CN111241372B - Method for predicting color harmony degree according to user preference learning - Google Patents

Method for predicting color harmony degree according to user preference learning Download PDF

Info

Publication number
CN111241372B
CN111241372B CN202010050943.0A CN202010050943A CN111241372B CN 111241372 B CN111241372 B CN 111241372B CN 202010050943 A CN202010050943 A CN 202010050943A CN 111241372 B CN111241372 B CN 111241372B
Authority
CN
China
Prior art keywords
color
comments
value
harmony
user
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.)
Active
Application number
CN202010050943.0A
Other languages
Chinese (zh)
Other versions
CN111241372A (en
Inventor
杨柏林
魏天祥
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Gongshang University
Original Assignee
Zhejiang Gongshang University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang Gongshang University filed Critical Zhejiang Gongshang University
Priority to CN202010050943.0A priority Critical patent/CN111241372B/en
Publication of CN111241372A publication Critical patent/CN111241372A/en
Application granted granted Critical
Publication of CN111241372B publication Critical patent/CN111241372B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/951—Indexing; Web crawling techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a method for predicting color harmony degree according to user preference learning. Based on user comments of a large website data set, the method firstly carries out semantic classification on the comments of the users in the data set on the color subjects, and divides the comments into a plurality of different harmony values according to specific keywords. And then extracting a plurality of color themes from the color themes and training in a three-layer back propagation neural network. The result output from the hidden layer generates an aesthetic ability value that is in line with the user's aesthetics through a probability density model based on Kernel distribution. And finally, calculating a final color harmony value through linear calculation. The harmony value estimated by the method has high accuracy due to the consideration of the aesthetic feelings of different users to the colors, and the method can be used in practical application scenes such as color suggestion with color palettes, recoloring of images, color style conversion and the like.

Description

Method for predicting color harmony degree according to user preference learning
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to a method for predicting color harmony degree according to user preference learning.
Background
Color matching can be used in a variety of fields, such as images, posters, apparel, and interior home designs. Where harmonious color matching is a determining factor for the popularity of the design. Designers often use accepted harmonious color matching empirically in the design, some of which are proven to be effective in color harmony theory, e.g., the color matching at a specific position in the color wheel is harmonious, and harmonious color matching such as similar colors, contrasting colors, three roles, etc. is established on the basis of the harmonious color matching. But in most human eyes they lack the eye-light and theoretical knowledge of artists, and therefore creating and evaluating a set of harmonious color themes becomes very difficult.
With the continuous rise of multimedia interaction technology, in recent years, online websites for constructing Color themes, such as Adobe Color CC and colorerrors, appear, and users can build and release self-designed Color palettes or patterns in the online websites. These services are useful to novices and designers because they can use others' palette designs as a reference and also customize palettes to suit personal aesthetics and taste. The websites are also attached with evaluation systems, and users can evaluate the color theme of other people to judge whether the color theme accords with the aesthetic preference of the users. But such an assessment is ambiguous and one cannot learn the true degree of harmony from these palettes.
Recently, O' Donovan collected the evaluation of 10743 different color topics by Amazon Mechanical turn (MTurk) and quantified the harmony of the different color topics between 1-5 points in numerical form by Lasso regression training the predictive model. Based on the method, Yang adopts a two-layer Maximum Likelihood Estimation (MLE) method and a Back Propagation Neural Network (BPNN) method to enable the prediction result to be more accurate. However, the data set adopted by the users is obtained by a large online website crawler, but the evaluation of the color subjects by the users is integrated into an evaluation by utilizing linear calculation, namely, each color subject in the data set has a unique color harmony value corresponding to the color subject. Although such a data set can be trained quickly when used as a label for a supervised model, the influence of the preferences of different users on the harmony degree of colors is not considered, so that the prediction models of the users are inaccurate.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a method for predicting the color harmony degree according to the learning of user preference.
The technical scheme adopted by the invention for solving the technical problem comprises the following steps:
a first part: data acquisition and preprocessing
The method comprises the following steps: and performing online crawler to acquire color theme id of all comments exceeding a set person, user comment id and specific comments in a certain website.
Step two: and carrying out semantic classification on the acquired comment information. And searching keywords in the user comments, and judging the harmonious degree to be expressed by the comments according to the keywords.
Step three: and deleting the blank comments and the user ids of the comments, numbering according to the user name sequence, and obtaining the id of the color theme, the user id and the harmony value converted from the comments.
A second part: constructing color harmony prediction model
The method comprises the following steps: color harmony prediction model constructed based on back propagation neural network
1-1: and extracting a plurality of color features including palette colors, average values, standard deviations, median values, maximum values, minimum values, mode numbers, color moments, maximum-minimum values of single channels in each color space and Euclidean distances of adjacent colors in the theme, and training the color features as an input layer of a back propagation neural network.
1-2: and training the color characteristics of the input layer through the three hidden layers.
1-3: a sigmod function is adopted as an activation function of the three hidden layers, and a gradient descent algorithm is adopted as a loss function of the neural network back propagation.
Step two: and constructing a Kernel probability distribution model based on user preference between a hidden layer and an output layer of the back propagation neural network.
Suppose that M users evaluate N color themes, and the color characteristics of the color theme are CnIndicating that the mth user gives the evaluation Z of the color thememnIs represented by ZmnIs a real tag, and ZmnIs less than M × N.
2-1: extension CnIs a vector of length M.
2-2: converting the color feature vector into color feature beta by using the back propagation neural network of the step onemnWherein, βmn∈(0,1)。
2-3: according to the real label ZmnAnd color characteristic betamnPreliminarily generate aesthetic ability value alpha'mn：
Figure BDA0002371164980000031
Where K represents a color rating with different color themes.
2-4: analyzing aesthetic ability value α'mnObtaining itProbability density function p (x | K, h) and cumulative distribution function F (x | K, h) based on Kernel distribution:
2-5: generating aesthetic ability values alpha corresponding to users one by one according to the cumulative distribution function F (x | K, h)m：
Figure BDA0002371164980000032
Wherein x 'is an aesthetic ability value x' { α1,α2,…,αmU is a random variable satisfying uniform distribution.
2-6: by alphamAnd betamnCalculating a harmony value L of the final color thememn：
Figure BDA0002371164980000033
Step three: comparison LmnAnd ZmnCalculating the Loss value and returning to the first step for reverse iteration until the Loss value is not changed any more.
Step four: outputting the iterated harmony value LmnAnd takes it as the final prediction result.
The invention has the beneficial effects that: the invention considers the color preference of different users, analyzes the comment of the user as the color theme through the probability density model based on Kernel distribution, quantifies the color preference of the user into the aesthetic capability value, and considers the color preference as an important factor capable of influencing color harmony. Compared with the prior method, the method has higher accuracy and can better reflect the aesthetic quality of the public to the color.
Detailed Description
The method quantifies the aesthetic passing models of different users for colors, finally reflects the preference of each user for colors through the aesthetic ability value, the user can input the color themes of five different colors with the same area, and the harmony degree of the color theme can be generated through the method and is represented by numerical values.
A first part: data acquisition and preprocessing
The method comprises the following steps: and performing online crawler in an online website COLOURLovers, and acquiring color theme id of more than 5 people for all comments in the website, the id of the user comments and specific comments.
Step two: and carrying out semantic classification on the acquired comment information. And searching whether keywords of the selected positive words or negative words exist in the user comment, and judging the harmonious degree of the comment to be expressed according to the keywords. For reviews that do not contain keywords, their harmony value is listed as 4. Table 1 shows detailed classification of comment keywords and harmony degree.
TABLE 1 mapping of color topic review keywords to harmonious degrees
Figure BDA0002371164980000041
Step three: blank comments are deleted, along with the user id of the comment, for reducing noise of the data set. And meanwhile, numbering according to the user name sequence to obtain the id of the color theme, the id of the user and the harmony value converted from the comment.
A second part: constructing color harmony prediction model
The method comprises the following steps: color harmony prediction model constructed based on back propagation neural network
1-1: 184 color features are extracted, including palette color, mean, standard deviation, median, maximum, minimum, mode, color moment, max-min of a single channel in each color space, and euclidean distance of adjacent colors in the theme, i.e., RGB, Lab, HSV, and LCH. These color features are trained as the input layer of the back propagation neural network.
1-2: the color characteristics of the input layer are trained through the three hidden layers, and the number of nodes of each hidden layer is determined through multiple experiments. The first layer of hidden layer has 300 neurons, the second layer of hidden layer has 20 neurons, and the third layer of hidden layer has 1 neuron.
1-3: using sigmod function as three layersThe activation function of the hidden layer, the gradient descent algorithm as the loss function of the neural network back propagation. The learning rate η is set to 0.00075 to balance the prediction accuracy and time consumption, and gradually decreases as the number of iterations increases (η)nextη × 0.99) to ensure that the model remains stable late in the iteration.
Step two: and constructing a Kernel probability distribution model based on user preference between a hidden layer and an output layer of the back propagation neural network.
If there are M users evaluating N color themes, the color characteristics of the color theme may be CnIndicating that the mth user gives the evaluation Z of the color thememnIs represented by, and ZmnIs less than M × N.
2-1: to ensure consistency of back-propagating neural network inputs and outputs, C is extendednIs a vector of length M
Figure BDA0002371164980000051
2-2: utilizing the back propagation neural network of the step one to carry out color feature vector
Figure BDA0002371164980000052
Conversion to betamn. Wherein, betamn∈(0,1)。
2-3: according to the real label ZmnAnd color characteristic betamnPreliminarily generate aesthetic ability value alpha'mn：
Figure BDA0002371164980000053
Where K represents seven different color ratings of the color theme.
2-4: analyzing aesthetic ability value α'mnObtaining a probability density function p (xK, h) and a cumulative distribution function F (x | K, h) based on Kernel distribution:
Figure BDA0002371164980000054
Figure BDA0002371164980000055
Figure BDA0002371164980000056
Figure BDA0002371164980000057
Figure BDA0002371164980000058
|x|≤1
wherein x is s aesthetic ability values x ═ α satisfying the Kernel distribution1,α2,…,αsσ is the standard deviation of x, h is the bandwidth parameter in the Kernel distribution function, K (x; h) is the Epanechnikov scaling Kernel function used in the model, and W (x) is the cumulative distribution function of K (x; h).
2-5: generating aesthetic ability values alpha corresponding to users one by one according to F (x | K, h)m：
Figure BDA0002371164980000061
Wherein x 'is an aesthetic ability value x' { α1,α2,…,αmU is a random variable u-Unif [0,1 ] satisfying uniform distribution]。
2-6: by alphamAnd betamnCalculating a harmony value L of the final color thememn：
Figure BDA0002371164980000062
Step three: comparison LmnAnd ZmnCalculating the Loss value and returning to the step oneReverse iteration is performed until the Loss value no longer changes.
Step four: outputting the iterated harmony value LmnAnd takes it as the final prediction result.
In order to verify the effect of the invention, the experimental result is measured by adopting a Mean Square Error (MSE) and Mean Absolute Error (MAE) method under the same data set, and the result of the invention is compared with the result of O' Donovan and the result of Yang. The comparison results are shown in the following table:
TABLE 2 comparison of different color harmony models using MSE and MAE
Figure BDA0002371164980000063

Claims (3)

1. A method of learning a predicted color harmony measure from user preferences, the method comprising the steps of:
a first part: data acquisition and preprocessing
The method comprises the following steps: performing online crawler to obtain color subject id of all comments in a certain website, id of user comments and specific comments, wherein the number of the comments exceeds that of a set person;
step two: carrying out semantic classification on the obtained comment information; searching key words in the user comments, and judging the harmonious degree to be expressed by the comments according to the key words;
step three: deleting blank comments and the user ids of the comments, numbering according to user name sequencing, and obtaining the id of the color theme, the user ids and the harmony value converted from the comments;
a second part: constructing color harmony prediction model
The method comprises the following steps: color harmony prediction model constructed based on back propagation neural network
1-1: extracting a plurality of color features, including palette colors, an average value, a standard deviation, a median, a maximum value, a minimum value, a mode, a color moment, a maximum-minimum value of a single channel in each color space and Euclidean distances of adjacent colors in a theme, and training the color features as an input layer of a back propagation neural network;
1-2: training the color characteristics of the input layer through the three hidden layers;
1-3: a sigmod function is adopted as an activation function of the three hidden layers, and a gradient descent algorithm is adopted as a loss function of the neural network back propagation;
step two: constructing a Kernel probability distribution model based on user preference between a hidden layer and an output layer of the back propagation neural network;
suppose that M users evaluate N color themes, and the color characteristics of the color theme are CnIndicating that the mth user gives the evaluation Z of the color thememnIs represented by ZmnIs a real tag, and ZmnIs less than mxn;
2-1: extension CnIs a vector of length M;
2-2: converting the color feature vector into color feature beta by using the back propagation neural network of the step onemnWherein, βmn∈(0,1)；
2-3: according to the real label ZmnAnd color characteristic betamnPreliminarily generate aesthetic ability value alpha'mn：
Figure FDA0003162717380000021
Wherein K represents a color rating, the color theme is different, and the color rating is also different;
2-4: analyzing aesthetic ability value α'mnAcquiring a probability density function p (x | K, h) and a cumulative distribution function F (x | K, h) based on Kernel distribution, wherein h is a bandwidth parameter in the Kernel distribution function;
2-5: generating aesthetic ability values alpha corresponding to users one by one according to the cumulative distribution function F (x | K, h)m：
Figure FDA0003162717380000022
Wherein x 'is an aesthetic ability value, and x' { α ═ α1,α2,…,αmU is a random variable satisfying uniform distribution;
2-6: by alphamAnd betamnCalculating a harmony value L of the final color thememn：
Figure FDA0003162717380000023
Step three: comparison LmnAnd ZmnCalculating the Loss value and returning to the first step for reverse iteration until the Loss value is not changed any more;
step four: outputting the iterated harmony value LmnAnd takes it as the final prediction result.
2. The method of claim 1, wherein:
in data acquisition and preprocessing:
the website in the first step is an online website COLOURLOVers, and color theme id of all people with more than 5 comments in the website, the id of the user comments and specific comments are obtained;
and the keywords in the step two refer to positive words or negative words, and if the comments do not contain the keywords, the corresponding harmonic value is the middle value of the harmonic value.
3. The method of claim 1, wherein:
in the construction of the color harmony prediction model:
1-2: training the color characteristics of the input layer through the three hidden layers, and determining the number of nodes of each hidden layer by utilizing multiple experiments; the first layer of hidden layer has 300 neurons, the second layer of hidden layer has 20 neurons, and the third layer of hidden layer has 1 neuron;
1-3: setting the learning rate eta to 0.00075 to balance the prediction accuracy and the time consumption.
CN202010050943.0A 2020-01-17 2020-01-17 Method for predicting color harmony degree according to user preference learning Active CN111241372B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010050943.0A CN111241372B (en) 2020-01-17 2020-01-17 Method for predicting color harmony degree according to user preference learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010050943.0A CN111241372B (en) 2020-01-17 2020-01-17 Method for predicting color harmony degree according to user preference learning

Publications (2)

Publication Number Publication Date
CN111241372A CN111241372A (en) 2020-06-05
CN111241372B true CN111241372B (en) 2021-11-05

Family

ID=70880906

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010050943.0A Active CN111241372B (en) 2020-01-17 2020-01-17 Method for predicting color harmony degree according to user preference learning

Country Status (1)

Country Link
CN (1) CN111241372B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113253961B (en) * 2021-06-02 2021-12-03 广州朗国电子科技股份有限公司 Electronic blackboard control method and system, electronic blackboard and readable medium

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102799669A (en) * 2012-07-17 2012-11-28 杭州淘淘搜科技有限公司 Automatic grading method for commodity image vision quality
WO2014070914A1 (en) * 2012-10-30 2014-05-08 Stylyze Llc Automated color processing and selection platform
CN108492294A (en) * 2018-03-23 2018-09-04 北京邮电大学 A kind of appraisal procedure and device of image color harmony degree
CN108961346A (en) * 2018-08-08 2018-12-07 浙江工商大学 The method that color harmony degree is predicted based on BP neural network
CN109389591A (en) * 2018-09-30 2019-02-26 西安电子科技大学 Color image quality evaluation method based on colored description

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10409822B2 (en) * 2014-05-06 2019-09-10 Shutterstock, Inc. Systems and methods for presenting ranked search results

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102799669A (en) * 2012-07-17 2012-11-28 杭州淘淘搜科技有限公司 Automatic grading method for commodity image vision quality
WO2014070914A1 (en) * 2012-10-30 2014-05-08 Stylyze Llc Automated color processing and selection platform
CN108492294A (en) * 2018-03-23 2018-09-04 北京邮电大学 A kind of appraisal procedure and device of image color harmony degree
CN108961346A (en) * 2018-08-08 2018-12-07 浙江工商大学 The method that color harmony degree is predicted based on BP neural network
CN109389591A (en) * 2018-09-30 2019-02-26 西安电子科技大学 Color image quality evaluation method based on colored description

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
An EL-LDA based general color harmony model for photo aesthetics assessment;Peng Lu等;《Signal Processing》;20141218;第120卷;第731-745页 *
基于颜色和谐度的可计算图像美学评估关键技术研究与应用;邝智杰;《中国优秀硕士学位论文全文数据库信息科技辑》;20150815;全文 *

Also Published As

Publication number Publication date
CN111241372A (en) 2020-06-05

Similar Documents

Publication Publication Date Title
CN109960759B (en) Recommendation system click rate prediction method based on deep neural network
CN109948165B (en) Fine granularity emotion polarity prediction method based on mixed attention network
CN108363753B (en) Comment text emotion classification model training and emotion classification method, device and equipment
CN108510012B (en) Target rapid detection method based on multi-scale feature map
CN107066445B (en) The deep learning method of one attribute emotion word vector
CN109657156A (en) A kind of personalized recommendation method generating confrontation network based on circulation
Guo et al. Emotional design method of product presented in multi-dimensional variables based on Kansei Engineering
CN111127146B (en) Information recommendation method and system based on convolutional neural network and noise reduction self-encoder
CN109101938B (en) Multi-label age estimation method based on convolutional neural network
CN111062775A (en) Recommendation system recall method based on attention mechanism
CN108446676B (en) Face image age discrimination method based on ordered coding and multilayer random projection
Lin et al. Intelligent decision support for new product development: a consumer-oriented approach
Hu et al. A quantitative aesthetic measurement method for product appearance design
CN110321805A (en) A kind of dynamic expression recognition methods based on sequential relationship reasoning
Wu et al. Product design award prediction modeling: Design visual aesthetic quality assessment via DCNNs
Wang et al. Clustering-based collaborative filtering for link prediction
CN114937182B (en) Image emotion distribution prediction method based on emotion wheel and convolutional neural network
CN111241372B (en) Method for predicting color harmony degree according to user preference learning
Yang et al. A Color‐Pair Based Approach for Accurate Color Harmony Estimation
CN111400525A (en) Intelligent fashionable garment matching and recommending method based on visual combination relation learning
Xue et al. A method of the product form design and color design of train seats based on GRA-Fuzzy theory
CN110188791A (en) Based on the visual emotion label distribution forecasting method estimated automatically
Chen et al. Mutual information-based dropout: Learning deep relevant feature representation architectures
Tan et al. An emotion index estimation based on facial action unit prediction
Yang Clothing design style recommendation using decision tree algorithm combined with deep learning

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