CN108289221B - The non-reference picture quality appraisement model and construction method of rejecting outliers - Google Patents
The non-reference picture quality appraisement model and construction method of rejecting outliers Download PDFInfo
- Publication number
- CN108289221B CN108289221B CN201810044673.5A CN201810044673A CN108289221B CN 108289221 B CN108289221 B CN 108289221B CN 201810044673 A CN201810044673 A CN 201810044673A CN 108289221 B CN108289221 B CN 108289221B
- Authority
- CN
- China
- Prior art keywords
- image
- value
- model
- database
- test
- 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.)
- Expired - Fee Related
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N17/00—Diagnosis, testing or measuring for television systems or their details
- H04N17/004—Diagnosis, testing or measuring for television systems or their details for digital television systems
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
- H04N19/154—Measured or subjectively estimated visual quality after decoding, e.g. measurement of distortion
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the non-reference picture quality appraisement models and construction method of a kind of rejecting outliers.The building of the model the following steps are included: (1) using in image data base image and its subjective scores value carry out convolutional neural networks training, obtain image quality evaluation model；(2) mass fraction test is carried out to the image in image data base with the model；(3) test result and subjective scores value are subjected to linear regression analysis, abnormal image sample are rejected from image data base, the database updated；(4) it is tested after the image pattern in more new database being carried out step (1) operation, what is achieved the desired results is final Environmental Evaluation Model；If fall flat, then repeatedly step (2)-(3), until achieving the desired results.Image quality evaluating method provided by the invention, applicable image range is wide, subjective consistency is good, and it is few to evaluate the required time.
Description
Technical field
The invention belongs to field of image processings, are based on rejecting outliers and convolutional neural networks more particularly, to one kind
Non-reference picture quality appraisement method.
Background technique
Image is the important carrier of information in daily life, the quality of picture quality directly affect people subjective feeling and
The acquisition of information content, meanwhile, picture quality is also more various image processing algorithm performance superiority and inferiority and optimization system parameter
Important indicator, therefore image quality evaluation (Image qualityassessment, IQA) is received significant attention in recent decades.
Image quality evaluating method is divided into subjective picture quality evaluation method and objective according to whether by manually carrying out evaluation
Image quality evaluating method.For ordinary circumstance, because the mankind are the final recipients of image information, subjective assessment is most
Excellent quality evaluating method, however subjective assessment needs a large amount of personnel to participate in, time and economic cost are big, and operation is inconvenient.
The advantages of Objective image quality is evaluated is to make up that subjective evaluation method is at high cost, problem that can not be real-time, and disadvantage is can not be with
Subjective evaluation keeps highly effective consistency.
Objective image quality evaluation method is according to whether having original reference image that can be further divided into full reference, half refer to
And non-reference picture quality appraisement.Two kinds be most widely used are full Y-PSNRs with reference to Objective image quality evaluation method
(PeakSignal-to-Noise Ratio, PSNR) and mean square error (Mean Squared Error, MSE), but both
Evaluation method can not all keep high consistency with subjective evaluation result, and the original image needed for evaluating obtain difficulty compared with
Greatly.Half only needs to extract part raw image data for evaluating with reference to evaluation method, and is then not necessarily to original graph without reference method
As that can evaluate test image.Therefore for full reference picture evaluation method, half evaluates with reference to and without reference
Method strong flexibility, the scope of application is wider, more researching value.
The research of non-reference picture quality appraisement method is still within the starting stage at present, and is lost for specific mostly
Proper class type, applicable image fault type are few, and subjective consistency is poor, and it is long to evaluate the required time.
Summary of the invention
It is few for poor, the applicable image fault type of the subjective consistency of image quality evaluating method in the prior art, comment
The technical problem of valence time length, the present invention provides a kind of non-reference picture quality appraisement methods based on rejecting outliers.
To achieve the above object, the present invention provides a kind of structures of the non-reference picture quality appraisement model of rejecting outliers
Construction method includes the following steps:
(1) all images in image data base are divided into training set, verifying collection and three subsets of test set, by the figure
As all images in database divide several times at random, several groups image data is obtained；Training set between group and group, verifying collection and
The percentage that image number in test set accounts for total number of images is identical, and between group and group in training set, verifying collection and test set
Image it is not fully identical；The several groups image data is at least two groups of image datas；
(2) in every group of image data in step (1) training set and verifying concentrate image and the image in database
In corresponding subjective scores value carry out convolutional neural networks training, obtain several initial pictures Environmental Evaluation Models；
(3) with several initial pictures Environmental Evaluation Models obtained in step (2) respectively to described in step (1)
All images in image data base carry out mass fraction test, obtain several groups test result；Every group of test result is each
Initial pictures Environmental Evaluation Model carries out mass fraction to all images in image data base described in step (1) and tests
It arrives；
(4) by every group of test result obtained in step (3), subjective scores value corresponding with image is linearly returned respectively
Return analysis, and carry out rejecting outliers with statistical method, the image pattern of score exception is obtained, if the image of score exception
There is abnormal number and accounts for the group number of test result being more than or equal to 10% in sample, then by the abnormal image sample from image data base
Middle rejecting obtains the database for updating image pattern；
(5) database for the update image pattern that step (4) obtains is subjected to convolutional neural networks training, obtains image matter
Measure evaluation model, and the image quality evaluation model tested, if the SROCC value for the image quality evaluation that test obtains or
LCC value reaches predetermined value, then the image quality evaluation model is final Environmental Evaluation Model；If testing obtained image
The SROCC value or LCC value of quality evaluation do not reach predetermined value, then the number that will update image pattern obtained in step (4)
Step (1)-(4) are carried out according to library, until database subjective scores value corresponding with its of obtained update image pattern is rolled up
The SROCC value and/or LCC for the image quality evaluation that image quality evaluation model obtains after tested are obtained after product neural metwork training
Value reaches predetermined value, which is the non-reference picture quality appraisement model of rejecting outliers.
Preferably, several groups image data described in step (1) is more than or equal to ten groups.
Preferably, convolutional neural networks described in step (2) training the following steps are included:
(2-1) cuts the image pattern that training set described in step (1) and verifying are concentrated, and obtains image block,
The image block is normalized, the image block after being normalized；
Image block after the normalization that (2-2) obtains step (2-1) obtains image matter using convolutional neural networks training
Measure evaluation model.
Preferably, carrying out rejecting outliers with statistical method described in step (4) is by judging regression data
Whether studentized residuals exceed threshold value, if the studentized residuals of regression data exceed threshold value, the corresponding image of the regression data
Sample is the image pattern of score exception.
Preferably, the image number in training set described in step (1), verifying collection and test set accounts for the percentage of total number of images
Than being respectively 60%, 20% and 20%.
Preferably, convolutional neural networks described in step (2) and step (5) include convolutional layer, pond layer and full articulamentum,
It is connected between any two layers by excitation layer；Wherein first convolutional layer is used for input picture block number evidence, and the full articulamentum in end is used for
Export image score test value.
Preferably, mass fraction described in step (3) be predicted as carrying out the image in described image database cutting and
The normalization stripping and slicing image of each image, is then input in image quality evaluation model described in step (2), obtains by normalization
To the test result of the normalization stripping and slicing image, the average value of all stripping and slicing image prediction scores is taken, the test of the image is obtained
Score.
Preferably, image data base described in step (1) is natural image database or ultrasound image data library.
Preferably, the natural image database is II database of LIVE.
According to another aspect of the present invention, a kind of non-reference picture quality appraisement model of rejecting outliers is provided, is wrapped
It includes:
Image grouping module: for all images in image data base to be divided into training set, verifying collection and test set three
A subset；All images in described image database are divided several times at random, obtain several groups image data；Between group and group
Training set, verifying collection and the percentage that accounts for total number of images of the image number in test set are identical, and training set between group and group, test
Image in card collection and test set is not fully identical；The several groups image data is at least two groups of image datas；
Initial pictures Environmental Evaluation Model obtains module: for concentrate the training set in every group of image data with verifying
Corresponding subjective scores value carries out convolutional neural networks training in the database for image and the image, obtains several initial pictures
Environmental Evaluation Model；
Test result obtain module: for by initial pictures Environmental Evaluation Model respectively to all figures in image data base
As carrying out mass fraction test, several groups test result is obtained；Every group of test result is each initial pictures Environmental Evaluation Model
It is obtained after carrying out mass fraction test to all images in image data base；
Rejecting outliers module: for subjective scores value corresponding with image linearly to be returned respectively by every group of test result
Return analysis, and carry out rejecting outliers with statistical method, the image pattern of score exception is obtained, if the image of score exception
There is abnormal number and accounts for the group number of test result being more than or equal to 10% in sample, then by the abnormal image sample from image data base
Middle rejecting obtains the database for updating image pattern；
Image quality evaluation model obtains module: the database for that will update image pattern carries out convolutional neural networks instruction
After white silk, image quality evaluation model is obtained, and test the image quality evaluation model, if the picture quality that test obtains
The SROCC value or LCC value of evaluation reach predetermined value, then the image quality evaluation model is final Environmental Evaluation Model；
If the SROCC value or LCC value of testing obtained image quality evaluation do not reach predetermined value, then the number that will update image pattern
After carrying out image grouping, initial pictures quality evaluation according to library, obtain test result and detection exceptional value, until obtained update figure
Decent database subjective scores value corresponding with its obtains image quality evaluation model after carrying out convolutional neural networks training
The SROCC value or LCC value of the image quality evaluation obtained after tested reach predetermined value, which is abnormal
It is worth the non-reference picture quality appraisement model of detection.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, can obtain down and show
Beneficial effect:
(1) the above-mentioned non-reference picture quality method based on rejecting outliers and convolutional neural networks provided by the invention,
Study is trained to image data base, extracts characteristics of image without artificial, easy to operate, validity is high.
(2) the above-mentioned non-reference picture quality method based on rejecting outliers and convolutional neural networks provided by the invention,
Model when need to only call training to complete when test, evaluation speed are fast.
(3) the above-mentioned non-reference picture quality method based on rejecting outliers and convolutional neural networks provided by the invention,
It is not necessarily to any original reference image in model use after the completion of training, does not need corresponding high quality original image, nothing
It need to manually be evaluated, easy to operate, cost is extremely low.
(4) the above-mentioned non-reference picture quality method based on rejecting outliers and convolutional neural networks provided by the invention,
Belong to objective quality evaluating method, but there is high consistency with subjective quality evaluation result, improves objective quality
The accuracy of evaluation.
(5) the above-mentioned non-reference picture quality method based on rejecting outliers and convolutional neural networks provided by the invention,
It is adaptable to all type of distortion for including in training sample energy effective evaluation.
(6) the above-mentioned non-reference picture quality method based on rejecting outliers and convolutional neural networks provided by the invention,
Abnormal label data in training sample is screened, the robustness of image quality evaluation model is improved.
Detailed description of the invention
Fig. 1 is that the non-reference picture quality provided in an embodiment of the present invention based on rejecting outliers and convolutional neural networks is commented
The flow diagram of valence method；
Fig. 2 is the non-reference picture quality appraisement based on rejecting outliers and convolutional neural networks that inventive embodiments provide
The convolutional neural networks structure chart that method uses.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
Not constituting a conflict with each other can be combined with each other.
As shown in Figure 1, the non-reference picture quality based on rejecting outliers and convolutional neural networks of the embodiment of the present invention
Evaluation method includes the following steps:
(1) using the image in database, corresponding subjective scores value carries out convolutional Neural net in the database with the image
Network training obtains accurate image quality evaluation network.Wherein convolutional neural networks include convolutional layer, pond layer and multiple
Full articulamentum is connected between any two layers by excitation layer, wherein first convolutional layer is used for input picture block number evidence, end connects entirely
Layer is connect for exporting image score predicted value.Fig. 2 illustrates the model of convolutional neural networks, including sequentially connected convolutional layer,
The full articulamentum of excitation layer, pond layer, excitation layer and two (wherein each excitation layer Fig. 2 is not showed that).
Training process includes following sub-step:
(1.1) whole sample images are randomly divided into training set, verifying collection and test according to 0.6,0.2,0.2 ratio
Collection divides N group at random altogether (value of N should be not less than 10)；
(1.2) to the training set and verifying collection in every group, every width sample image is not overlapped cutting, obtains M 32*
32 image block carries out local normalization to all image blocks；
(1.3) for the image block after all normalization in every group of step (1.2), by its affiliated original sample image
Subjective average mark as its label value, then input network and be trained, finally obtain N number of Environmental Evaluation Model.
(2) mass fraction prediction is carried out to database images again with trained network.Specifically: to all sample graphs
It is then that M normalization stripping and slicing image of each image is defeated as carrying out cutting described in step (1) and part normalization one by one
Enter into trained model, obtain M evaluation score, all image block fractional values are averaged, it is final to obtain testing image
Fractional value.Every group model can correspond to a prediction fractional value, and each image shares N number of prediction fractional value.Every group of test result be
Each image quality evaluation model carries out what mass fraction was tested to all images in image data base.
(3) it will predict that score and artificial evaluation score carry out linear regression, and carry out rejecting outliers with statistical method,
Obtain the image of score exception.Specifically: all N groups are predicted that score and artificial evaluation score carry out linearly with SPSS software
It returns and residual values calculates, using the studentized residuals in statistics come the exceptional value in diagnostic evaluation score, find all N groups
There is abnormal number in exceptional value image in data, statistics.
(4) exceptional value obtained in previous step is rejected from sample database, repeats aforementioned 1-3 step.Specifically: according to
The abnormal number of the abnormal image and appearance found out in step (3), the image by number greater than 2 times are rejected from sample database, shape
Then the sample database of Cheng Xin repeats step (1)-(3) with new sample database.
(5) training will be repeated twice and the abnormal sample obtained later of detection is trained and tests, test is to convolution
Test set in neural metwork training is tested, and final quality evaluation network is obtained.Specifically: to by exceptional value twice
Detection and obtained sample database carries out the training in step (1) after rejecting obtains that test set is passed through network after network model
It is predicted, and linearly dependent coefficient and the calculating of Si Boerman rank correlation coefficient is carried out to prediction score and artificial evaluation score,
Obtain final evaluation model and result.
Embodiment 1
The present embodiment 1 the following steps are included:
(1) using the natural image and fractional value provided in the widely applied database LIVE II of natural image quality evaluation
As original sample and label value, convolutional neural networks training is carried out, initial image quality evaluation network is obtained.Wherein convolution
Neural network includes convolutional layer, pond layer and multiple full articulamentums, is connected between any two layers by excitation layer, wherein first volume
Lamination is used for input picture block number evidence, and the full articulamentum in end is for exporting image score predicted value.Fig. 2 illustrates convolutional Neural net
The model of network, including full articulamentum (the wherein each excitation of sequentially connected convolutional layer, excitation layer, pond layer, excitation layer and two
Layer Fig. 2 is not showed that).Last full articulamentum output is the corresponding image quality score of input picture.
Training process includes following sub-step:
(1.1) whole sample images are randomly divided into training set, verifying collection and test according to 0.6,0.2,0.2 ratio
Collection, random point 20 groups altogether；
(1.2) to the training set and verifying collection in every group, every width sample image is cut, the figure of N number of 32*32 is obtained
As block, local normalization is carried out to all image blocks.Specifically, used normalized function are as follows:Wherein I (I, j) is original graph picture point,It is
Picture point that treated, C are a positive constants, and to prevent σ (i, j) from being zero, P and Q is part normalization window size, C set
It is set to 1/255, P and Q is set as 7；
(1.3) for the image block after all normalization in every group of step (1.2), by its affiliated original sample image
Subjective average mark as its label value, then input network and be trained, as shown in Fig. 2, setting model parameter are as follows: first
Convolutional layer template size is set as 1*7*7*50, and pond layer template size is set as 1*26*26*1, and two full connection node layers are set
It is set to 800.Layer excitation in pond uses linear incentive, other layer of excitation is all made of ReLU.Wherein learning rate is preferably arranged to
0.0005,3000 learning rates of every training fall to the 0.9 of former learning rate.Finally obtain 20 Environmental Evaluation Models.
(2) mass fraction prediction is carried out to database images again with trained network.Specifically: to all sample graphs
It is then that M normalization stripping and slicing image of each image is defeated as carrying out cutting described in step (1) and part normalization one by one
Enter each trained model and obtain M evaluation score, all image block fractional values are averagely obtained into final point of testing image
Numerical value.Every group model can correspond to a prediction fractional value, and each image shares 20 prediction fractional values.
(3) it will predict that score and artificial evaluation score carry out linear regression, and carry out rejecting outliers with statistical method,
Obtain the image of score exception.Specifically: all 20 groups of prediction scores and artificial evaluation score are subjected to line with SPSS software
Property return and residual values calculate, using the studentized residuals in statistics come the exceptional value in diagnostic evaluation score, pass through judgement
Whether the studentized residuals of regression data exceed threshold value, if the studentized residuals of regression data exceed threshold value, the regression data
Corresponding image pattern is the image pattern of score exception.The threshold value can by setting obtain, set the threshold value as
2.576, it finds the exceptional value image in all 20 groups of data and abnormal number occurs.
(4) exceptional value obtained in previous step is rejected from sample database, repeats aforementioned 1-3 step.Specifically: according to
The abnormal number of the abnormal image and appearance found out in step (3), the image by number greater than 2 times are rejected from sample database, shape
Then the sample database of Cheng Xin repeats step (1)-(3) with new sample database.
(5) training will be repeated twice and the abnormal sample obtained later of detection is trained and tests, obtain final matter
Amount evaluation network.Specifically: the instruction in step (1) is carried out to the sample database obtained by rejecting outliers twice and after rejecting
Practice, predict test set by network after obtaining network model, and prediction score and artificial evaluation score is carried out linear
Related coefficient and Si Boerman rank correlation coefficient calculate, and obtain final evaluation model and result.Image quality evaluation model warp
The SROCC value and/or LCC value for testing obtained image quality evaluation reach predetermined value, and the image quality evaluation model is as different
The non-reference picture quality appraisement model of constant value detection.SROCC value and LCC value can be preset, when SROCC value and LCC value
When greater than 0.942, which is the non-reference picture quality appraisement model of rejecting outliers.Table 1 is shown
Quality evaluation result of the same test image pattern under distinct methods.
The SROCC value and LCC value of 1 distinct methods image quality evaluation of table
| Methods | PSNR | SSIM | FSIM | DIIVINE | BRISQUE | CORNIA | CNN_1 | CNN_2 |
| SROCC | 0.866 | 0.913 | 0.964 | 0.916 | 0.940 | 0.942 | 0.960 | 0.964 |
| LCC | 0.856 | 0.906 | 0.960 | 0.917 | 0.942 | 0.935 | 0.967 | 0.971 |
1 first three columns evaluation method of table is full reference image quality appraisement method, and latter four kinds are non-reference picture quality appraisement
Method, most next two columns be respectively this method carry out once and twice after rejecting outliers as a result, thickened portion is no reference
Optimal situation in image quality evaluation.It is based on the nothing of rejecting outliers and convolutional neural networks (CNN) as can be seen from the table
Reference image quality appraisement method is better than other methods, the optimal feelings being even more than in full reference image quality appraisement method
Condition.Compared to being obviously improved effect for other non-reference picture quality appraisement methods proposed before.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (9)
1. a kind of construction method of the non-reference picture quality appraisement model of rejecting outliers, which is characterized in that including walking as follows
It is rapid:
(1) all images in image data base are divided into training set, verifying collection and three subsets of test set, by described image number
Divide at random several times according to all images in library, obtains several groups image data；Training set, verifying collection and test between group and group
The percentage that the image number of concentration accounts for total number of images is identical, and the figure between group and group in training set, verifying collection and test set
As not fully identical；The several groups image data is at least two groups of image datas；
(2) right in the database to the image and the image of training set and verifying concentration in every group of image data in step (1)
The subjective scores value answered carries out convolutional neural networks training, obtains several initial pictures Environmental Evaluation Models；
(3) with several initial pictures Environmental Evaluation Models obtained in step (2) respectively to image described in step (1)
All images in database carry out mass fraction test, obtain several groups test result；Every group of test result is each initial
Image quality evaluation model carries out mass fraction to all images in image data base described in step (1) and tests to obtain
's；
(4) by every group of test result obtained in step (3), subjective scores value corresponding with image carries out linear regression point respectively
Analysis, and rejecting outliers are carried out with statistical method, the image pattern of score exception is obtained, if the image pattern of score exception
The group number that the number for exception occur accounts for test result then picks the abnormal image sample more than or equal to 10% from image data base
It removes, obtains the database for updating image pattern；
(5) database for the update image pattern that step (4) obtains is subjected to convolutional neural networks training, obtains picture quality and comments
Valence model, and the image quality evaluation model is tested, if the SROCC value or LCC of the image quality evaluation that test obtains
Value reaches predetermined value, then the image quality evaluation model is final Environmental Evaluation Model；If testing obtained image matter
The SROCC value or LCC value for measuring evaluation do not reach predetermined value, then the data that will update image pattern obtained in step (4)
Library carries out step (1)-(4), until database subjective scores value corresponding with its of obtained update image pattern carries out convolution
The SROCC value or LCC value that the image quality evaluation that image quality evaluation model obtains after tested is obtained after neural metwork training reach
To predetermined value, which is the non-reference picture quality appraisement model of rejecting outliers.
2. the construction method of the non-reference picture quality appraisement model of rejecting outliers as described in claim 1, feature exist
In several groups image data described in step (1) is more than or equal to ten groups.
3. the construction method of the non-reference picture quality appraisement model of rejecting outliers as described in claim 1, feature exist
In the training of, convolutional neural networks described in step (2) the following steps are included:
(2-1) cuts the image pattern that training set described in step (1) and verifying are concentrated, and image block is obtained, to this
Image block is normalized, the image block after being normalized；
Image block after the normalization that (2-2) obtains step (2-1) is obtained picture quality and is commented using convolutional neural networks training
Valence model.
4. the construction method of the non-reference picture quality appraisement model of rejecting outliers as described in claim 1, feature exist
In carrying out rejecting outliers with statistical method described in step (4) is by judging that the studentized residuals of regression data are
No to exceed threshold value, if the studentized residuals of regression data exceed threshold value, the corresponding image pattern of the regression data is score
Abnormal image pattern.
5. the construction method of the non-reference picture quality appraisement model of rejecting outliers as described in claim 1, feature exist
In the percentage that the image number in training set described in step (1), verifying collection and test set accounts for total number of images is respectively
60%, 20% and 20%.
6. the construction method of the non-reference picture quality appraisement model of rejecting outliers as described in claim 1, feature exist
In, convolutional neural networks described in step (2) and step (5) include convolutional layer, pond layer and full articulamentum, it is any between two layers
It is connected by excitation layer；Wherein first convolutional layer is used for input picture block number evidence, and the full articulamentum in end is for exporting image score
Test value.
7. the construction method of the non-reference picture quality appraisement model of rejecting outliers as described in claim 1, feature exist
In then the test of mass fraction described in step (3) will to be cut and being normalized the image in described image database
The normalization stripping and slicing image of each image is input in image quality evaluation model described in step (2), is obtained the normalization and is cut
The test result of block image takes the average value of all stripping and slicing image prediction scores, obtains the test result of the image.
8. the construction method of the non-reference picture quality appraisement model of rejecting outliers as described in claim 1, feature exist
In image data base described in step (1) is natural image database or ultrasound image data library；
Preferably, the natural image database is II database of LIVE.
9. a kind of non-reference picture quality appraisement model of rejecting outliers characterized by comprising
Image grouping module, for all images in image data base to be divided into training set, verifying collection and three sons of test set
Collection；All images in described image database are divided several times at random, obtain several groups image data；Training between group and group
The percentage that collection, verifying collection and the image number in test set account for total number of images is identical, and training set, verifying collection between group and group
It is not fully identical with the image in test set；The several groups image data is at least two groups of image datas；
Initial pictures Environmental Evaluation Model obtains module, the image for concentrating the training set in every group of image data with verifying
Corresponding subjective scores value carries out convolutional neural networks training in the database with the image, obtains several initial pictures quality
Evaluation model；
Test result obtain module, for by initial pictures Environmental Evaluation Model respectively to all images in image data base into
The test of row mass fraction, obtains several groups test result；Every group of test result is each initial pictures Environmental Evaluation Model to figure
It is obtained after carrying out mass fraction test as all images in database；
Rejecting outliers module, for subjective scores value corresponding with image to carry out linear regression point respectively by every group of test result
Analysis, and rejecting outliers are carried out with statistical method, the image pattern of score exception is obtained, if the image pattern of score exception
The group number that the number for exception occur accounts for test result then picks the abnormal image sample more than or equal to 10% from image data base
It removes, obtains the database for updating image pattern；
Image quality evaluation model obtains module, and the database for that will update image pattern carries out convolutional neural networks training
Afterwards, image quality evaluation model is obtained, and the image quality evaluation model is tested, if the picture quality that test obtains is commented
The SROCC value or LCC value of valence reach predetermined value, then the image quality evaluation model is final Environmental Evaluation Model；If
The SROCC value or LCC value for testing obtained image quality evaluation do not reach predetermined value, then the data that will update image pattern
After library carries out image grouping, initial pictures quality evaluation, obtains test result and detection exceptional value, until obtained more new images
The database of sample subjective scores value corresponding with its obtains image quality evaluation model warp after carrying out convolutional neural networks training
The SROCC value or LCC value for testing obtained image quality evaluation reach predetermined value, which is exceptional value
The non-reference picture quality appraisement model of detection.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810044673.5A CN108289221B (en) | 2018-01-17 | 2018-01-17 | The non-reference picture quality appraisement model and construction method of rejecting outliers |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810044673.5A CN108289221B (en) | 2018-01-17 | 2018-01-17 | The non-reference picture quality appraisement model and construction method of rejecting outliers |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108289221A CN108289221A (en) | 2018-07-17 |
| CN108289221B true CN108289221B (en) | 2019-08-30 |
Family
ID=62835190
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810044673.5A Expired - Fee Related CN108289221B (en) | 2018-01-17 | 2018-01-17 | The non-reference picture quality appraisement model and construction method of rejecting outliers |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108289221B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109308692B (en) * | 2018-07-30 | 2022-05-17 | 西北大学 | OCT image quality evaluation method based on improved Resnet and SVR mixed model |
| CN109741313B (en) * | 2018-12-29 | 2023-08-22 | 南京信息工程大学 | No-reference image quality evaluation method for independent component analysis and convolutional neural network |
| CN112149818B (en) * | 2019-06-27 | 2024-04-09 | 北京数安鑫云信息技术有限公司 | Threat identification result evaluation method and device |
| CN110464380B (en) * | 2019-09-12 | 2021-10-29 | 李肯立 | Method for controlling quality of ultrasonic section image of fetus in middle and late pregnancy |
| CN111311545A (en) * | 2020-01-19 | 2020-06-19 | 上海箱云物流科技有限公司 | Container detection method, device and computer readable storage medium |
| CN111368875B (en) * | 2020-02-11 | 2023-08-08 | 浙江昕微电子科技有限公司 | Method for evaluating quality of non-reference super-resolution image based on stacking |
| CN112669209B (en) * | 2020-12-24 | 2023-06-16 | 华中科技大学 | Three-dimensional medical image super-resolution reconstruction method and system |
| CN113191424A (en) * | 2021-04-28 | 2021-07-30 | 中国石油大学(华东) | Color fusion image quality evaluation method based on multi-model fusion |
| CN118071210B (en) * | 2024-04-17 | 2024-06-21 | 成都理工大学 | Ecological environment vulnerability assessment method combining CNN and PPM |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105160678A (en) * | 2015-09-02 | 2015-12-16 | 山东大学 | Convolutional-neural-network-based reference-free three-dimensional image quality evaluation method |
| CN105208374A (en) * | 2015-08-24 | 2015-12-30 | 宁波大学 | Non-reference image quality objective evaluation method based on deep learning |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN100559881C (en) * | 2008-05-09 | 2009-11-11 | 中国传媒大学 | A kind of method for evaluating video quality based on artificial neural net |
| CN101404778B (en) * | 2008-07-16 | 2011-01-19 | 河北师范大学 | Integrated non-reference video quality appraisement method |
| CN101742353B (en) * | 2008-11-04 | 2012-01-04 | 工业和信息化部电信传输研究所 | No-reference video quality evaluating method |
| US9591316B2 (en) * | 2014-03-27 | 2017-03-07 | Intel IP Corporation | Scalable video encoding rate adaptation based on perceived quality |
-
2018
- 2018-01-17 CN CN201810044673.5A patent/CN108289221B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105208374A (en) * | 2015-08-24 | 2015-12-30 | 宁波大学 | Non-reference image quality objective evaluation method based on deep learning |
| CN105160678A (en) * | 2015-09-02 | 2015-12-16 | 山东大学 | Convolutional-neural-network-based reference-free three-dimensional image quality evaluation method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108289221A (en) | 2018-07-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108289221B (en) | The non-reference picture quality appraisement model and construction method of rejecting outliers | |
| CN107123123B (en) | Image segmentation quality evaluating method based on convolutional neural networks | |
| CN105208374B (en) | A kind of non-reference picture assessment method for encoding quality based on deep learning | |
| CN112232476B (en) | Method and device for updating test sample set | |
| CN105160678A (en) | Convolutional-neural-network-based reference-free three-dimensional image quality evaluation method | |
| Bare et al. | An accurate deep convolutional neural networks model for no-reference image quality assessment | |
| CN108665460A (en) | Image quality evaluating method based on combination neural net and Classification Neural | |
| CN107959848B (en) | Universal no-reference video quality evaluation algorithms based on Three dimensional convolution neural network | |
| CN113177937B (en) | Improved YOLOv 4-tiny-based cloth defect detection method | |
| CN106303507A (en) | Video quality evaluation without reference method based on space-time united information | |
| CN105574901B (en) | A kind of general non-reference picture quality appraisement method based on local contrast pattern | |
| CN104036493B (en) | No-reference image quality evaluation method based on multifractal spectrum | |
| CN106651829B (en) | A kind of non-reference picture method for evaluating objective quality based on energy and texture analysis | |
| CN101562675A (en) | No-reference image quality evaluation method based on Contourlet transform | |
| CN108053396A (en) | A kind of more distorted image quality without with reference to evaluation method | |
| CN103517065A (en) | Method for objectively evaluating quality of degraded reference three-dimensional picture | |
| CN103745466A (en) | Image quality evaluation method based on independent component analysis | |
| CN109584203A (en) | Reorientation image quality evaluating method based on deep learning and semantic information | |
| Kwon et al. | Distorted representation space characterization through backpropagated gradients | |
| CN107146216A (en) | A kind of non-reference picture method for evaluating objective quality based on gradient self-similarity | |
| CN112016618A (en) | Measurement method for generalization capability of image semantic segmentation model | |
| CN112084842A (en) | Hydrological remote sensing image target identification method based on depth semantic model | |
| CN104574424A (en) | No-reference image blur degree evaluation method based on multiresolution DCT edge gradient statistics | |
| CN106683079A (en) | No-reference image objective quality evaluation method based on structural distortion | |
| CN104899893B (en) | The picture quality detection method of view-based access control model attention |
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: 20190830 Termination date: 20220117 |