CN109829924A - A kind of image quality evaluating method based on body feature analysis - Google Patents
A kind of image quality evaluating method based on body feature analysis Download PDFInfo
- Publication number
- CN109829924A CN109829924A CN201910046769.XA CN201910046769A CN109829924A CN 109829924 A CN109829924 A CN 109829924A CN 201910046769 A CN201910046769 A CN 201910046769A CN 109829924 A CN109829924 A CN 109829924A
- Authority
- CN
- China
- Prior art keywords
- image
- value
- feature
- follows
- lab
- 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
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000004458 analytical method Methods 0.000 title claims abstract description 12
- 230000009466 transformation Effects 0.000 claims description 20
- 210000000746 body region Anatomy 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 16
- 238000001514 detection method Methods 0.000 claims description 15
- 238000012549 training Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 10
- 238000001228 spectrum Methods 0.000 claims description 9
- 238000000605 extraction Methods 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 239000000284 extract Substances 0.000 claims description 5
- 238000001303 quality assessment method Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000013441 quality evaluation Methods 0.000 abstract description 14
- 230000000694 effects Effects 0.000 abstract description 4
- 230000007812 deficiency Effects 0.000 abstract 1
- 238000011156 evaluation Methods 0.000 description 8
- 230000006870 function Effects 0.000 description 5
- 238000002474 experimental method Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000007792 addition Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000013016 learning Effects 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000000697 sensory organ Anatomy 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Abstract
The invention discloses a kind of image quality evaluating methods based on body feature analysis, it solves conventional images quality evaluating method and deficiency, which leads to problems such as image quality evaluation accuracy rate not high, to be considered to image subject feature, this method mainly carries out image quality evaluation from body features and image blur global characteristics etc. such as the depth of field, space configuration feature, background color complexity, the Lab features of main body, and effect is better than general universality method.In addition, the present invention has chosen the method based on sparse representation as classifier, it is more preferable than traditional classifier effect.
Description
Technical field
The invention belongs to image aesthetic qualities to evaluate field, specifically a kind of image matter based on body feature analysis
Measure evaluation method.
Background technique
Image quality evaluation is the predecessor of image aesthetic quality evaluation, it simulates human visual system objectively to image
Quality is evaluated, and traditional images quality evaluation is primarily used to whether detection image has distortion in transmission process, such as: mould
Paste, distortion, noise etc..And the evaluation of image aesthetic quality is then that the aesthetic feeling degree of image is evaluated from sense organ, it is in computer
With the crossing domain of the subjects such as psychology, design science, the evaluation index being related to is more abstract, such as: image depth, main body structure
Figure, degree harmonious in colour, scene, shadow etc..
Datta in 2005 has delivered the article for calculating aesthetics in ECCV meeting, causes the extensive concern of industry,
In subsequent 10 years, concern image aesthetic quality pricer is more and more, and the article of image aesthetic quality evaluation is major
Accounting is also higher and higher in meeting, and due to lacking the database of aesthetic evaluation, early stage people only carry out two classification to image,
After AVA database is born, image is classified more also gradually to be studied with the problem of image recurrence.Until 2015, image was aesthstic
Quality evaluation encounters bottleneck, and accuracy rate cannot be obviously improved, and with the rise into 2 years deep learnings, image aesthetic quality is commented
Valence becomes popular domain again.
Recently as the fast development of artificial intelligence, in image retrieval, streetscape navigation, U.S. face, advertisement design, media etc.
The status in field, image aesthetic evaluation is also more and more important, and the core algorithm of many application demands is required with image aesthetics
Evaluation.Currently, the application of image aesthetic quality evaluation is still in infancy, there has also been some relatively successful stories, such as:
Online Video website: iqiyi.com can be automatically extracted out in video using image aesthetic quality evaluation algorithms most suitable as cover
That frame image, a large amount of manual labor can be reduced after the technology maturation and accuracy rate will obtain the promotion of matter；
Apple company is added to the function of automatic cutting, and the grinding in terms of composition by image aesthetics in handset image editor
Study carefully, automatic cutting image finally obtains the highest image of score.
Summary of the invention
Problem to be solved by this invention is: considering for conventional images quality evaluating method image subject feature insufficient
Lead to problems such as image quality evaluation accuracy rate not high, proposes a kind of image quality evaluation side based on body feature analysis
Method, this method is mainly from body features and images such as the depth of field, space configuration feature, background color complexity, the Lab features of main body
Fuzziness global characteristics etc. carry out image quality evaluation.
The technical scheme is that a kind of image quality evaluating method based on body feature analysis, including following steps
It is rapid:
Step S1 assesses database by image aesthetic quality and completes image quality measure model training, including following son
Step,
Step S11 carries out main body identification to every width training image in image aesthetic quality assessment database, determines main body
Region and background area；
Step S12 carries out body feature extraction, wherein body feature in conjunction with the body region and background area detected
It include: the Lab feature of the depth of field, space configuration feature, background color complexity, main body；
Step S13 calculates global characteristics image blur；
Step S14, picture quality classifying dictionary of the generation based on sparse representation are D；
Step S2 realizes picture quality discriminant classification, including following sub-step to test image；
Step S21 carries out body feature to the test image of input and global characteristics extracts, all features are connected
Y can be expressed as by obtaining total feature vector；
Step S22 differentiates test image classification quality, provides the objective function of its optimization first:
In formula, x is variable,It is the result of equation Optimization Solution；λ is a regularization parameter, is a constant, and D is
Picture quality classifying dictionary solves above-mentioned equation using Lasso method；
Calculate the error amount r that test image belongs to the i-th classi(y), calculation method is as follows:
Wherein,Indicate a column vector, it byIt generates, i.e. column vectorIn the value of corresponding i-th class be not 0,
Other values are set to 0 entirely；The judgment method that the test image belongs to picture quality classification c is as follows:
Further, in step S11 using Bogdan Alexe salient region detection method determine body region and
Background area, specific implementation is as follows,
(1.1) image preprocessing extracts R, G, B triple channel image of original image, carries out respectively to triple channel image more
Image f under different scale transformation is further divided into several identical candidate frames of size by subdimension transformation, is determined significant
The number U in property region；
(1.2) salient region detection is carried out to each candidate frame, detection method uses conspicuousness detection method, tool
Body are as follows: Fourier transformation is carried out to the image f under different scale transformation first and obtains its amplitude spectrum A (f) and phase spectrum P (f),
It defines Log amplitude spectrum L (f)=log (A (f)), h is the mean filter convolution kernel of a n*n, and g is a Gaussian Blur convolution
Core, then salient region S (f) calculation method is as follows:
R (f)=L (f)-h*L (f)
S (f)=g*F-1[exp(R(f)+iP(f))]2
Here i represents imaginary number, and exp represents exponential function；
(1.3) corresponding salient region S (f) is calculated to each candidate frame under different scale transformation, is pressed in conjunction with threshold value
The significance value of each candidate frame salient region is calculated according to following method, and the biggish preceding U candidate frame of significance value is made
For main body region, other regions are as background area；Wherein the calculation formula of significance value is as follows:
In above formula, the transformed different scale of S representative image, ω indicates that candidate frame, p indicate the pixel in candidate frame,Indicate the value of p point in salient region, θSIndicate the threshold value at size S.
Further, the specific implementation of step S12 is as follows,
(2.1) space configuration feature: being divided into 9 small cuboids for original image, makes it in groined type, and the 4 of intersection
A point, that is, golden section point determines that coordinate of main center's point in original image is (x1,y1), 4 golden section point coordinates point
It Wei not (x2,y2)、(x3,y3)、(x4,y4)、(x5,y5)；Specific step is as follows:
2.1a calculates separately main center's point and 4 golden section point distance L1、L2、L3、L4；
Wherein, ω, h are respectively the width and height of original image；
The straight line and trunnion axis angulation α that 2.1b, calculating main center's point and 4 golden section points are linked to be1,α2,
α3,α4；
αm-1=arctan [(ym-y1)/(xm-x1)] m=2,3,4,5
2.1c calculates the different subjects area accounting be overlapped with the rectangle on four angles of groined type
Wherein,Width and height of respectively j-th of the body region in i-th of rectangle, ω, h are respectively image
Width and height；
2.1d, space configuration feature is by L1,L2,L3,L4,α1,α2,α3,α4,It is connected into a column
Vector is constituted；
(2.2) depth of field: for the corresponding body region of significance value maximum in step (1.3)；
2.2a is calculated image blur: being become respectively to R, G of image, channel B using fast fourier transform algorithm
It changes, then selects maximum value therein, 1/5th of maximum value are set as to judge the threshold value of image blur, then calculate figure
Accounting of the pixel number (being less than threshold value) in whole image is obscured as in；
2.2b calculates the depth of field: by step 2.2a, calculating separately the corresponding main body of maximum significance value in step (1.3)
The fuzziness F in the region and fuzziness B of background area, using the ratio of F and B as the characteristic value of the depth of field；
(2.3) background color complexity: being divided into N number of equidistant section for R, G, B triple channel of original image respectively, this
Sample then has N in the three-dimensional color space of RGB3Color space is planted, the corresponding background of maximum significance value in traversal step (1.3)
Each pixel, count pixel in background and occupy the number ni of color space, then answered ni*100 as background color
The characteristic value of miscellaneous degree；
(2.4) the Lab feature of main body: original image is converted to Lab space from rgb space first, then to Lab space
Feature extraction is carried out, it is as follows that RGB is transformed into Lab space process:
The value first converted R, G, B of original image in the corresponding X, Y, Z axis of Lab space.
Wherein,
Conversion formula are as follows:
Wherein,
L is converted by the value in X, Y, Z axis*、a*、b*。
Wherein,
In addition, L*、a*、b*Respectively represent brightness, tone, colour temperature, Xn、Yn、ZnGeneral default value are as follows: 95.047,100.0,
108.883 calculating separately following features value:
L (brightness):
Wherein, L1For average brightness of the corresponding main body of significance value maximum in step (1.3) in Lab space；L2For
Average brightness of the image in Lab space；
A (tone characteristics):
Wherein, a1For tone average value of the corresponding main body of significance value maximum in step (1.3) in Lab space；a2For
Tone average value of the image in Lab space；
B (colour temperature feature):
Wherein, b1For colour temperature average value of the corresponding main body of significance value maximum in step (1.3) in Lab space；b2For
Colour temperature average value of the image in Lab space.
By A, B, C series connection constitute the Lab feature of main body for a column vector.
Further, the pyramidal natural image non-reference fuzzy evaluation of Fourier transformation and airspace is based in step S13
Method calculates global characteristics image blur, and calculating process is as follows:
Original image cutting is identical 9 blocks of size, is expressed as f by S13a1,..,f9；
S13b calculates power image P to original image and nine blocks according to the following formula:
P=10 (| G |2+1)
Wherein, G represents the Fourier transformation of correspondence image or block；
S13c obtains 10 power images by step S13b；Each power image is divided into size identical 9 again
A block, in addition itself, totally 10 power images；The value of power image is divided into K section according to size, counts each section
The number of interior performance number, using the statistical result as the feature vector of global characteristics image blur, the length of this feature vector
For 100K.
Further, the implementation that picture quality classifying dictionary D is obtained in step S14 is as follows,
Assuming that the classification comprising picture quality has N kind in image aesthetic quality assessment database, for the i-th class image, press
Corresponding classifying dictionary D is constructed according to following methodsi, to the jth width training image in the i-th class, by it according to step S11, S12 and
S13 step generates body feature (including: the Lab feature of the depth of field, space configuration feature, background color complexity, main body) and complete
Office's feature (image blur), and all series connection obtains a column vector d by all featuresj, then Di=[d1,…,dM], M
It is the number of training image in the i-th class；So total picture quality classifying dictionary can be expressed as D=[D1,…,DN]。
Compared with prior art, the invention has the advantages that and the utility model has the advantages that
1, for the picture comprising main body class, the present invention have chosen the depth of field, space configuration feature, background color complexity,
The body features such as the Lab feature of main body carry out image quality evaluation, and effect is better than general universality method.
2, the present invention has chosen the method based on sparse representation as classifier, more preferable than traditional classifier effect.
Detailed description of the invention
Fig. 1 is the flow chart for the portrait aesthetic quality evaluation analyzed the present invention is based on body feature.
Fig. 2 is the image after marking main body after identifying main body.
Fig. 3 is space configuration feature schematic diagram.
Fig. 4 is corresponding image after Fourier transformation.
Fig. 5 is test image.
Specific embodiment
Implementation of the invention is illustrated with a specific example below:
The database that the invention uses is AADB, and the database is by 10000 images selecting on the website Flickr.com
Composition, the score of each picture is obtained by 11 kinds of aesthetic evaluation indexs, we select portrait as study main body,
200 high quality portrait photos and 200 low quality portrait photos are picked from the data set, be all in two class images by
150 are used as training dataset, and 50 are used as test data set.It carries out in accordance with the following steps
S11: main body identification is carried out to every width training image, there is used herein the detections of the salient region of Bogdan Alexe
Method (What is an object?, Bogdan Alexe, etc) and determine body position, specific steps include:
(1.1) image preprocessing: extracting R, G, B triple channel image of original portrait image, respectively to triple channel image into
The transformation of 5 subdimension of row, i.e., zoom to corresponding scale size S*S for original image, here S=16, and 24,32,48,64.Every kind
Threshold θ under scaleSAre as follows: 0.43,0.32,0.34,0.35,0.26, determine that body region number is 8.Further by it is above-mentioned not
Several identical candidate frames of size are divided into the image f under change of scale；
(1.2) salient region detection is carried out to each candidate frame, detection method uses conspicuousness detection method
(" salient region detection: spectrum Remanent Model ", Saliency Detection:A Spectral Residual Approach,
Xiaodi Hou).In the method, its amplitude is obtained to by the image f progress Fourier transformation under different scale transformation first
A (f) and phase spectrum P (f) are composed, as shown in Figure 4.It defines Log amplitude spectrum L (f)=log (A (f)), h is the mean value filter of a n*n
Wave convolution kernel, n=3, g are a Gaussian Blur convolution kernel, size 10*10.Then salient region S (f) calculation method is as follows:
R (f)=L (f)-h*L (f)
S (f)=g*F-1[exp(R(f)+iP(f))]2
Here i represents imaginary number, and exp represents exponential function.
(1.3) corresponding salient region S (f) is calculated to each candidate frame, is calculated as follows in conjunction with threshold value every
The significance value of a candidate frame salient region, and using the maximum candidate frame of significance value as body region, other regions are made
For background area.The calculation formula of significance value is as follows:
In above formula, different scale after S representative image scaling, S=16 here, 24,32,48,64.ω indicates candidate
Frame, p indicate the pixel in candidate frame,Indicate the value of p point in salient region, θSIndicate the threshold value at size S, point
It Wei 0.43,0.32,0.34,0.35,0.26.Wherein a picture main body recognition result is as shown in Fig. 2.
S12: further, in conjunction with the body region and background area detected, body feature extraction is carried out, wherein main body
Feature includes: the Lab feature of space configuration feature, the depth of field, background color complexity, main body.
(2.1) space configuration feature: as shown in figure 3, original image is divided into nine small cuboids by us, it is made to be in
Groined type, four point, that is, golden section points of intersection.Determine that coordinate of main center's point in original image is (x1,y1), four
Golden section point coordinate is respectively (x2,y2)、(x3,y3)、(x4,y4)、(x5,y5).Specific step is as follows:
2.1a calculates separately main center's point and four golden section point distance L1、L2、L3、L4,
Wherein, ω, h are respectively the width and height of original image.
The straight line and trunnion axis angulation α that 2.1b, calculating main center's point and four golden section points are linked to be1,α2,
α3,α4。
αm-1=arctan [(ym-y1)/(xm-x1)] m=2,3,4,5
2.1c calculates different subjects (have 8 in fig 2, therefore, j=1 ..., 8) and the length on four angles of groined type
Rectangular be overlapped area accounting
Wherein,Width and height of respectively j-th of the body region in i-th of rectangle, ω, h are respectively original
The width and height of image.
Space configuration feature is by L1、L2、L3、L4,α1,α2,α3,α4 (j=1 ..., 8) it is connected into one
Column vector is constituted.
(2.2) depth of field: later in three body features, our situation corresponding to selective analysis most significant feature walk
Suddenly the corresponding areas case of maximum significance value in (1.3).The calculating of depth of field feature is divided into two steps, first step by us
Suddenly the calculation method of fuzziness is described, main body and background are then carried out fuzziness calculating by second step respectively, and final arrives
Depth of field characteristic value needed for us.
2.2a calculating image blur: being become respectively to R, G of image, channel B using fast fourier transform algorithm
It changes, then selects maximum value therein, we are set as judging the threshold value of image blur for 1/5th of maximum value, then count
Accounting of the pixel number (being less than threshold value) in whole image is obscured in nomogram picture, the accounting of attached drawing 2 is 0.0021.
2.2b calculates the depth of field: by step 2.2a, we calculate separately the fuzziness F and overall region of body region
Fuzziness B, using the ratio of F and B as the characteristic value of the depth of field, the depth of field characteristic value of attached drawing 2 is 1.0476.
(2.3) background color complexity: R, G, B triple channel of original image are divided into 16 equidistant areas by we respectively
Between, then there is 16*16*16=4096 kind color space in the three-dimensional color space of RGB in this way.We traverse each of background
Pixel counts the number ni that pixel in background image occupies color space, then using ni*100 as background color complexity
Characteristic value.Ni*100 is 3.1738 in example.
(2.4) the Lab feature of main body: original image is converted to Lab space from rgb space first by we, then to Lab
Space carries out feature extraction.It is as follows that RGB is transformed into Lab space process:
The value first converted R, G, B of original image in the corresponding X, Y, Z axis of Lab space.
Wherein,
Conversion formula are as follows:
Wherein,
L is converted by the value in X, Y, Z axis*、a*、b*。
Wherein,
Wherein L*、a*、b*Respectively represent brightness, tone, colour temperature, Xn、Yn、ZnGeneral default value are as follows: 95.047,100.0,
108.883.Calculate separately following features value:
L (brightness):
Wherein, L1Based on average brightness in Lab space；L2The average brightness for being image in Lab space.
A (tone characteristics):
Wherein, a1Based on tone average value in Lab space；a2The tone average value for being image in Lab space.
B (colour temperature feature):
Wherein, b1Based on colour temperature average value in Lab space；b2The colour temperature average value for being image in Lab space.
L in the present embodiment, A and B are respectively 101.3082,132.9119,133.2750.
By A, B, C series connection constitute the Lab feature of main body for a column vector.
S13: further, it is based on the pyramidal natural image non-reference fuzzy assessment method of Fourier transformation and airspace
(No-reference blur assessment in natural images using Fourier transform and
Spatial pyramids, Eftichia Mavridaki, etc) global characteristics image blur is calculated, calculating process is such as
Under:
Original image cutting is identical nine blocks of size, is expressed as f by S13a1,..,f9。
S13b calculates power image P to original image and nine blocks according to the following formula:
P=10 (| G |2+1)
Wherein, G represents the Fourier transformation of correspondence image or block.
S13c, by step S13b, we can obtain ten power images.Each power image is divided into size again
Identical nine blocks, in addition itself, totally ten power images.The value of power image is divided into 5 sections according to size, is counted
The number of performance number in each section, using the statistical result as the feature vector of global characteristics image blur, this feature to
The length of amount is 500.
S14: further, we generate the image quality measure model based on sparse representation.Assuming that in tranining database
Classification comprising picture quality has N kind, for the i-th class image, constructs corresponding classifying dictionary D by the following methodi.To i-th
Jth width training image in class, it (includes: the depth of field, space structure that it, which is generated body feature according to step S11, S12 and S13 step,
Figure feature, background color complexity, the Lab feature of main body) and global characteristics (image blur), and all features are all gone here and there
Join and obtains a column vector dj, then Di=[d1,…,dM].M is the number of training image in the i-th class.So total classification
Dictionary can be expressed as D=[D1,…,DN]。
Classification in the present embodiment in tranining database comprising picture quality has 2 kinds, and total classifying dictionary can be expressed as D
=[D1,D2]。
Step S2 specifically:
S21: body feature is carried out to input test image and global characteristics extract, the step and step S11, S12 and S13
Identical, we select a width test image, as shown in Fig. 5, calculate its feature vector y according to above-mentioned steps；
S22: differentiating the picture quality, provides the objective function of its optimization first:
In formula, x is variable,It is the result of equation Optimization Solution.λ is a regularization parameter, is a constant, λ
=8 × 10-5, D is total classifying dictionary, by all single classifying dictionary DiIt cascades, i here takes 1 and 2 respectively.Into
One step is by Lasso method (Least Absolute Shrinkage and Selection Operator) to above-mentioned equation
It is solved.Calculate the error amount r that test image belongs to the 1st class and the 2nd classi(y), calculation method is as follows:
Wherein,Indicate a column vector, it byIt generates, i.e. column vectorIn the value of corresponding i-th class be not 0,
His value is set to 0 entirely.The judgment method that the test image belongs to picture quality classification c is as follows:
In the present embodiment, it is high quality graphic that the image generic, which is finally calculated,.
Finally we are by our method and Kuo-Yen Lo (Assessment of Photo Aesthetics with
Efficiency method) is compared, and uses identical SVM as classifier, the standard of Kuo-Yen Lo method in experiment
True rate is 73%, our experimental method classification accuracy is 82%, it was demonstrated that the validity of proposed method.
1 Experimental comparison's data of table
| Experimental method | Classification accuracy |
| Kuo-Yen Lo | 73% |
| Dictionary learning | 82% |
Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention
The technical staff in domain can make various modifications or additions to the described embodiments or replace by a similar method
In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.
Claims (5)
1. a kind of image quality evaluating method based on body feature analysis, which comprises the steps of:
Step S1 assesses database by image aesthetic quality and completes image quality measure model training, including following sub-step,
Step S11 carries out main body identification to every width training image in image aesthetic quality assessment database, determines body region
The background area and；
Step S12 carries out body feature extraction in conjunction with the body region and background area detected, and wherein body feature includes:
The depth of field, space configuration feature, background color complexity, the Lab feature of main body；
Step S13 calculates global characteristics image blur；
Step S14, picture quality classifying dictionary of the generation based on sparse representation are D；
Step S2 realizes picture quality discriminant classification, including following sub-step to test image；
Step S21 carries out body feature to the test image of input and global characteristics extracts, all features are connected to obtain
Total feature vector is expressed as y；
Step S22 differentiates test image classification quality, provides the objective function of its optimization first:
In formula, x is variable,It is the result of equation Optimization Solution；λ is a regularization parameter, is a constant, D is image
Quality classification dictionary solves above-mentioned equation using Lasso method；
Calculate the error amount r that test image belongs to the i-th classi(y), calculation method is as follows:
Wherein,Indicate a column vector, it byIt generates, i.e. column vectorIn the value of corresponding i-th class be not 0, it is other
Value is set to 0 entirely；The judgment method that the test image belongs to picture quality classification c is as follows:
2. a kind of image quality evaluating method based on body feature analysis as described in claim 1, it is characterised in that: step
Body region and background area, specific implementation are determined using the salient region detection method of Bogdan Alexe in S11
It is as follows,
(1.1) image preprocessing extracts R, G, B triple channel image of original image, carries out multiple ruler to triple channel image respectively
Image f under different scale transformation is further divided into several identical candidate frames of size, determines conspicuousness area by degree transformation
The number U in domain；
(1.2) salient region detection is carried out to each candidate frame, detection method uses conspicuousness detection method, specifically
Are as follows: Fourier transformation is carried out to the image f under different scale transformation first and obtains its amplitude spectrum A (f) and phase spectrum P (f), it is fixed
Adopted Log amplitude spectrum L (f)=log (A (f)), h are the mean filter convolution kernels of a n*n, and g is a Gaussian Blur convolution kernel,
Then salient region S (f) calculation method is as follows:
R (f)=L (f)-h*L (f)
S (f)=g*F-1[exp(R(f)+iP(f))]2
Here i represents imaginary number, and exp represents exponential function；
(1.3) corresponding salient region S (f) is calculated to each candidate frame under different scale transformation, in conjunction with threshold value according to such as
Lower method calculates the significance value of each candidate frame salient region, and using the biggish preceding U candidate frame of significance value as master
Body region, other regions are as background area；Wherein the calculation formula of significance value is as follows:
In above formula, the transformed different scale of S representative image, ω indicates that candidate frame, p indicate the pixel in candidate frame,
Indicate the value of p point in salient region, θSIndicate the threshold value at size S.
3. a kind of image quality evaluating method based on body feature analysis as described in claim 1, it is characterised in that: step
The specific implementation of S12 is as follows,
(2.1) space configuration feature: being divided into 9 small cuboids for original image, makes it in groined type, 4 points of intersection
That is golden section point determines that coordinate of main center's point in original image is (x1,y1), 4 golden section point coordinates are respectively
(x2,y2)、(x3,y3)、(x4,y4)、(x5,y5)；Specific step is as follows:
2.1a calculates separately main center's point and 4 golden section point distance L1、L2、L3、L4；
Wherein, ω, h are respectively the width and height of original image；
The straight line and trunnion axis angulation α that 2.1b, calculating main center's point and 4 golden section points are linked to be1,α2,α3,α4；
αm-1=arctan [(ym-y1)/(xm-x1)] m=2,3,4,5
2.1c calculates the different subjects area accounting be overlapped with the rectangle on four angles of groined type
Wherein,Width and height of respectively j-th of the body region in i-th of rectangle, ω, h are respectively the width of image
And height；
2.1d, space configuration feature is by L1,L2,L3,L4,α1,α2,α3,α4,It is connected into a column vector structure
At；
(2.2) depth of field: for the corresponding body region of significance value maximum in step (1.3)；
2.2a is calculated image blur: being converted respectively to R, G of image, channel B using fast fourier transform algorithm,
Then maximum value therein is selected, 1/5th of maximum value are set as to judge the threshold value of image blur, then calculates image
In obscure accounting of the pixel number (be less than threshold value) in whole image；
2.2b calculates the depth of field: by step 2.2a, calculating separately the corresponding body region of maximum significance value in step (1.3)
Fuzziness F and background area fuzziness B, using the ratio of F and B as the characteristic value of the depth of field；
(2.3) background color complexity: R, G, B triple channel of original image are divided into N number of equidistant section respectively, existed in this way
Then there is N in the three-dimensional color space of RGB3Color space is planted, the corresponding background of maximum significance value is every in traversal step (1.3)
One pixel counts the number ni that pixel in background occupies color space, then using ni*100 as background color complexity
Characteristic value；
(2.4) original image: being converted to Lab space from rgb space first by the Lab feature of main body, is then carried out to Lab space
Feature extraction, it is as follows that RGB is transformed into Lab space process:
The value first converted R, G, B of original image in the corresponding X, Y, Z axis of Lab space；
Wherein,
Conversion formula are as follows:
Wherein,
L is converted by the value in X, Y, Z axis*、a*、b*；
Wherein,
In addition, wherein L*、a*、b*Respectively represent brightness, tone, colour temperature, Xn、Yn、ZnGeneral default value are as follows: 95.047,100.0,
108.883 calculating separately following features value:
L (brightness):
Wherein, L1For average brightness of the corresponding main body of significance value maximum in step (1.3) in Lab space；L2For image
Average brightness in Lab space；
A (tone characteristics):
Wherein, a1For tone average value of the corresponding main body of significance value maximum in step (1.3) in Lab space；a2For image
Tone average value in Lab space；
B (colour temperature feature):
Wherein, b1For colour temperature average value of the corresponding main body of significance value maximum in step (1.3) in Lab space；b2For image
Colour temperature average value in Lab space；
By A, B, C series connection constitute the Lab feature of main body for a column vector.
4. a kind of image quality evaluating method based on body feature analysis as described in claim 1, it is characterised in that: step
Global characteristics image mould is calculated based on the pyramidal natural image non-reference fuzzy assessment method of Fourier transformation and airspace in S13
Paste degree, calculating process are as follows:
Original image cutting is identical 9 blocks of size, is expressed as f by S13a1,..,f9；
S13b calculates power image P to original image and nine blocks according to the following formula:
P=10 (| G |2+1)
Wherein, G represents the Fourier transformation of correspondence image or block；
S13c obtains 10 power images by step S13b；Each power image is divided into identical 9 blocks of size again,
In addition itself, totally 10 power images；The value of power image is divided into K section according to size, counts each section internal strength
The number of rate value, using the statistical result as the feature vector of global characteristics image blur, the length of this feature vector is
100K。
5. a kind of image quality evaluating method based on body feature analysis as described in claim 1, it is characterised in that: step
The implementation that picture quality classifying dictionary D is obtained in S14 is as follows,
Assuming that the classification comprising picture quality has N kind in image aesthetic quality assessment database, for the i-th class image, under
It states method and constructs corresponding classifying dictionary Di, to the jth width training image in the i-th class, by it according to step S11, S12 and S13
Step generates body feature (including: the Lab feature of the depth of field, space configuration feature, background color complexity, main body) and the overall situation is special
It levies (image blur), and all series connection obtains a column vector d by all featuresj, then Di=[d1,…,dM], M is
The number of training image in i class；So total picture quality classifying dictionary can be expressed as D=[D1,…,DN]。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910046769.XA CN109829924B (en) | 2019-01-18 | 2019-01-18 | Image quality evaluation method based on principal feature analysis |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910046769.XA CN109829924B (en) | 2019-01-18 | 2019-01-18 | Image quality evaluation method based on principal feature analysis |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109829924A true CN109829924A (en) | 2019-05-31 |
| CN109829924B CN109829924B (en) | 2020-09-08 |
Family
ID=66860861
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910046769.XA Expired - Fee Related CN109829924B (en) | 2019-01-18 | 2019-01-18 | Image quality evaluation method based on principal feature analysis |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109829924B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110580678A (en) * | 2019-09-10 | 2019-12-17 | 北京百度网讯科技有限公司 | image processing method and device |
| CN110807759A (en) * | 2019-09-16 | 2020-02-18 | 幻想动力(上海)文化传播有限公司 | Method and device for evaluating photo quality, electronic equipment and readable storage medium |
| CN110853032A (en) * | 2019-11-21 | 2020-02-28 | 北京航空航天大学 | Unmanned aerial vehicle video aesthetic quality evaluation method based on multi-mode deep learning |
| CN110889718A (en) * | 2019-11-15 | 2020-03-17 | 腾讯科技（深圳）有限公司 | Method and apparatus for screening program, medium, and electronic device |
| CN111507941A (en) * | 2020-03-24 | 2020-08-07 | 杭州电子科技大学 | Composition characterization learning method for aesthetic quality evaluation |
| CN111784702A (en) * | 2020-06-16 | 2020-10-16 | 南京理工大学 | Grading method for image segmentation quality |
| CN112991308A (en) * | 2021-03-25 | 2021-06-18 | 北京百度网讯科技有限公司 | Image quality determination method and device, electronic equipment and medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106127744A (en) * | 2016-06-17 | 2016-11-16 | 广州市幸福网络技术有限公司 | Display foreground and background border Salience estimation and system |
| CN106485259A (en) * | 2015-08-26 | 2017-03-08 | 华东师范大学 | A kind of image classification method based on high constraint high dispersive principal component analysiss network |
| CN106570862A (en) * | 2016-10-25 | 2017-04-19 | 中国人民解放军信息工程大学 | Super-resolution reconstruction quality evaluation method and apparatus thereof |
| CN106778788A (en) * | 2017-01-13 | 2017-05-31 | 河北工业大学 | The multiple features fusion method of aesthetic evaluation is carried out to image |
-
2019
- 2019-01-18 CN CN201910046769.XA patent/CN109829924B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106485259A (en) * | 2015-08-26 | 2017-03-08 | 华东师范大学 | A kind of image classification method based on high constraint high dispersive principal component analysiss network |
| CN106127744A (en) * | 2016-06-17 | 2016-11-16 | 广州市幸福网络技术有限公司 | Display foreground and background border Salience estimation and system |
| CN106570862A (en) * | 2016-10-25 | 2017-04-19 | 中国人民解放军信息工程大学 | Super-resolution reconstruction quality evaluation method and apparatus thereof |
| CN106778788A (en) * | 2017-01-13 | 2017-05-31 | 河北工业大学 | The multiple features fusion method of aesthetic evaluation is carried out to image |
Non-Patent Citations (1)
| Title |
|---|
| 周正 等: "基于角点特征检测的视频图像质量评价方法", 《计算机工程》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110580678A (en) * | 2019-09-10 | 2019-12-17 | 北京百度网讯科技有限公司 | image processing method and device |
| CN110807759A (en) * | 2019-09-16 | 2020-02-18 | 幻想动力(上海)文化传播有限公司 | Method and device for evaluating photo quality, electronic equipment and readable storage medium |
| CN110807759B (en) * | 2019-09-16 | 2022-09-06 | 上海甜里智能科技有限公司 | Method and device for evaluating photo quality, electronic equipment and readable storage medium |
| CN110889718A (en) * | 2019-11-15 | 2020-03-17 | 腾讯科技（深圳）有限公司 | Method and apparatus for screening program, medium, and electronic device |
| CN110853032A (en) * | 2019-11-21 | 2020-02-28 | 北京航空航天大学 | Unmanned aerial vehicle video aesthetic quality evaluation method based on multi-mode deep learning |
| CN110853032B (en) * | 2019-11-21 | 2022-11-01 | 北京航空航天大学 | Unmanned aerial vehicle video tag acquisition method based on multi-mode deep learning |
| CN111507941A (en) * | 2020-03-24 | 2020-08-07 | 杭州电子科技大学 | Composition characterization learning method for aesthetic quality evaluation |
| CN111784702A (en) * | 2020-06-16 | 2020-10-16 | 南京理工大学 | Grading method for image segmentation quality |
| CN111784702B (en) * | 2020-06-16 | 2022-09-27 | 南京理工大学 | Grading method for image segmentation quality |
| CN112991308A (en) * | 2021-03-25 | 2021-06-18 | 北京百度网讯科技有限公司 | Image quality determination method and device, electronic equipment and medium |
| CN112991308B (en) * | 2021-03-25 | 2023-11-24 | 北京百度网讯科技有限公司 | Image quality determining method and device, electronic equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109829924B (en) | 2020-09-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109829924A (en) | A kind of image quality evaluating method based on body feature analysis | |
| CN106778788B (en) | The multiple features fusion method of aesthetic evaluation is carried out to image | |
| CN106529447B (en) | Method for identifying face of thumbnail | |
| CN108629338B (en) | Face beauty prediction method based on LBP and convolutional neural network | |
| CN104050471B (en) | Natural scene character detection method and system | |
| CN111753828B (en) | Natural scene horizontal character detection method based on deep convolutional neural network | |
| CN109684922B (en) | Multi-model finished dish identification method based on convolutional neural network | |
| CN104008375B (en) | The integrated face identification method of feature based fusion | |
| CN109767422A (en) | Pipe detection recognition methods, storage medium and robot based on deep learning | |
| CN106960176B (en) | Pedestrian gender identification method based on transfinite learning machine and color feature fusion | |
| CN101789005A (en) | Image searching method based on region of interest (ROI) | |
| CN104268590B (en) | The blind image quality evaluating method returned based on complementary combination feature and multiphase | |
| CN107092884B (en) | Rapid coarse-fine cascade pedestrian detection method | |
| CN108960404B (en) | Image-based crowd counting method and device | |
| CN110400293B (en) | No-reference image quality evaluation method based on deep forest classification | |
| CN103336835B (en) | Image retrieval method based on weight color-sift characteristic dictionary | |
| CN108829711B (en) | Image retrieval method based on multi-feature fusion | |
| CN109344818B (en) | Light field significant target detection method based on deep convolutional network | |
| CN112906550B (en) | Static gesture recognition method based on watershed transformation | |
| CN110796022B (en) | Low-resolution face recognition method based on multi-manifold coupling mapping | |
| CN105956570B (en) | Smiling face's recognition methods based on lip feature and deep learning | |
| CN108615229B (en) | Collision detection optimization method based on curvature point clustering and decision tree | |
| CN103473545A (en) | Text-image similarity-degree measurement method based on multiple features | |
| CN107169508A (en) | A kind of cheongsam Image emotional semantic method for recognizing semantics based on fusion feature | |
| CN109918542A (en) | A kind of convolution classification method and system for relationship diagram data |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200908 |