, assuming n trees in random forestAnd finally, performing normalization processing on all the obtained importance scores:
⑨ calculating tone mapped image ITMContrast of (2), which is taken as a featureThe contrast C is calculated in the following manner:
wherein g (x, y) represents a pixel value of the point (x, y), BRI represents a luminance mean value of the image, and M and N represent a length and a width of the image, respectively;
Tone mapping image I at 747 sheets randomlyTMSelecting 600 images according to the steps ① to ①In the same manner to obtain each tone-mapped image ITMFeature vector ofAs input to the training set, denoted as ITRAIN,M is more than or equal to 600 and more than or equal to 1, and the corresponding subjective score MOS is also used as the input of the training set and is recorded as OTRAIN,OTRAIN=[MOS1,MOS2,...,MOSm]Constructing a model;
the tone-mapped image to be evaluated is converted into a grayscale image and then steps ② through stepsCalculated feature vectorThe output of the model, denoted as MOS, is obtained as the input to the modelTAnd define MOSTA higher value of (d) represents a better quality of the image to be evaluated.
Claims (6)
1. A no-reference tone mapping image quality evaluation method based on image segmentation is characterized by comprising the following steps:
① selecting 747 ESPL-LIVE image libraries to obtain any tone mapping image, and converting the tone mapping image into gray level image;
② texture-dividing the tone-mapped image, and recording the divided complex region image as GCompFlat area image denoted as Gflat;
③, performing brightness segmentation on the gray level image by using a maximum entropy threshold segmentation method, and marking the segmented middle brightness region as N, the low dark region as D and the high bright region as B;
④ tensor resolution is performed on the tone mapped image, and the first descendant of the core tensor is recorded as G1For the complex area image GCompThe tensor resolution is carried out and the first filial generation of the core tensor is marked as C1;
⑤ pairs G1Calculating to obtain gray level gradient co-occurrence matrix, and calculating G on the gray level gradient co-occurrence matrix1Corresponding 15 common features, respectively small gradient dominance f1Great gradient advantage f2Non-uniformity of gray distribution f3Gradient distribution nonuniformity f4Energy f5Correlation f6Entropy of gray scale f7Gradient entropy f8Mixed entropy f9Difference moment f10Inverse differential moment f11Mean value of gray level f12Mean value of gradient f13Standard deviation of gray scale f14Gradient standard deviation f15The 15 features are used as feature vectorsIt is shown that,
⑥ calculate 9 features of the tone-mapped image related to the color moments, the first moment of the RGB three channels of the tone-mapped image, the second moment of the RGB three channels of the tone-mapped image, f37-f39Respectively representing the three moments of the RGB three channels of the tone-mapped image and using these 9 featuresIt is shown that,wherein f is31-f33;
For flat area image GflatThese 9 features are also computed and the feature vector is usedIt is shown that,wherein f is40-f42Respectively representing a flat area image GflatFirst moment of the three channels of RGB of (1), f43-f45Respectively representing a flat area image GflatSecond moment of the three channels of RGB of (1), f46-f48Respectively representing a flat area image GflatThe third moment of the RGB three channels of (1);
⑦ the gray level median G of the gray level imageMIDThreshold T for dividing highlight and intermediate luminance regionsBAnd threshold T for low dark and medium brightness regionsDAs a characteristic, is described as Respectively calculating the information entropies of the three areas B, D and N, and recording the information entropies as characteristics
⑧ will be characterized byAndmerge into a feature of 30 dimensions in length And toQuantifying the importance degree of the characteristic value by using a random forest model to obtain a characteristic importance vector [ w ] with the length of 30 dimensions1,..wi,..w30]Wherein w isiRepresents fiAt f1-f30Degree of importance in dimensional features, followed by a comparison of w1And w16,w2And w17,...,w15And w30Size; when w is1>w16When, let k11, otherwise k1When w is 162>w17When, let k22, otherwise k2When w is 1715>w30When, let k1515, otherwise k1530; finally, let FW be1=FeatG+C(k1),...FWx=FeatG+C(kx),...FW15=FeatG+C(k15) Wherein, FeatG+C(kx) Is represented in30-dimensional feature of (2)xIs characterized in that x is more than or equal to 15 and more than or equal to 1; for G1And in C1The feature with higher importance is screened out from the same feature of the texture mentioned in the section, and the screened texture feature is recorded as
Will be characterized byAndmerge into one feature of 18 dimensions in length And toBy performing feature screening in the same manner as described above, for the same chromaticity feature in the tone-mapped image and the flat-region image, the feature having higher importance is screened out, and the texture feature after screening is recorded as Wherein, FSjJ is more than or equal to 9 and more than or equal to 1 for the screened chromaticity characteristics;
Randomly selecting 600 images from 747 tone-mapped images according to steps ① to stepsIn the same manner, to obtain a feature vector for each tone-mapped imageAs input to the training set, denoted as ITRAIN,M is more than or equal to 600 and more than or equal to 1, and the corresponding subjective score MOS is also used as the input of the training set and is recorded as OTRAIN,OTRAIN=[MOS1,MOS2,...,MOSm]Constructing a model;
the tone-mapped image to be evaluated is converted into a grayscale image and then steps ② through stepsCalculated feature vectorThe output of the model, denoted as MOS, is obtained as the input to the modelTAnd define MOSTA higher value of (d) represents a better quality of the image to be evaluated.
2. The method according to claim 1, wherein the texture segmentation method in step ② is as follows:
② _1a, performing edge extraction on the gray level image by using a canny operator, and recording the image after the edge extraction as a;
② _1b, expanding a to make the image form a connected region as much as possible, recording the expanded image as b,wherein S is a disc with the radius of 1 pixel, and Z is the displacement generated when the expansion element S is translated;
② _1c, edge-filling b with a line segment of 10 pixels in length, and recording the filled image as c;
② _1d, filling the c by using a hole filling algorithm, and marking the filled image as d;
② _1e, removing a region with the area less than 1500 pixels or the length and the width less than 10 in the image d by using a denoising algorithm, and recording the denoised image as e;
② _1f, traversing the pixel points in e, recording the position set of the point with the pixel value of 255, taking the point with the same position in the tone mapping image to form a complex area image, and forming the rest points in the tone mapping image into a flat area image.
3. The method according to claim 1, wherein the specific brightness division manner in step ③ is as follows:
③ _1a, calculating the grayscale median of grayscale image by image grayscale histogram, and recording as GMIDMake the gray value greater than GMIDIs denoted as GBIs less than GMIDIs denoted as GDAnd the probability for each gray value is found and is denoted as pi,i=1...255;
③ _1b, at GBCalculating the maximum entropy threshold of gray scale according to the maximum entropy division method, and recording the obtained threshold as TBMake the gray value greater than TBThe area (D) is marked as a highlight area B, and the gray value is smaller than TBIs denoted as a first intermediate luminance region N1;
③ _1c, at GDCalculating the maximum entropy threshold of gray level in the region, and recording the obtained threshold as TDMake the gray value greater than TDIs designated as a second intermediate luminance region N2Make the gray value less than TDIs marked as a low dark region D, and N is1And N2Adding the areas into an area N; in summary, the image is divided into a high brightness region B, a low dark region D, and a middle brightness region N.
4. The method for evaluating the quality of a tone-mapped image based on exposure analysis according to claim 1 or 2, wherein the specific tensor resolution in step ④ is as follows:
④ _1a, first image ITMExpressed as third order tensorl1And l2Width and height of the image, respectively; then, the third-order tensor is expanded according to the mode 1, the mode 2 and the mode 3 respectively to obtain corresponding matrixes
④ _1b, performing SVD on the matrix to obtain an orthogonal matrix U(1),U(2),U(3)Size division ofIs other than1×l1,l2×l23 × 3; the specific calculation mode is [ U ](i),S(i),V(i)]=SVD(X(i)) (i ═ 1,2,3), where SVD (·) represents the singular value decomposition of the function;
Since the sub-tensors of the core tensor G satisfy the ordering, i.e. G1>G2>G3Here, G will be1Is defined as child one, G2Is defined as the child two, G3Defining as the third generation; according to the order of the core tensor, G1Carries more information and energy, so it is ready for G1And calculating the texture features.
5. The method of claim 1, wherein the measure of importance of the features in step ⑧ is determined by:
⑧ _1a, the Gini index of the node m in each decision tree is calculated as
Where K denotes a feature having K classes, pkRepresenting the proportion of the class k in the node m;
⑧ _1b, recalculating feature XjThe significance of the node m, i.e., the amount of change in Gini index before and after branching of the node m, is VIMjgmini=GIm-GIl-GIrWherein, GIlAnd GIrRespectively representing Gini indexes of two new nodes after branching;
⑧_1cthen calculating the feature importance of a decision tree, and enabling the decision tree i to have the feature XjIs denoted as M, then XjThe importance of the ith tree is:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910993399.0A CN110910347B (en) | 2019-10-18 | 2019-10-18 | Tone mapping image non-reference quality evaluation method based on image segmentation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910993399.0A CN110910347B (en) | 2019-10-18 | 2019-10-18 | Tone mapping image non-reference quality evaluation method based on image segmentation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110910347A true CN110910347A (en) | 2020-03-24 |
CN110910347B CN110910347B (en) | 2023-06-06 |
Family
ID=69814601
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910993399.0A Active CN110910347B (en) | 2019-10-18 | 2019-10-18 | Tone mapping image non-reference quality evaluation method based on image segmentation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110910347B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113362354A (en) * | 2021-05-07 | 2021-09-07 | 安徽国际商务职业学院 | Method, system, terminal and storage medium for evaluating quality of tone mapping image |
WO2021223540A1 (en) * | 2020-05-08 | 2021-11-11 | 华为技术有限公司 | Processing method, encoding device, and decoding device for high dynamic range (hdr) video |
CN116894851A (en) * | 2023-09-11 | 2023-10-17 | 吉林农业大学 | Soil CT image pore extraction method and system |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120201456A1 (en) * | 2009-10-08 | 2012-08-09 | International Business Machines Corporation | Transforming a digital image from a low dynamic range (ldr) image to a high dynamic range (hdr) image |
CN103353982A (en) * | 2013-05-15 | 2013-10-16 | 中山大学 | Method for tone mapping based on histogram equalization |
CN108010024A (en) * | 2017-12-11 | 2018-05-08 | 宁波大学 | It is a kind of blind with reference to tone mapping graph image quality evaluation method |
-
2019
- 2019-10-18 CN CN201910993399.0A patent/CN110910347B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120201456A1 (en) * | 2009-10-08 | 2012-08-09 | International Business Machines Corporation | Transforming a digital image from a low dynamic range (ldr) image to a high dynamic range (hdr) image |
CN103353982A (en) * | 2013-05-15 | 2013-10-16 | 中山大学 | Method for tone mapping based on histogram equalization |
CN108010024A (en) * | 2017-12-11 | 2018-05-08 | 宁波大学 | It is a kind of blind with reference to tone mapping graph image quality evaluation method |
Non-Patent Citations (2)
Title |
---|
侯春萍;李浩;岳广辉: "局部和全局特征融合的色调映射图像质量评价" * |
刘佳;郁梅;姜浩;马华林;: "结合美学特征的无参考色调映射图像质量评价" * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2021223540A1 (en) * | 2020-05-08 | 2021-11-11 | 华为技术有限公司 | Processing method, encoding device, and decoding device for high dynamic range (hdr) video |
CN113362354A (en) * | 2021-05-07 | 2021-09-07 | 安徽国际商务职业学院 | Method, system, terminal and storage medium for evaluating quality of tone mapping image |
CN113362354B (en) * | 2021-05-07 | 2024-04-30 | 安徽国际商务职业学院 | Quality evaluation method, system, terminal and storage medium for tone mapping image |
CN116894851A (en) * | 2023-09-11 | 2023-10-17 | 吉林农业大学 | Soil CT image pore extraction method and system |
CN116894851B (en) * | 2023-09-11 | 2023-11-14 | 吉林农业大学 | Soil CT image pore extraction method and system |
Also Published As
Publication number | Publication date |
---|---|
CN110910347B (en) | 2023-06-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110046673B (en) | No-reference tone mapping image quality evaluation method based on multi-feature fusion | |
Wang et al. | Simple low-light image enhancement based on Weber–Fechner law in logarithmic space | |
CN101853504B (en) | Image quality evaluating method based on visual character and structural similarity (SSIM) | |
CN109978854B (en) | Screen content image quality evaluation method based on edge and structural features | |
CN112381897B (en) | Low-illumination image enhancement method based on self-coding network structure | |
Tian et al. | A multi-order derivative feature-based quality assessment model for light field image | |
CN110910347A (en) | Image segmentation-based tone mapping image no-reference quality evaluation method | |
CN112950596B (en) | Tone mapping omnidirectional image quality evaluation method based on multiple areas and multiple levels | |
CN113822830B (en) | Multi-exposure image fusion method based on depth perception enhancement | |
CN105513033B (en) | A kind of super resolution ratio reconstruction method that non local joint sparse indicates | |
CN109218716B (en) | No-reference tone mapping image quality evaluation method based on color statistics and information entropy | |
CN110706196B (en) | Clustering perception-based no-reference tone mapping image quality evaluation algorithm | |
Chen et al. | Blind quality index for tone-mapped images based on luminance partition | |
Steffens et al. | Cnn based image restoration: Adjusting ill-exposed srgb images in post-processing | |
Wang et al. | Low-light image enhancement based on virtual exposure | |
CN107146220A (en) | A kind of universal non-reference picture quality appraisement method | |
CN104346809A (en) | Image quality evaluation method for image quality dataset adopting high dynamic range | |
Feng et al. | Low-light image enhancement algorithm based on an atmospheric physical model | |
Xu et al. | Remote-sensing image usability assessment based on ResNet by combining edge and texture maps | |
CN113409247B (en) | Multi-exposure fusion image quality evaluation method | |
Zhao et al. | A multi-scale U-shaped attention network-based GAN method for single image dehazing | |
CN111047618A (en) | Multi-scale-based non-reference screen content image quality evaluation method | |
Dwivedi et al. | Single image dehazing using extended local dark channel prior | |
CN112132774A (en) | Quality evaluation method of tone mapping image | |
Zheng et al. | Overwater image dehazing via cycle-consistent generative adversarial network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |