CN107680077A - A kind of non-reference picture quality appraisement method based on multistage Gradient Features - Google Patents
A kind of non-reference picture quality appraisement method based on multistage Gradient Features Download PDFInfo
- Publication number
- CN107680077A CN107680077A CN201710760654.8A CN201710760654A CN107680077A CN 107680077 A CN107680077 A CN 107680077A CN 201710760654 A CN201710760654 A CN 201710760654A CN 107680077 A CN107680077 A CN 107680077A
- Authority
- CN
- China
- Prior art keywords
- image
- gradient
- hidden layer
- predicted
- gradient features
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30168—Image quality inspection
Abstract
The invention discloses a kind of non-reference picture quality appraisement method based on multistage Gradient Features, image to be predicted is read in first;Treat prognostic chart picture to be pre-processed, the image to be predicted of colour is converted into gray level image, and change data type;Secondly the multistage Gradient Features of extraction gray level image are as characteristics of image to be predicted;Training set image is carried out with image identical to be predicted pretreatment and feature extraction, obtain training set characteristics of image, and the human subject after treatment of training set characteristics of image and correspondence image point is input to Score on Prediction model, it is trained, the Score on Prediction model trained;Characteristics of image to be predicted is input in the Score on Prediction model trained and obtains predicting fraction.The present invention is wider without reference to image, application;Complexity is relatively low, and real-time is higher;Utilize the adaptive normalization characteristic of training pattern so that the model of training has higher applicability for different images.
Description
Technical field
The present invention relates to image procossing, image quality evaluation field, be specifically one kind under conditions of without reference information from
The dynamic method to be scored picture quality.
Background technology
With communication and the development of network technology, digital picture is full of in various social network sites, mobile phone and camera, at me
Life in it is ubiquitous.But image is obtaining, transmit with processing etc. during can produce various forms and various degree
Distortion.In order that user receives the image with better quality, simultaneously quantized image distortion becomes automatic prognostic chart picture quality
It is very crucial.
Image quality evaluation is divided into subjective picture quality evaluation and evaluated with Objective image quality.Subjective picture quality evaluation by
People observes and given a mark, and takes human subject's opinion average mark as picture quality, although evaluation is the most accurate, takes time and effort,
Quick large-scale image scoring can not be realized.Objective image quality evaluation is divided into full reference, half reference, non-reference picture matter again
Amount evaluation.Wherein complete all or part of information with reference to the undistorted reference picture of half reference image quality appraisement needs, this
The hypothesis of sample is handled with communication and unreasonable in real world images, therefore non-reference picture quality appraisement is due to without necessarily referring to figure
Any information of picture can prognostic chart picture quality, attracted it is substantial amounts of research and concern.Non-reference picture quality appraisement method point
For for certain distortion without reference method with general without with reference to algorithm, it is general without with reference to algorithm due to for various distortions
All there is preferable evaluation effect, turn into the main flow of research.
General non-reference picture quality appraisement method is typically using natural scene statistical nature come prognostic chart picture quality.It is first
First need to read image to be predicted, then carry out image preprocessing, then extract the natural scene statistical nature of image and then incite somebody to action
These features are input to the Score on Prediction model trained, the image quality score predicted.Wherein need what is solved to ask
Topic mainly has two:Image characteristics extraction and the selection of Score on Prediction model.Wherein image characteristics extraction process is simpler, feature
Less, picture quality can more be reflected, then the real-time of algorithm is higher.And the quality of Score on Prediction model directly determines prediction
The relevant coherency of fraction and real image quality.Score on Prediction model mainly has SVM (SVM), adaptively strengthens BP
Neutral net (AdaBoost BP), deep neural network (DNN) etc..Wherein SVM is due to simple efficiently extensive utilization.
The content of the invention
Goal of the invention:In order to solve the problems, such as that prior art is present, it is not necessary to any information of reference picture, again can be more
Image characteristics extraction is more completely carried out soon, and the present invention provides a kind of non-reference picture quality appraisement based on multistage Gradient Features
Method.
Technical scheme:(referring to amended claims)
The present invention provides a kind of non-reference picture quality appraisement method based on multistage Gradient Features, has below beneficial to effect
Fruit:
(1) present invention is more extensive without necessarily referring to image information, adaptable scope.
(2) present invention extraction is characterized as the image gradient features simply easily calculated, and complexity is relatively low, has high real-time.
(3) present invention extracts the single order information of the gradient of image using processing, is equally extracted the higher-order gradients letter of image
Breath, can the preferably detailed information such as the curvature of response diagram picture and texture, raising quality evaluation precision.
(4) present invention is chosen by the use of the DNN neutral nets for stacking own coding as forecast model, combines SVM and SAE two
The advantages of kind model, on the basis of subjective and objective fraction uniformity is improved, still keep higher real-time.
(5) the adaptive normalization characteristic of training pattern is utilized so that the model of training has higher for different images
Applicability.
Brief description of the drawings
Fig. 1 is the total algorithm flow chart of the present invention;
Fig. 2 is image characteristics extraction flow chart of the present invention;
Fig. 3 is the DNN neural network structure figures of the present invention.
Embodiment
The invention will be further described with specific embodiment below in conjunction with the accompanying drawings.
For a database, such as LIVE IQA image data bases, we can make the 80% of distorted image therein
Test set, i.e., image to be predicted are used as training set, 20%.
As shown in figure 1, the non-reference picture quality appraisement method based on multistage Gradient Features, comprises the following steps:
1) image to be predicted is read in;
2) treat prognostic chart picture to be pre-processed, if image to be predicted is coloured image, the image to be predicted of colour is turned
Gray level image is turned to, and value type is converted into double types from unit8 or unit16;
3) gray level image obtained to step 2) carries out feature extraction, and it is special as image to be predicted to extract multistage Gradient Features
Sign;The multistage Gradient Features include:First-order Gradient feature, second order Gradient Features, three rank Gradient Features, the First-order Gradient are special
Sign, second order Gradient Features, three rank Gradient Features are again comprising gradient amplitude, relative gradient orientation and relative gradient amplitude.Because
The characteristics of image of extraction contains the order of information of image, and relative processing has been done to image, and order of information correspondence image
Details and textural characteristics, can the preferably detailed information such as the curvature of response diagram picture and texture, the feature of extraction can be more complete,
So as to improve quality evaluation precision.The image gradient features that must be characterized as simply easily calculating are extracted, complexity is relatively low, has high
Real-time.
If image to be predicted is I (x, y), extraction First-order Gradient feature specifically includes following steps:
31) Gauss local derviation wave filter calculated level direction x gradient component I is utilizedx' (i, j) and vertical direction y gradient
Component Iy' (i, j), the Gauss local derviation filter form are:
Wherein,For Gaussian kernel, γ represents x or Vertical Square in the horizontal direction
Local derviation is sought to y, i, j are respectively horizontal stroke, the ordinate value of pixel in the picture, and σ is the standard deviation of Gaussian kernel;
32) gradient magnitude is set as I ' (i, j), according to the gradient component I 31) obtainedx' (i, j) and Iy' (i, j), calculate I '
(i, j):
33) gradient orientations are set as ∠ I ' (i, j), according to the gradient component I 31) obtainedx' (i, j) and Iy' (i, j), calculate
∠ I ' (i, j):
Because orientation information is relative, we are in M × N localized masses in size, define single order relative gradient orientation ∠
I ' (i, j)RO:
∠ I ' (i, j)RO=∠ I ' (i, j)-∠ I ' (i, j)AVE (4)
Wherein, ∠ I ' (i, j)AVEFor local average orientation value, ∠ I ' (i, j) are definedAVE:
Wherein, Ix' (i, j)AVEWith Iy' (i, j)AVEThe respectively average local derviation of horizontal direction x and vertical direction y single order:
W is relative (i, j) offset collection in I (i, j) local neighborhood M × N localized masses;M, n represents relative localized mass
The line number and columns of center pixel (i, j) skew;
34) relative gradient amplitude is set as I ' (i, j)RM, according to Ix' (i, j), Iy' (i, j), Ix' (i, j)AVE、Iy' (i,
j)AVECalculate I ' (i, j)RM:
When extracting gradient amplitude image with Gradient Phase characteristics of image, the histogram of the pixel of gradient image is directly extracted
The variance of distribution, processing is simpler, has stronger real-time.
35) local derviation is sought again on single order local derviation figure is obtained, can obtain second order local derviation, can obtain multistage local derviation successively:
For two, three rank image gradient features, repeat step 32), 33), 34) obtain High-order Image Gradient Features;
36) for the First-order Gradient feature of image zooming-out to be predicted, second order Gradient Features and three rank Gradient Features, choose every
The distribution variance of kind feature represents this feature, therefore is directed to each image, extracts the characteristic vector that a dimension is 18.Utilize side
Difference characterizes the probability distribution of the pixel of each gradient image, and the feature of extraction is less, more succinctly, to the subsequent figures picture of forecast model
Training time is shorter.
4) to several natural undistorted images, different type list distortion processing, such as JP2K compression artefacts, JPEG pressure are carried out
Contracting distortion, WN white noise sound distortions etc., obtain the database of different type of distortion, and as training set image, training set image is entered
Row step 2), pretreatment and feature extraction 3), obtain training set characteristics of image, and by the mankind master of known training set image
See to divide and be input to Score on Prediction model, be trained, the Score on Prediction model trained, human subject divides for one
Primary data storehouse, such as LIVE IQA, TID2013 are known;
The Score on Prediction model, compared to more existing DNN neutral nets, increases to stack the DNN neutral nets of own coding
Add an adaptive normalization layer, also have adjusted each hidden layer neuron number.Its training step includes:
41) operation of step 3) is carried out to training set image, training set characteristics of image is obtained, according to training set characteristics of image
Obtain training set set of image characteristics Xtrain, XtrainFor the matrix of N × 18, wherein N is the number of training set image;For
The human subject of given data storehouse point is known, if subjective point of training set is Ytrain;
42) it is as follows as forecast model, forecast model design that DNN neutral nets are chosen:
A) utilize and stack the first hidden layer of own coding design, the second hidden layer, the 3rd hidden layer and the 4th hidden layer, first
Hidden layer, the second hidden layer, the neuron number of the 3rd hidden layer and the 4th hidden layer are respectively 18,100,18,18;First is hidden
Layer is hidden by the use of sigmod functions as transfer function to realize the automatic normalization of extracted feature, the second hidden layer, the 3rd hide
Layer and the 4th hidden layer realize three layers of sparse own coding;
B) unsupervised training is carried out to forecast model:To the second hidden layer, the 3rd hidden layer and the 4th hidden layer in a)
It is trained in the form of sparse own coding, preceding layer output inputs as later layer, trains one by one, obtains the weight of network;
C) supervision fine setting:A layer line is added after the first hidden layer, the second hidden layer, the 3rd hidden layer and the 4th hidden layer
Property Recurrent networks, the weight to be obtained in b) are used as the initial weight of supervised training, YtrainInstructed as linear regression output
Practice, finely tune weighting network;
D) SVM is returned:After supervision is finely tuned, the feature of the 4th hidden layer output is input in SVM, YtrainAs
Output training SVM, after the completion of preserve the model trained, wherein SVM is exported as the final pre- of whole forecast model
Survey result.
5) characteristics of image to be predicted obtained in step 3) is input to the Score on Prediction mould trained in step 4)
Obtain predicting fraction in type.
The present embodiment carries out unsupervised training using the DNN neutral nets for stacking own coding, while utilizes SVM
To help DNN to exercise supervision training, small parameter perturbations are carried out, the advantages of combining two kinds of models of SVM and SAE, learning model tool
There are stronger learning ability, while the normalization that the model can be adaptive to the characteristic parameter of extraction so that the model of training
There is higher applicability for different images;On the basis of subjective and objective fraction uniformity is improved, still keep higher
Real-time.
Claims (5)
1. a kind of non-reference picture quality appraisement method based on multistage Gradient Features, it is characterised in that comprise the following steps:
1) image to be predicted is read in;
2) treat prognostic chart picture to be pre-processed, if image to be predicted is coloured image, the image to be predicted of colour is converted into
Gray level image, and value type is converted into double types from unit8 or unit16;
3) gray level image obtained to step 2) carries out feature extraction, extracts multistage Gradient Features as characteristics of image to be predicted;
4) to several natural undistorted images, different type list distortion processing is carried out, obtains the database of different type of distortion, made
For training set images, step 2), pretreatment and feature extraction 3) are carried out to training set image, obtain training set characteristics of image,
And the human subject of known training set image point is input to Score on Prediction model, it is trained, the fraction trained
Forecast model;
5) characteristics of image to be predicted obtained in step 3) is input in the Score on Prediction model trained in step 4)
Obtain predicting fraction.
2. the non-reference picture quality appraisement method according to claim 1 based on multistage Gradient Features, it is characterised in that
The multistage Gradient Features include:First-order Gradient feature, second order Gradient Features, three rank Gradient Features, the First-order Gradient feature,
Second order Gradient Features, three rank Gradient Features are again comprising gradient amplitude, relative gradient orientation and relative gradient amplitude.
3. the non-reference picture quality appraisement method according to claim 2 based on multistage Gradient Features, it is characterised in that
If image to be predicted is I (x, y), extraction First-order Gradient feature specifically includes following steps:
31) Gauss local derviation wave filter calculated level direction x gradient component I is utilizedx' (i, j) and vertical direction y gradient component
Iy' (i, j), the Gauss local derviation filter form are:
Wherein,For Gaussian kernel, γ represents that x or vertical direction y are asked in the horizontal direction
Local derviation, i, j are respectively horizontal stroke, the ordinate value of pixel in the picture, and σ is the standard deviation of Gaussian kernel;
32) gradient magnitude is set as I ' (i, j), according to the gradient component I 31) obtainedx' (i, j) and Iy' (i, j), calculating I ' (i,
j):
33) gradient orientations are set as ∠ I ' (i, j), according to the gradient component I 31) obtainedx' (i, j) and Iy' (i, j), calculate ∠ I '
(i, j):
Due to orientation information be it is relative, we in size be in M × N localized masses define single order relative gradient orientation ∠ I ' (i,
j)RO:
∠ I ' (i, j)RO=∠ I ' (i, j)-∠ I ' (i, j)AVE
Wherein, ∠ I ' (i, j)AVEFor local average orientation value, ∠ I ' (i, j) are definedAVE:
Wherein, Ix' (i, j)AVEWith Iy' (i, j)AVEThe respectively average local derviation of horizontal direction x and vertical direction y single order:
W is relative (i, j) offset collection in I (i, j) local neighborhood M × N localized masses;M, n represents the center of relative localized mass
The line number and columns of pixel (i, j) skew;
34) relative gradient amplitude is set as I ' (i, j)RM, according to Ix' (i, j), Iy' (i, j), Ix' (i, j)AVE、Iy' (i, j)AVEMeter
Calculate I ' (i, j)RM:
。
4. the non-reference picture quality appraisement method according to claim 3 based on multistage Gradient Features, it is characterised in that
Extract High-order Image Gradient Features the step of be:
35) local derviation is sought again on single order local derviation figure is obtained, can obtain second order local derviation, can obtain multistage local derviation successively:For
2nd, three rank image gradient features, repeat step 32), 33), 34) obtain High-order Image Gradient Features;
36) for the First-order Gradient feature of image zooming-out to be predicted, second order Gradient Features and three rank Gradient Features, every kind of spy is chosen
The distribution variance of sign represents this feature, therefore is directed to each image, extracts the characteristic vector that a dimension is 18.
5. the non-reference picture quality appraisement method according to claim 4 based on multistage Gradient Features, it is characterised in that
To stack the DNN neutral nets of own coding, training step includes the Score on Prediction model:
41) operation of step 3) is carried out to training set image, training set characteristics of image is obtained, is obtained according to training set characteristics of image
Training set set of image characteristics Xtrain, XtrainFor the matrix of N × 18, wherein N is the number of training set image;For known
The human subject of database point is known, if subjective point of training set is Ytrain;
42) it is as follows as forecast model, forecast model design that DNN neutral nets are chosen:
A) using the first hidden layer of own coding design, the second hidden layer, the 3rd hidden layer and the 4th hidden layer is stacked, first hides
The neuron number of layer, the second hidden layer, the 3rd hidden layer and the 4th hidden layer is respectively 18,100,18,18;First hidden layer
Realize the automatic normalization of extracted feature by the use of sigmod functions as transfer function, the second hidden layer, the 3rd hidden layer and
4th hidden layer realizes three layers of sparse own coding;
B) unsupervised training is carried out to forecast model:To the second hidden layer, the 3rd hidden layer and the 4th hidden layer in a) with dilute
Dredge own coding form to be trained, preceding layer output inputs as later layer, trains one by one, obtains the weight of network;
C) supervision fine setting:A layer line is added after the first hidden layer, the second hidden layer, the 3rd hidden layer and the 4th hidden layer to return
Return network, the weight to be obtained in b) is used as the initial weight of supervised training, YtrainIt is trained as linear regression output, it is micro-
Adjust weighting network;
D) SVM is returned:After supervision is finely tuned, the feature of the 4th hidden layer output is input in SVM, YtrainAs output
Train SVM, after the completion of preserve the model trained, wherein SVM exports the final prediction knot as whole forecast model
Fruit.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710760654.8A CN107680077A (en) | 2017-08-29 | 2017-08-29 | A kind of non-reference picture quality appraisement method based on multistage Gradient Features |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710760654.8A CN107680077A (en) | 2017-08-29 | 2017-08-29 | A kind of non-reference picture quality appraisement method based on multistage Gradient Features |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107680077A true CN107680077A (en) | 2018-02-09 |
Family
ID=61135538
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710760654.8A Pending CN107680077A (en) | 2017-08-29 | 2017-08-29 | A kind of non-reference picture quality appraisement method based on multistage Gradient Features |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107680077A (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108734686A (en) * | 2018-05-28 | 2018-11-02 | 成都信息工程大学 | Multi-focus image fusing method based on Non-negative Matrix Factorization and visual perception |
CN108830829A (en) * | 2018-05-08 | 2018-11-16 | 天津大学 | Combine the reference-free quality evaluation algorithm of a variety of edge detection operators |
CN108898591A (en) * | 2018-06-22 | 2018-11-27 | 北京小米移动软件有限公司 | Methods of marking and device, electronic equipment, the readable storage medium storing program for executing of picture quality |
CN108985351A (en) * | 2018-06-27 | 2018-12-11 | 北京中安未来科技有限公司 | It is a kind of that the method and apparatus of blurred picture are identified based on gradient direction sparse features information, calculate equipment and storage medium |
CN108985350A (en) * | 2018-06-27 | 2018-12-11 | 北京中安未来科技有限公司 | It is a kind of that the method and apparatus of blurred picture are identified based on gradient magnitude sparse features information, calculate equipment and storage medium |
CN109064478A (en) * | 2018-07-17 | 2018-12-21 | 暨南大学 | A kind of astronomical image contour extraction method based on extreme learning machine |
CN109784358A (en) * | 2018-11-23 | 2019-05-21 | 南京航空航天大学 | A kind of non-reference picture quality appraisement method merging manual features and depth characteristic |
CN110009141A (en) * | 2019-03-22 | 2019-07-12 | 国网山东省电力公司经济技术研究院 | Climbing event prediction method and system based on SDAE feature extraction and svm classifier model |
CN110473181A (en) * | 2019-07-31 | 2019-11-19 | 天津大学 | Screen content image based on edge feature information without ginseng quality evaluating method |
CN110728656A (en) * | 2019-09-06 | 2020-01-24 | 西安电子科技大学 | Meta-learning-based no-reference image quality data processing method and intelligent terminal |
CN111307798A (en) * | 2018-12-11 | 2020-06-19 | 成都智叟智能科技有限公司 | Article checking method adopting multiple acquisition technologies |
CN111553879A (en) * | 2020-03-24 | 2020-08-18 | 上海交通大学 | Mobile phone photo selection area subjective comparison evaluation method and system |
WO2020172979A1 (en) * | 2019-02-25 | 2020-09-03 | 深圳市商汤科技有限公司 | Data processing method and apparatus, electronic device, and storage medium |
CN112652000A (en) * | 2020-12-30 | 2021-04-13 | 南京航空航天大学 | Method for judging small-scale motion direction of image |
Citations (1)
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 |
-
2017
- 2017-08-29 CN CN201710760654.8A patent/CN107680077A/en active Pending
Patent Citations (1)
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 |
Non-Patent Citations (7)
Title |
---|
ANISH MITTAL ET AL.: "No-Reference Image Quality Assessment in the Spatial Domain", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
LIXIONG LIU ET AL.: "Blind image quality assessment by relative gradient statistics and adaboosting neural network", 《SIGNAL PROCESSING:IMAGE COMMUNICATION》 * |
QIAOHONG LI ET AL.: "NO-REFERENCE IMAGE QUALITY ASSESSMENT BASED ON HIGH ORDER DERIVATIVES", 《2016 IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA AND EXPO》 * |
WUFENG XUE ET AL.: "Blind Image Quality Assessment Using Joint Statistics of Gradient Magnitude and Laplacian Features", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
YUMING LI ET AL.: "No-Reference Image Quality Assessment Using Shearlet Transform and Stacked Autoencoders", 《2015 IEEE INTERNATIONAL SYMPOSIUM ON CIRCUITS AND SYSTEMS》 * |
化毅: "基于梯度关联性和对比度掩膜的图像质量评价方法", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
陈伟: "《煤矿监控图像分析方法研究》", 31 December 2011 * |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108830829A (en) * | 2018-05-08 | 2018-11-16 | 天津大学 | Combine the reference-free quality evaluation algorithm of a variety of edge detection operators |
CN108830829B (en) * | 2018-05-08 | 2021-09-07 | 天津大学 | Non-reference quality evaluation algorithm combining multiple edge detection operators |
CN108734686A (en) * | 2018-05-28 | 2018-11-02 | 成都信息工程大学 | Multi-focus image fusing method based on Non-negative Matrix Factorization and visual perception |
CN108898591A (en) * | 2018-06-22 | 2018-11-27 | 北京小米移动软件有限公司 | Methods of marking and device, electronic equipment, the readable storage medium storing program for executing of picture quality |
CN108985351A (en) * | 2018-06-27 | 2018-12-11 | 北京中安未来科技有限公司 | It is a kind of that the method and apparatus of blurred picture are identified based on gradient direction sparse features information, calculate equipment and storage medium |
CN108985350A (en) * | 2018-06-27 | 2018-12-11 | 北京中安未来科技有限公司 | It is a kind of that the method and apparatus of blurred picture are identified based on gradient magnitude sparse features information, calculate equipment and storage medium |
CN108985351B (en) * | 2018-06-27 | 2021-11-26 | 北京中安未来科技有限公司 | Method and device for recognizing blurred image based on gradient direction sparse characteristic information, computing equipment and storage medium |
CN109064478A (en) * | 2018-07-17 | 2018-12-21 | 暨南大学 | A kind of astronomical image contour extraction method based on extreme learning machine |
CN109784358A (en) * | 2018-11-23 | 2019-05-21 | 南京航空航天大学 | A kind of non-reference picture quality appraisement method merging manual features and depth characteristic |
CN109784358B (en) * | 2018-11-23 | 2023-07-11 | 南京航空航天大学 | No-reference image quality evaluation method integrating artificial features and depth features |
CN111307798A (en) * | 2018-12-11 | 2020-06-19 | 成都智叟智能科技有限公司 | Article checking method adopting multiple acquisition technologies |
WO2020172979A1 (en) * | 2019-02-25 | 2020-09-03 | 深圳市商汤科技有限公司 | Data processing method and apparatus, electronic device, and storage medium |
CN110009141A (en) * | 2019-03-22 | 2019-07-12 | 国网山东省电力公司经济技术研究院 | Climbing event prediction method and system based on SDAE feature extraction and svm classifier model |
CN110473181A (en) * | 2019-07-31 | 2019-11-19 | 天津大学 | Screen content image based on edge feature information without ginseng quality evaluating method |
CN110728656A (en) * | 2019-09-06 | 2020-01-24 | 西安电子科技大学 | Meta-learning-based no-reference image quality data processing method and intelligent terminal |
CN111553879A (en) * | 2020-03-24 | 2020-08-18 | 上海交通大学 | Mobile phone photo selection area subjective comparison evaluation method and system |
CN112652000A (en) * | 2020-12-30 | 2021-04-13 | 南京航空航天大学 | Method for judging small-scale motion direction of image |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107680077A (en) | A kind of non-reference picture quality appraisement method based on multistage Gradient Features | |
CN111369563B (en) | Semantic segmentation method based on pyramid void convolutional network | |
CN108510485B (en) | Non-reference image quality evaluation method based on convolutional neural network | |
Lv et al. | Difference of Gaussian statistical features based blind image quality assessment: A deep learning approach | |
CN108717568A (en) | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network | |
CN108304826A (en) | Facial expression recognizing method based on convolutional neural networks | |
CN111696101A (en) | Light-weight solanaceae disease identification method based on SE-Inception | |
CN108846323A (en) | A kind of convolutional neural networks optimization method towards Underwater Targets Recognition | |
CN110110596B (en) | Hyperspectral image feature extraction, classification model construction and classification method | |
CN104268593A (en) | Multiple-sparse-representation face recognition method for solving small sample size problem | |
CN103942571B (en) | Graphic image sorting method based on genetic programming algorithm | |
Liu et al. | The classification and denoising of image noise based on deep neural networks | |
CN111178312B (en) | Face expression recognition method based on multi-task feature learning network | |
CN111047543A (en) | Image enhancement method, device and storage medium | |
CN105718932A (en) | Colorful image classification method based on fruit fly optimization algorithm and smooth twinborn support vector machine and system thereof | |
CN109614866A (en) | Method for detecting human face based on cascade deep convolutional neural networks | |
CN112215054A (en) | Depth generation countermeasure method for underwater acoustic signal denoising | |
CN110321805A (en) | A kind of dynamic expression recognition methods based on sequential relationship reasoning | |
CN111694977A (en) | Vehicle image retrieval method based on data enhancement | |
CN109344966A (en) | A kind of method of the full Connection Neural Network of efficient tensorization | |
CN113221913A (en) | Agriculture and forestry disease and pest fine-grained identification method and device based on Gaussian probability decision-level fusion | |
CN110781936B (en) | Construction method of threshold learnable local binary network based on texture description and deep learning and remote sensing image classification method | |
CN112263224A (en) | Medical information processing method based on FPGA edge calculation | |
CN106570910B (en) | Based on the image automatic annotation method from coding characteristic and Neighborhood Model | |
CN108960275A (en) | A kind of image-recognizing method and system based on depth Boltzmann machine |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180209 |
|
RJ01 | Rejection of invention patent application after publication |