CN109583500A - A kind of aesthetic images quality prediction system and method based on depth drift-diffusion method - Google Patents
A kind of aesthetic images quality prediction system and method based on depth drift-diffusion method Download PDFInfo
- Publication number
- CN109583500A CN109583500A CN201811453642.1A CN201811453642A CN109583500A CN 109583500 A CN109583500 A CN 109583500A CN 201811453642 A CN201811453642 A CN 201811453642A CN 109583500 A CN109583500 A CN 109583500A
- Authority
- CN
- China
- Prior art keywords
- image
- aesthetic
- score
- images
- aesthetic images
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
- G06F18/2193—Validation; Performance evaluation; Active pattern learning techniques based on specific statistical tests
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Abstract
The invention discloses a kind of aesthetic images quality prediction system and method based on depth drift-diffusion method: including: using depth drift-diffusion (DDD) method inspired by psychologist, on Caffe deep learning frame, it is distributed according to training image score, prediction aesthetic images score distribution;According to training aesthetic images classification data, aesthetic images class categories value is predicted;According to the aesthstic score of image in training image set, aesthetic images scoring is predicted.The forecast image is inputted into the aesthetic images score distribution method model, export the distribution of aesthetic images score, the scoring of image classification class label, aesthetic images, it solves the problems, such as that traditional Aesthetic picture quality method can not provide high accuracy to the aesthetic images evaluation score distribution in non-gaussian distribution, has reached the technical effect of aesthetic images penetration quality dynamic prediction.
Description
Technical field
The present invention relates to image analysis technology fields, are a kind of aesthetic images quality based on depth drift-diffusion method
Forecasting system and method.
Background technique
As in recent decades, computer aesthetic images are evaluated in computer vision, artificial intelligence, psychology, neurology department
Etc. fields have obtained more and more concerns.And the picture number in this mobile internet era, internet increases very
Fastly.The target image for finding high aesthetic quality is good to us.The assessment of the aesthetics used in ray image retrieval system
Engine can accomplish this point.After with the shooting of mobile phone madness, people are often difficult to choose the society that photo sharing arrives them
It hands on network.Therefore, aesthetic images assessment technique can help them to automatically select most beautiful image to share.
Currently, aesthetic images quality evaluation is still a urgent problem to be solved.Under normal conditions, following reason makes its tool
Challenging: (1) high there are differences in biggish field with low aesthetic quality;(2) high-caliber aesthetic rule and low-level
Characteristics of image it is opposed;(3) people evaluates the subjective attribute of the aesthetic quality of image.
In recent years, the evaluation of aesthetic images quality shows following apparent trend: (1) from traditional machine learning to depth
Degree study;(2) target is expressed as one-dimensional two tag along sort or numerical value assigns to multidimensional score distribution;(3) analysis of aesthetic perception from
Objectively evaluate subjective assessment.
Art methods can predict Gaussian Profile well, but fail at higher and lower both ends of scoring: i.e. low
With high non-gaussian distribution.
In terms of deep learning.Summarize the traditional-handwork aesthetic features of machine sort or homing method.Now, main stream approach
Based on depth convolutional neural networks, it is better than conventional method label Numerical evaluation in aesthetics classification and single aspect.
In terms of score distribution.Recently, certain methods, which are suggested, carries out binary classification using modification or the score generated distribution
With aesthstic Numerical evaluation.Researcher proposes a kind of improved support vector regression algorithm, in extensive AVA data set
Score distribution before publication and depth convolutional network are universal in the two small aesthstic data set of prediction.
Have to propose based on the CNN of the cumulative distribution of Jensen-Shannon divergence (CJS-CNN) [20] at present and predict
The aesthetics scoring distribution of mankind's grading, wherein the reliability sensitivity learning method of the kurtosis based on scoring distribution.This work is
The state-of-the-art technology method of the aesthetic Score on Prediction of image.
In view of the subjectivity of aesthetic images evaluation, different from image recognition, people can aesthetically provide difference
Image score.It is the optimal Gaussian distribution model of performance, and the mankind are for aesthetic images for only 62% image in AVA
The distribution of evaluation is often non-gaussian distribution.
Some investigators propose to measure and visualize the common recognition of aesthetic score using the degree of bias-kurtosis figure (S-K figure).Point
The skewness and kurtosis of number distribution is the power exponent compared with mean value and variance.With the help of S-K figure, they are from AVA data set
In have found four kinds of aesthstic score distribution patterns.Non-gaussian distribution or Gamma distribution can simulate four kinds of modes, especially extensively
Kurtosis.
Current aesthetic images evaluation is because being limited to the processing mode of data set, picture appraisal result obtains the originals such as direct
Cause is only capable of evaluating known image and makes general evaluation system analysis, when handling mass data collection, lacks to evaluate aesthetic images and move
The considerations of state changes, therefore have a significant impact to the accuracy of each aesthetic images score distribution of determination prediction.
In view of real image evaluation is a dynamic process, the distribution of aesthetic images evaluation score has more apparent special
Sign, therefore traditional Aesthetic image evaluation method can not make accurate result to aesthetic images prediction of quality.
Summary of the invention
Technology of the invention solves the problems, such as: overcoming the shortage of prior art, provides a kind of based on depth drift-diffusion method
Aesthetic images quality prediction system carries out the calculating of depth drift-diffusion method by evaluating aesthetic images, obtains a kind of beauty
Image quality estimation is learned as a result, improving the accuracy to the aesthetic images score distribution of non-gaussian distribution, solves traditional beauty
The problem of picture quality method can not provide high accuracy to the aesthetic images evaluation score distribution in non-gaussian distribution is learned, is reached
The technical effect of high accuracy can be provided for the aesthetic images evaluation score distribution of non-gaussian distribution by having arrived.
The technology of the present invention solution: a kind of aesthetic images quality prediction system based on depth drift-diffusion method, packet
It includes: data acquisition module, preprocessing module and neural network module;
Data acquisition module: determine that known AVA image aesthetics classification data concentrates required training image collection sample, verifying
Image set sample and test chart image set sample;
Preprocessing module: pre-processing known AVA image aesthetics categorized data set, to meet depth drift-expansion
The characteristic for dissipating model, by treated, data are sent in neural network module;
Neural network module: the proof diagram image set sample is inputted in neural network, utilizes depth by building neural network
Neural network method is spent, determines the distribution of aesthstic score, class categories value and the aesthetics scoring of the proof diagram image set sample;
Aesthetic score forecast of distribution module: the test chart image set sample is inputted into neural network, obtains the test chart
The aesthetic images mass values such as the distribution of aesthetic images score, class categories value, aesthetics scoring in image set sample are final prediction
As a result.
In data acquisition module, training image collection sample, proof diagram image set sample and test chart image set known to the determination
Sample, image classification model, specifically include: the aesthetic images data set of entitled AVA being randomly divided into training image collection sample, altogether
234599 images, remaining image is as test chart image set sample, and totally 19930 images, proof diagram image set sample are non-A/V A
Other images of aesthetic images data set, because the AVA data set used be known to, fixed data set, therefore used picture
Quantity is fixed numbers.
Using known AVA image aesthetics categorized data set execute image aesthetics classification task, classification task be based on
The GoogLeNet image classification model and ResNet image classification model of ImageNet image data set classification task.
In the preprocessing module, the formula based on depth Drift-diffusion Model are as follows:
Wherein, v indicates aesthetic images score distribution, and MiddleScore indicates that the median score of known distribution, m and n divide
Of positive attractor and negative attractor in multitask regression block is added when Dai Biao not be adjusted to known image disaggregated model
Number, i indicate the score in score distribution range, EposAnd EnegPositive attractor and negative attractor are respectively indicated, W is modified
White noise;
The positive attractor and negative attractor are determined indirectly by mark of the AVA data set to aesthetic images quality, specific public
Formula is as follows:
E=0.5*e- 0.5*U (0,10) (2)
W=λ * U (- 1,1) (3)
λ=sigmoid (θTf) (4)
Wherein, U (0,10) is known 10 point value tables in AVA data set, and U (- 1,1) is that 10 point value tables are mapped in (- 1,1)
On distribution, λ is intermediate parameters, and θ and f are used to indicate picture appraisal and its dynamic letter of dynamic aesthetics image score forecast of distribution
Number, θTIt indicates to carry out transposition expression to the picture appraisal indicated in the matrix form.
The proof diagram image set sample is inputted into neural network, determines each aesthetic images quality in the authentication image
Value, specifically includes:
The neural network includes: to be adjusted with known GoogLeNet image classification model, the aesthetic images score
Forecast of distribution model includes the network layer of the not no parameter of nine layers of convolutional neural networks layer for having parameter and two layers;With known ResNet
Image classification model is adjusted, and the aesthetic images score distribution prediction model includes nine layers of convolutional neural networks for having parameter
The network layer of layer and two layers of not no parameter;
The aesthetic images score distribution prediction model, which includes: 50 layers, to be had the convolutional neural networks layer of parameter and not to have for two layers
There is the network layer of parameter;Wherein, the convolutional neural networks layer of parameter is characterized extract layer for 50 layers, layer second from the bottom is length
For 1024 full articulamentum, layer last includes the full articulamentum that two length are respectively 256, is connected respectively with layer second from the bottom
It connects;θ in layer output formula (4) second from the bottomTF, the m and n that the last layer will export respectively in formula (1).Formula is substituted into again
(1), aesthetic images score distribution is acquired.
The proof diagram image set sample is inputted into neural network, determines each aesthetic images quality in the authentication image
Value, specifically includes:
The neural network includes: to be adjusted with known GoogLeNet image classification model, has parameter including nine layers
The network layer of the not no parameter of convolutional neural networks layer and two layers;It is adjusted with known ResNet image classification model, the beauty
Learn the network layer that image score forecast of distribution model includes the not no parameter of nine layers of convolutional neural networks layer for having parameter and two layers;
The aesthetic images score distribution prediction model, which includes: 50 layers, to be had the convolutional neural networks layer of parameter and not to have for two layers
There is the network layer of parameter;Wherein, the convolutional neural networks layer of parameter is characterized extract layer for 50 layers, layer second from the bottom is length
For 1024 full articulamentum, layer last includes the full articulamentum that two length are respectively 256, is connected respectively with layer second from the bottom
It connects, the θ in layer output formula (4) second from the bottomTF, the m and n that the last layer will export respectively in formula (1), then substitute into formula
(1), aesthetic images score distribution is acquired.
A kind of aesthetic images qualitative forecasting method of the invention, comprising the following steps:
(1) the aesthetic images data set of entitled AVA is randomly divided into training image collection sample, totally 234599 images, remained
Under image as test chart image set sample, totally 19930 images, proof diagram image set sample are non-A/V A aesthetic images data set
Other images;
(2) image aesthetics classification task is executed with known AVA image aesthetics categorized data set, classification task is based on
The GoogLeNet image classification model and ResNet image classification model of ImageNet image data set classification task;
(3) modified according to known image classification model to original data set, by data set by image score be distributed into
Row classification;Meanwhile the score of each image is ranked up by regular hour sequence, obtains satisfactory image score distribution
Sequence then calculates the aesthetic score mean value in each time point, and the score as the moment, finally obtains required meet
The data set of depth drift-diffusion method.By calculating the position of the point of maximum and minimum, positive attractor and negative suction are obtained
The position of introduction and number are obtained with the deep learning method that can describe dynamic image classification;
(4) classify to testing image, wherein 80% picture is divided into training set, be used for training pattern parameter, it is remaining
Accuracy rate of 20% image as test set, for detection model;Image random assortment is obtaining the result is that being uniformly distributed
, therefore the prediction being distributed for aesthetic score in time series is suitable for All Time point;
(5) image beauty is carried out to testing image using GoogLeNet image classification model and ResNet image classification model
Image score forecast of distribution is learned, the picture being divided to is inputted into two image classification models, exports the image for each time point
Aesthetic score distribution counts the aesthetic images score on whole result time points, obtains final required aesthetic images score
Distribution.
The advantages of the present invention over the prior art are that:
The present invention, which first uses, is based on depth Drift-diffusion Model, true according to training image collection sample, proof diagram image set sample
Determine aesthetic images quality prediction system, then image is predicted according to aesthetic images score distribution prediction model, obtains each
Aesthetic images mass value.It is realized using the above-mentioned aesthetic images score distribution prediction model of the present invention for different aesthetic scores point
The aesthetic images score distribution of cloth type, improves determining accuracy.In addition the present invention is not only simplified without being pre-processed
Specific steps, and the pattern features of damaged image itself can be prevented, and then improve the distribution of aesthetic images score, class categories
The accuracy of the aesthetic images mass values such as value, aesthetics scoring.
Detailed description of the invention
It in order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, below will be to institute in embodiment
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only some implementations of the invention
Example, for those of ordinary skill in the art, without any creative labor, can also be according to these attached drawings
Obtain other attached drawings.
Fig. 1 is aesthetic images of embodiment of the present invention quality prediction system composition block diagram;
Fig. 2 is that aesthetic images of embodiment of the present invention dynamic fractional predicts flow chart;
Fig. 3 is aesthetic images of embodiment of the present invention dynamic fractional figure;
Fig. 4 is depth of embodiment of the present invention Drift-diffusion Model figure.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
As shown in Figure 1, the function of various pieces is next discussed in detail.
Data acquisition module: determine that known AVA image aesthetics classification data concentrates required training image collection sample, verifying
Image set sample and test chart image set sample;
Preprocessing module: pre-processing known AVA image aesthetics categorized data set, to meet depth drift-expansion
The characteristic for dissipating model, by treated, data are sent in neural network module;
Neural network module: the proof diagram image set sample is inputted in neural network, utilizes depth by building neural network
Neural network method is spent, determines the distribution of aesthstic score, class categories value and the aesthetics scoring of the proof diagram image set sample;
Aesthetic score forecast of distribution module: the test chart image set sample is inputted into neural network, obtains the test chart
The aesthetic images mass values such as the distribution of aesthetic images score, class categories value, aesthetics scoring in image set sample are final prediction
As a result.
As shown in Fig. 2, the function of various pieces is next discussed in detail.
Step S1: training image collection sample, proof diagram image set sample and test chart image set sample are determined;
The aesthetic images data set of entitled AVA is randomly divided into training image collection sample, totally 234599 images, remaining
Image is as test chart image set sample, and totally 19930 images, proof diagram image set sample are its of non-A/V A aesthetic images data set
His image;
Step S2: it based on depth drift-diffusion (DDD) model in psychology, and is done on known image disaggregated model
Certain adjustment out, to determine aesthetic images score distribution prediction technique according to the training image collection sample;
Known image disaggregated model is the GoogLeNet image classification mould based on ImageNet image data set classification task
Type and ResNet image classification model;
Known models are needed to carry out certain modification, adapt it to used image set;
By taking the method being adjusted with known ResNet image classification model as an example, the aesthetic images score distribution prediction
Model includes: the network layer of the not no parameter of 50 layers of convolutional neural networks layer for having parameter and two layers;Wherein, there is parameter for 50 layers
Convolutional neural networks layer be characterized extract layer, layer second from the bottom is the full articulamentum that length is 1024, and layer last includes
Two length are respectively 256 full articulamentum, are connect respectively with layer second from the bottom.θ in layer output formula (4) second from the bottomTF, the m and n that the last layer will export respectively in formula (1).Formula (1) is substituted into again, can acquire aesthetic images score distribution.
Related formula is as follows:
Wherein, v indicates aesthetic images score distribution, and MiddleScore indicates that the median score of known distribution, m and n divide
Of positive attractor and negative attractor in multitask regression block is added when Dai Biao not be adjusted to known image disaggregated model
Number, EposAnd EnegPositive attractor and negative attractor are respectively indicated, W is modified white noise.
E=0.5*e- 0.5*U (0,10) (2)
W=λ * U (- 1,1) (3)
λ=sigmoid (θTf) (4)
Wherein, U (0,10) is known 10 point value tables in AVA data set, and U (- 1,1) is that 10 point value tables are mapped in (- 1,1)
On distribution, λ is intermediate parameters, and θ indicates the picture appraisal of the aesthstic image score forecast of distribution of dynamic, and f indicates dynamic aesthetics figure
As the kinematic function that score distribution is predicted, [0,1,2,3,4,5,6,7,8,9] is expressed as if 0 to 9 point of evaluation, such as t moment
Prediction result is [n0, n1, n2, n3, n4, n5, n6, n7, n8, n9], then f (t)=[n0, n1, n2, n3, n4, n5, n6, n7, n8, n9]。
Its kinematic function is expressed as evaluating multiple aesthetic images the function for being gradually abstracted the behavior that image constantly adds.
Its function abscissa be evaluation number, ordinate be 10 point values, function input be aesthetic images evaluate, function output at present
The mean value of whole aesthetic images evaluation of acquisition, therefore maximum point of the kinematic function in domain is in formula (1)
Positive attractor, the minimum point of kinematic function is the negative attractor in formula (1).
Step S3: being calculated a kind of aesthetic images score distribution prediction technique and model, the module include:
Sample determining module, for determining training image collection sample, proof diagram image set sample and test chart image set sample;
Aesthetic images score distribution prediction model determining module, for the aesthetic images based on depth Drift-diffusion Model
Score distribution prediction technique determines each aesthetic images mass value according to the training image collection sample;
Prediction module, for the proof diagram image set sample to be inputted the aesthetic images score distribution prediction technique, really
Each aesthetic images score distribution predicted value and image classification class label in the fixed authentication image;
Error amount determining module, for true according to each aesthetic images score distribution predicted value and image classification class label
The error amount of fixed each image;
Judgment module, for judging whether the error amount of each image is less than setting error;It is if it is less than setting error, then defeated
Aesthetic images score distribution prediction model out;If it is larger than or equal to setting error, then it is pre- to execute the aesthetic images score distribution
Survey method determines aesthetic images score distribution according to the training image collection sample;
Output module, it is defeated for the test chart image set sample to be inputted the aesthetic images score distribution prediction model
Each aesthetic images value in the test chart image set sample out.
The details of each module is as follows:
Module P1: sample determining module
The aesthetic images data set of entitled AVA is randomly divided into training image collection sample, totally 234599 images, remaining
Image is as test chart image set sample, and totally 19930 images, proof diagram image set sample are its of non-A/V A aesthetic images data set
His image.
Module P2: aesthetic images score distribution prediction model determining module
It is used in the aesthetic images score distribution prediction technique of depth Drift-diffusion Model, according to the training image collection
Sample determines each aesthetic images mass value.To improve accuracy, module is to the aesthetic images score distribution having in Gaussian Profile
Image set and in non-gaussian analyze aesthetic images score distribution image set predicted.Meanwhile to image known to two kinds point
Class model is adjusted, and is predicted image set.
Module P3: prediction module
For the proof diagram image set sample to be inputted the aesthetic images score distribution prediction technique, the verifying is determined
Each aesthetic images score distribution predicted value and image classification class label in image;
By taking the method being adjusted with known ResNet image classification model as an example, the aesthetic images score distribution prediction
Model includes:
The network layer of the not no parameter of 50 layers of convolutional neural networks layer for having parameter and two layers;Wherein, there is parameter for 50 layers
Convolutional neural networks layer be characterized extract layer.
First layer as input layer carries out the adjustment of a scale size to image, is allowed to all meet former ImageNet
The size of image set in image classification experiment, as long 224 pixels, wide 224 pixel.
Layer second from the bottom is the full articulamentum that length is 1024, layer last include two length be respectively 256 it is complete
Articulamentum is connect with layer second from the bottom respectively.θ in layer output formula (4) second from the bottomTF, the last layer are public by output respectively
M and n in formula (1).Formula (1) is substituted into again, is obtained the prediction of aesthetic images score distribution, score distribution is re-designed as one-dimensional
Two metatags obtain aesthetic images classification prediction, use the average value of aesthetic images score distribution pre- as aesthetic images scoring
The result of survey.
Model is trained using Caffe frame and test model.The prediction of model will be repeated repeatedly according to regression problem
Generation, and training pattern is used for using stochastic gradient descent.
Operation is iterated to model every time, inputs 48 images every time, the weight decaying of neural network is set as
0.0005, maximum number of iterations 120000 downloads method training pattern using gradient.
Module P4: error amount determining module
Each aesthetic images score distribution predicted value and image classification class label determine the error amount of each image set, when accidentally
When difference is larger, stochastic gradient descent method (SGD) will make error carry out to lower direction.
Module P5: judgment module
For judging whether the error amount of each image is less than setting error;If it is less than setting error, then aesthstic figure is exported
As score distribution prediction model;If it is larger than or equal to setting error, then the aesthetic images score distribution prediction technique, root are executed
Aesthetic images score distribution is determined according to the training image collection sample.
Module P6: output module
For by the test chart image set sample, returned using multitask and layer of classifying according to one-dimensional vector respectively to aesthetics
Image score distribution is predicted, to obtain the aesthetic images matter such as the distribution of aesthetic images score, class categories value, aesthetics scoring
Magnitude.
As shown in figure 3, an example of aesthetic images score distribution prediction, is expressed as three in three convex nodes and evaluates
Factor, three concave section points be expressed as three evaluation difference factors.
As shown in figure 4, depth Drift-diffusion Model is divided into input picture, residual error network portion, multitask recurrence part three
A part, wherein residual error network portion includes 50 identical residual error network modules, and each module is by convolutional layer, batch normalizing
Change layer, the basic units such as activation primitive layer constitute, multitask returns part and is made of full articulamentum, and size is 256 × 1 to connect entirely
Connect the regression block that layer is used for multiple tasks.
The present invention is based on depth drift-diffusion (DDD) model, neural network can be good at extracting the feature of image,
The prediction of aesthetic images score distribution is yielded good result.For presently, there are to be in non-gaussian distribution common image
Fraction Model proposes a kind of preferable aesthetic images score distribution prediction model.The level of front includes that 50 compressions swash
Flexible module extracts the aesthetic features of image, extracts the multitask feature that returns and classify, last network in rear several levels
These features are passed through the weighted value of back-propagation algorithm more new model.Model passes through the tune to known image classification model
It is whole, and the calculating to positive attractor and negative attractor is added, obtaining can be to the aesthetics of common Gaussian Profile and non-gaussian distribution
Image score prediction model.And on this basis, the aesthetic images score distribution of acquisition is arranged and is calculated, obtain aesthetics
The data such as the distribution of image score, class categories value, aesthetics scoring.
Used herein a specific example illustrates the principle and implementation of the invention, and above embodiments are said
It is bright to be merely used to help understand method and model and its core concept of the invention;Meanwhile for the general technology people of this field
Member, according to the thought of the present invention, there will be changes in the specific implementation manner and application range.In conclusion this explanation
Book content should not be construed as limiting the invention.
Claims (8)
1. a kind of aesthetic images quality prediction system based on depth drift-diffusion method characterized by comprising data are adopted
Collect module, preprocessing module and neural network module;
Data acquisition module: determine that known AVA image aesthetics classification data concentrates required training image collection sample, authentication image
Collect sample and test chart image set sample;
Preprocessing module: pre-processing known AVA image aesthetics categorized data set, to meet depth drift-diffusion mould
The characteristic of type, by treated, data are sent in neural network module;
Neural network module: building neural network inputs the proof diagram image set sample in neural network, utilizes depth mind
Through network method, the distribution of aesthstic score, class categories value and the aesthetics scoring of the proof diagram image set sample are determined;
Aesthetic score forecast of distribution module: the test chart image set sample is inputted into neural network, obtains the test chart image set
The aesthetic images mass values such as the distribution of aesthetic images score, class categories value, aesthetics scoring in sample are final prediction result.
2. aesthetic images quality prediction system according to claim 1, it is characterised in that: described in data acquisition module
It determines known training image collection sample, proof diagram image set sample and test chart image set sample, image classification model, specifically includes:
The aesthetic images data set of entitled AVA is randomly divided into training image collection sample, totally 234599 images, remaining image conduct
Test chart image set sample, totally 19930 images, proof diagram image set sample are other images of non-A/V A aesthetic images data set, because
The AVA data set used is known, fixed data set, therefore used picture number is fixed numbers.
3. aesthetic images quality prediction system according to claim 1, it is characterised in that: use known AVA image aesthetics
Categorized data set executes image aesthetics classification task, and classification task is based on ImageNet image data set classification task
GoogLeNet image classification model and ResNet image classification model.
4. aesthetic images quality prediction system according to claim 1, it is characterised in that: in the preprocessing module, base
In the formula of depth Drift-diffusion Model are as follows:
Wherein, v indicates aesthetic images score distribution, and MiddleScore indicates the median score of known distribution, m and n generation respectively
The number of positive attractor and negative attractor in multitask regression block, i are added when table is adjusted known image disaggregated model
Indicate the score in score distribution range, EposAnd EnegPositive attractor and negative attractor are respectively indicated, W is modified white noise
Sound.
5. aesthetic images quality prediction system according to claim 4, it is characterised in that: the positive attractor and negative attraction
Son determines indirectly by mark of the AVA data set to aesthetic images quality, specific formula is as follows:
E=0.5*e- 0.5*U (0,10) (2)
W=λ * U (- 1,1) (3)
λ=sigmoid (θTf) (4)
Wherein, U (0,10) is known 10 point value tables in AVA data set, and U (- 1,1) is that 10 point value tables are mapped on (- 1,1)
Distribution, λ are intermediate parameters, and θ and f are used to indicate picture appraisal and its kinematic function of the aesthstic image score forecast of distribution of dynamic,
θTIt indicates to carry out transposition expression to the picture appraisal indicated in the matrix form.
6. aesthetic images quality prediction system according to claim 1, it is characterised in that: by the proof diagram image set sample
Neural network is inputted, each aesthetic images mass value in the authentication image is determined, specifically includes:
The neural network includes: to be adjusted with known GoogLeNet image classification model, the aesthetic images score distribution
Prediction model includes the network layer of the not no parameter of nine layers of convolutional neural networks layer for having parameter and two layers;With known ResNet image
Disaggregated model is adjusted, the aesthetic images score distribution prediction model include nine layers of convolutional neural networks layer for having a parameter and
The network layer of two layers of not no parameter;
The aesthetic images score distribution prediction model, which includes: 50 layers, to be had the convolutional neural networks layer of parameter and does not join for two layers
Several network layers;Wherein, the convolutional neural networks layer of parameter is characterized extract layer for 50 layers, layer second from the bottom is for length
1024 full articulamentum, layer last include the full articulamentum that two length are respectively 256, are connected respectively with layer second from the bottom
It connects;θ in layer output formula (4) second from the bottomTF, the m and n that the last layer will export respectively in formula (1).Formula is substituted into again
(1), aesthetic images score distribution is acquired.
7. aesthetic images quality prediction system according to claim 1, it is characterised in that: by the proof diagram image set sample
Neural network is inputted, each aesthetic images mass value in the authentication image is determined, specifically includes:
The neural network includes: to be adjusted with known GoogLeNet image classification model, there is the convolution of parameter including nine layers
The network layer of the not no parameter of neural net layer and two layers;It is adjusted with known ResNet image classification model, the aesthetics figure
As score distribution prediction model includes the network layer of the not no parameter of nine layers of convolutional neural networks layer for having parameter and two layers;
The aesthetic images score distribution prediction model, which includes: 50 layers, to be had the convolutional neural networks layer of parameter and does not join for two layers
Several network layers;Wherein, the convolutional neural networks layer of parameter is characterized extract layer for 50 layers, layer second from the bottom is for length
1024 full articulamentum, layer last include the full articulamentum that two length are respectively 256, are connected respectively with layer second from the bottom
It connects, the θ in layer output formula (4) second from the bottomTF, the m and n that the last layer will export respectively in formula (1), then substitute into formula
(1), aesthetic images score distribution is acquired.
8. a kind of aesthetic images qualitative forecasting method based on depth drift-diffusion method, it is characterised in that: including following step
It is rapid:
(1) the aesthetic images data set of entitled AVA is randomly divided into training image collection sample, totally 234599 images, remaining
Image is as test chart image set sample, and totally 19930 images, proof diagram image set sample are its of non-A/V A aesthetic images data set
His image;
(2) image aesthetics classification task is executed with known AVA image aesthetics categorized data set, classification task is based on ImageNet
The GoogLeNet image classification model and ResNet image classification model of image data set classification task;
(3) it is modified according to known image classification model to original data set, data set is divided by image score
Class;Meanwhile the score of each image is ranked up by regular hour sequence, obtains satisfactory image score distribution sequence
Column, then calculate the aesthetic score mean value in each time point, and the score as the moment, finally obtain and required meet depth
Spend the data set of drift-diffusion method.By calculating the position of the point of maximum and minimum, positive attractor and negative attraction are obtained
The position of son and number are obtained with the deep learning method that can describe dynamic image classification;
(4) classify to testing image, wherein 80% picture is divided into training set, be used for training pattern parameter, residue 20%
Accuracy rate of the image as test set, for detection model;Image random assortment, it is obtaining the result is that equally distributed, because
This is suitable for All Time point for the prediction that aesthetic score in time series is distributed;
(5) image aesthetics figure is carried out to testing image using GoogLeNet image classification model and ResNet image classification model
As score distribution prediction, the picture being divided to is inputted into two image classification models, exports the image aesthetics for each time point
Score distribution counts the aesthetic images score on whole result time points, obtains final required aesthetic images score distribution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811453642.1A CN109583500A (en) | 2018-11-30 | 2018-11-30 | A kind of aesthetic images quality prediction system and method based on depth drift-diffusion method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811453642.1A CN109583500A (en) | 2018-11-30 | 2018-11-30 | A kind of aesthetic images quality prediction system and method based on depth drift-diffusion method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109583500A true CN109583500A (en) | 2019-04-05 |
Family
ID=65925913
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811453642.1A Pending CN109583500A (en) | 2018-11-30 | 2018-11-30 | A kind of aesthetic images quality prediction system and method based on depth drift-diffusion method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109583500A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110349673A (en) * | 2019-06-27 | 2019-10-18 | 吉林大学 | A kind of group's physique assessment method based on Gaussian Mixture distribution |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102156817A (en) * | 2011-04-11 | 2011-08-17 | 国家海洋信息中心 | Rapid prediction method for drifting path and diffusion of marine spilled oil |
WO2012104844A2 (en) * | 2011-02-06 | 2012-08-09 | Oplon, Pure Science Ltd | Improved biocidal zeolites |
CN107610123A (en) * | 2017-10-11 | 2018-01-19 | 中共中央办公厅电子科技学院 | A kind of image aesthetic quality evaluation method based on depth convolutional neural networks |
-
2018
- 2018-11-30 CN CN201811453642.1A patent/CN109583500A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2012104844A2 (en) * | 2011-02-06 | 2012-08-09 | Oplon, Pure Science Ltd | Improved biocidal zeolites |
CN102156817A (en) * | 2011-04-11 | 2011-08-17 | 国家海洋信息中心 | Rapid prediction method for drifting path and diffusion of marine spilled oil |
CN107610123A (en) * | 2017-10-11 | 2018-01-19 | 中共中央办公厅电子科技学院 | A kind of image aesthetic quality evaluation method based on depth convolutional neural networks |
Non-Patent Citations (1)
Title |
---|
박태서: "Cognitive analysis of visual aesthetic perception", 《COLLEGE OF HUMANITIES PROGRAM IN COGNITIVE SCIENCE》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110349673A (en) * | 2019-06-27 | 2019-10-18 | 吉林大学 | A kind of group's physique assessment method based on Gaussian Mixture distribution |
CN110349673B (en) * | 2019-06-27 | 2022-08-05 | 吉林大学 | Group constitution evaluation method based on Gaussian mixture distribution |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107330437B (en) | Feature extraction method based on convolutional neural network target real-time detection model | |
CN107122375B (en) | Image subject identification method based on image features | |
CN108875600A (en) | A kind of information of vehicles detection and tracking method, apparatus and computer storage medium based on YOLO | |
CN110490177A (en) | A kind of human-face detector training method and device | |
CN108629326A (en) | The action behavior recognition methods of objective body and device | |
CN112541532B (en) | Target detection method based on dense connection structure | |
CN109919252A (en) | The method for generating classifier using a small number of mark images | |
CN110879982A (en) | Crowd counting system and method | |
CN110070116A (en) | Segmented based on the tree-shaped Training strategy of depth selects integrated image classification method | |
CN109978875A (en) | A kind of capacitor open defect recognition methods and identification device | |
CN111126864A (en) | Street quality assessment method based on man-machine confrontation score | |
CN109376787A (en) | Manifold learning network and computer visual image collection classification method based on it | |
CN109410184A (en) | Live streaming pornographic image detection method based on dense confrontation network semi-supervised learning | |
CN106056059A (en) | Multidirectional SLGS characteristic description and performance cloud weight fusion face recognition method | |
CN110110663A (en) | A kind of age recognition methods and system based on face character | |
CN108256473A (en) | Methods, devices and systems, computer readable storage medium for data amplification | |
CN109903339A (en) | A kind of video group personage's position finding and detection method based on multidimensional fusion feature | |
CN110503078A (en) | A kind of remote face identification method and system based on deep learning | |
CN109919246A (en) | Pedestrian's recognition methods again based on self-adaptive features cluster and multiple risks fusion | |
CN110147833A (en) | Facial image processing method, apparatus, system and readable storage medium storing program for executing | |
CN110008853A (en) | Pedestrian detection network and model training method, detection method, medium, equipment | |
CN112818849A (en) | Crowd density detection algorithm based on context attention convolutional neural network of counterstudy | |
CN111428655A (en) | Scalp detection method based on deep learning | |
Wang et al. | Distortion recognition for image quality assessment with convolutional neural network | |
CN115393666A (en) | Small sample expansion method and system based on prototype completion in image classification |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190405 |