CN110222722A - Interactive image stylization processing method, calculates equipment and storage medium at system - Google Patents
Interactive image stylization processing method, calculates equipment and storage medium at system Download PDFInfo
- Publication number
- CN110222722A CN110222722A CN201910396504.2A CN201910396504A CN110222722A CN 110222722 A CN110222722 A CN 110222722A CN 201910396504 A CN201910396504 A CN 201910396504A CN 110222722 A CN110222722 A CN 110222722A
- Authority
- CN
- China
- Prior art keywords
- image
- network
- training
- sub
- target
- 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
- G06F18/214—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
- G06F18/23—Clustering 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/60—Editing figures and text; Combining figures or text
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of interactive image stylization processing method, system, calculate equipment and storage medium, which comprises obtain multiple first training images and multiple second training images;It constructs and scratches figure network, adaptivenon-uniform sampling is carried out to the target and background of the first training image, obtains first object image;Stylized neural network is constructed, first object image and the second training image are inputted into stylized neural network, stylized neural network is trained;According to interactively entering for user, adaptivenon-uniform sampling is carried out using the target and background of stingy figure network handles processing image, obtains the second target image;Stylized processing is carried out to the second target image using the stylized neural network after training, obtains final stylized image.The present invention according to the needs of users, can interactively carry out the image stylization in specified region, simulate the drafting style of different art forms, enhance the form of expression of visual information in image, promote the attraction of image.
Description
Technical field
The present invention relates to a kind of interactive image stylization processing method, system, equipment and storage medium are calculated, belonged to
In image procossing and rendering field.
Background technique
Graphic Arts stylization is non-photorealistic rendering (Non-in computer graphics in recent years
Photorealistic Rendering, NPR) main research, it using computer as tool, with algorithm simulation go out difference
The drafting style of art form enhances the form of expression of visual information in image, attracts the attention of people.From last century 90 years
In generation, feeling of unreality is rendered to be rapidly developed as an individual branches of graphics so far.The range of design is also no longer
Simple area of computer aided artistic painting, also relates to the industries such as film, education, Entertainment, cartoon.
The stylization processing of image is increasingly liked there are many image processing softwares in the market by everybody, wherein
The user of Photoshop is most, professional most strong.User passes through some filters (such as watercolor, poster included in Photoshop
Edge, splash and texture) corresponding stylization processing can be carried out to image.With the appearance of mobile phone A ndroid and ios system,
Handset image processing software is also liked by everybody.360camera, Baidu's evil spirit figure, a plurality of handset image processing such as Meitu Xiu Xiu
Software also joined various image stylization effects, such as HDR, pseudo-classic, sketch etc..
However the image stylization filter of current image processing software existing on the market and relevant research are substantially all
It is stylized processing to be carried out to general image and treatment effect is bad, user is thus caused not only cannot independently to select to need in image
The part rendered, and rendering effect is often unable to reach expection, also needs to carry out a series of subsequent processings.In addition, mesh
There is the problems such as inefficiency, poor robustness mostly in preceding common image stylization algorithm.
Summary of the invention
In view of this, the present invention provides a kind of interactive image stylization processing method, system, calculating equipment and storage
Medium according to the needs of users, can interactively carry out the image stylization in specified region, simulate different art forms
Drafting style, enhance image in visual information the form of expression, the attraction of image is promoted, in multimedia, art teaching etc.
Field has practical value.
The first purpose of this invention is to provide a kind of interactive image stylization processing method.
Second object of the present invention is to provide a kind of interactive image stylization processing system.
Third object of the present invention is to provide a kind of calculating equipment.
Fourth object of the present invention is to provide a kind of storage medium.
The first purpose of this invention can be reached by adopting the following technical scheme that:
A kind of interactive image stylization processing method, which comprises
Obtain multiple first training images and multiple second training images;Wherein, first training image is to include mesh
The image of mark and background, second training image are style image;
It constructs and scratches figure network, adaptivenon-uniform sampling is carried out to the target and background of the first training image, obtains first object figure
Picture;
Stylized neural network is constructed, first object image and the second training image are inputted into stylized neural network, it is right
Stylized neural network is trained;
According to interactively entering for user, the target and background using stingy figure network handles processing image is adaptively divided
It cuts, obtains the second target image;
Stylized processing is carried out to the second target image using the stylized neural network after training, obtains final stylization
Image.
Further, described construct scratches figure network, carries out adaptivenon-uniform sampling to the target and background of the first training image, obtains
To first object image, specifically include:
In RGB color, the first training is schemed respectively using the full covariance mixed Gauss model of K Gaussian component
The target and background of picture is modeled;
The target and background of the first training image is partitioned into using iteration Energy minimization;Wherein, the energy is most
Smallization algorithm is initialized by K-means algorithm, respectively by belong to target and background pixel cluster be K class, K class it is equal
Value and covariance are estimated to obtain by belonging to the rgb value of the pixel of target and background;
According to the target and background of the first training image, a figure is established, weight figure is handled by maximum-flow algorithm, is used in combination
Minimal cut algorithm is split;
Iteration optimization mixed Gauss model and segmentation result obtain first object image.
Further, the gibbs energy of first training image is as follows:
E (α, k, θ, z)=U (α, k, θ, z)+V (α, z)
Wherein:
α is the label of pixel, and 0 is background, and 1 is target, and 2 be possible background, and 3 be possible target;K=1 ... K, table
Show the classification of K-means algorithm cluster, θ is the weight of Gaussian component, and z is the image data of the first training image;
U is the area item of energy function, is indicated in the image data of the first training image, a pixel is classified as mesh
The punishment of mark or background:
According to gauss hybrid models, negative logarithm is taken to obtain the probability that some pixel belongs to target or background:
Wherein, π is the weight that single Gaussian component contributes probability, and μ is the mean vector of each Gaussian component, and ∑ is association
Variance matrix;
V is boundary energy item, indicates discontinuous punishment between neighborhood territory pixel m and n, the boundary energy item is in rgb space
In, it is as follows for measuring the similitude of neighborhood territory pixel m and n:
Wherein, β=(2 < (zm-zn)2>)-1, β determines by the contrast of the first training image, and γ is constant, and C is adjacent face
The set of color pair.
Further, after the iteration optimization mixed Gauss model and segmentation result, further includes: use border
Matting carries out the boundary of segmentation smooth.
Further, first object image and the second training image are inputted style by the stylized neural network of the building
Change neural network, stylized neural network be trained, is specifically included:
It establishes and generates network;Wherein, the generation network is by encoder sub-network, AdaIN sub-network and decoder subnet
Network composition, encoder sub-network utilize several layer buildings before pre-training VGG19 network;
By first object image and the second training image input coding device sub-network, the output of encoder sub-network is obtained,
Input by the output of encoder sub-network as AdaIN sub-network, obtains the output of AdaIN sub-network, by AdaIN sub-network
Input of the output as decoder sub-network, by decoder sub-network output stylization image;
Using encoder sub-network as loss network is calculated, according to the stylized image of decoder sub-network output, AdaIN
The output of sub-network and the second training image calculate loss;
Using back-propagation algorithm, the optimal way finely tuned using level, by calculating transmitting gradient from the last layer, by
Layer transmitting, updates all parameters.
Further, described by first object image and the second training image input coding device sub-network, obtain encoder
The output of sub-network, the input by the output of encoder sub-network as AdaIN sub-network, obtains the output of AdaIN sub-network,
Input by the output of AdaIN sub-network as decoder sub-network exports stylized image by decoder sub-network, specific to wrap
It includes:
By first object image and the second training image input coding device sub-network, in feature space to first object image
It is encoded with the second training image, the corresponding fisrt feature mapping of output first object image is corresponding with the second training image
Second feature mapping;
The fisrt feature mapping and second feature mapping input AdaIN sub-network that encoder sub-network is exported, export mesh
Mark Feature Mapping;
The target signature that AdaIN sub-network is exported maps input decoder sub-network, is converted by decoder sub-network
To image space, stylized image is exported.
Further, the loss is made of content loss and style loss, as follows:
L=Lc+λLs
Wherein, LcFor content loss, LsFor style loss, λ is weight;
The content loss LcBetween the stylized image exported for the output of AdaIN sub-network and decoder sub-network
Euclidean distance is as follows:
Lc=| | f (g (t))-t | |2
Wherein, t is the output of AdaIN sub-network, and g (t) is the stylized image of decoder sub-network output;
The style loss is as follows:
Wherein, σ is standard deviation, and μ is average value,For a wherein layer network for pre-training VGG19 network, s is the second instruction
Practice the image data of image.
Second object of the present invention can be reached by adopting the following technical scheme that:
A kind of interactive image stylization processing system, the system comprises:
Image collection module, for obtaining multiple first training images and multiple second training images;Wherein, described first
Training image is the image comprising target and background, and second training image is style image;
Figure network struction module is scratched, for constructing stingy figure network, the target and background of the first training image is carried out adaptive
It should divide, obtain first object image;
Stylized neural network constructs module, and for constructing stylized neural network, first object image and second are instructed
Practice image and input stylized neural network, stylized neural network is trained;
Module is scratched, for interactively entering according to user, utilizes the target and background of stingy figure network handles processing image
Adaptivenon-uniform sampling is carried out, the second target image is obtained;
Stylized processing module, for carrying out stylization to the second target image using the stylized neural network after training
Processing obtains final stylized image.
Third object of the present invention can be reached by adopting the following technical scheme that:
A kind of calculating equipment, including processor and for the memory of storage processor executable program, the processing
When device executes the program of memory storage, above-mentioned interactive image stylization processing method is realized.
Fourth object of the present invention can be reached by adopting the following technical scheme that:
A kind of storage medium is stored with program, when described program is executed by processor, realizes above-mentioned interactive image wind
It formats processing method.
The present invention have compared with the existing technology it is following the utility model has the advantages that
The present invention scratches figure network and stylized neural network by constructing, can interactively entering according to user, using scratching
The target and background that figure network handles handle image carries out adaptivenon-uniform sampling, obtains the target image that user specifies region, then benefit
Stylized processing is carried out to the target image with stylized neural network, so that the drafting style of different art forms is simulated,
The form of expression for enhancing visual information in image, promotes the attraction of image, has in fields such as multimedia, art teachings practical
Value.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
The structure shown according to these attached drawings obtains other attached drawings.
Fig. 1 is the flow chart of the interactive image stylization processing method of the embodiment of the present invention 1.
Fig. 2 is that the flow chart of figure network is scratched in the building of the embodiment of the present invention 1.
Fig. 3 is the building of the embodiment of the present invention 1 and the flow chart of the stylized neural network of training.
Fig. 4 is the schematic diagram of the image to be processed of the embodiment of the present invention 1.
Fig. 5 is that the user of the embodiment of the present invention 1 interactively enters the schematic diagram of image to be processed.
Fig. 6 is the image schematic diagram that the stingy figure network of the embodiment of the present invention 1 exports.
Fig. 7 is the schematic diagram of the reference style image of the embodiment of the present invention 1.
Fig. 8 is the image schematic diagram that the stylized neural network of the embodiment of the present invention 1 exports.
Fig. 9 is the structural block diagram of the interactive image stylization processing system of the embodiment of the present invention 2.
Figure 10 is the structural block diagram of the stingy figure network struction module of the embodiment of the present invention 2.
Figure 11 is that the stylized neural network of the embodiment of the present invention 2 constructs the structural block diagram of module.
Figure 12 is the structural block diagram of the calculating equipment of the embodiment of the present invention 3.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiments of the present invention, instead of all the embodiments, based on the embodiments of the present invention, ordinary skill people
Member's every other embodiment obtained without creative efforts, shall fall within the protection scope of the present invention.
Embodiment 1:
As shown in Figure 1, present embodiments providing a kind of interactive image stylization processing method, this method includes following step
It is rapid:
S101, multiple first training images and multiple second training images are obtained.
In this step, the first training image is the image comprising target (prospect) and background, and the second training image is wind
Table images, the target of the present embodiment are face.
First training image and the second training image can be obtained by acquisition, such as include target by camera shooting
With the image of background, and shooting style image, can also be obtained from database lookup, such as store up wrap in databases in advance
Image and style image containing target and background, searching for from database can be obtained.
S102, it constructs and scratches figure network, adaptivenon-uniform sampling is carried out to the target and background of the first training image, obtains the first mesh
Logo image.
In this step, adaptivenon-uniform sampling is carried out to the target and background of the first training image and uses Grabcut algorithm,
The target of the first training image is extracted, by Grabcut algorithm to obtain first object image.
As shown in Fig. 2, step S102 is specifically included:
S1021, in RGB color, using the full covariance mixed Gaussian mould of a Gaussian component of K (one take as K=5)
Type (Gaussian Mixed Model, abbreviation GMM) respectively models the target and background of the first training image.
Gibbs (Gibbs) energy of first training image is as follows:
E (α, k, θ, z)=U (α, k, θ, z)+V (α, z)
Wherein:
α is the label of pixel, and 0 is background, and 1 is target, and 2 be possible background, and 3 be possible target;K=1 ... K, table
Show the classification of K-means algorithm cluster, θ is the weight of Gaussian component, and z is the image data of the first training image;
U is the area item of energy function, is indicated in the image data of the first training image, a pixel is classified as mesh
The punishment of mark or background:
According to gauss hybrid models, negative logarithm is taken to obtain the probability that some pixel belongs to target or background:
Wherein, π is the weight that single Gaussian component contributes probability, and μ is the mean vector of each Gaussian component (because having
Tri- channels RGB, therefore be three element vectors), ∑ is covariance matrix (because there is tri- channels RGB, therefore being 3x3 matrix);
V is boundary energy item, indicates discontinuous punishment between neighborhood territory pixel m and n, the boundary energy item is in rgb space
In, it is as follows for measuring the similitude of neighborhood territory pixel m and n:
Wherein, β=(2 < (zm-zn)2>)-1, β is determined by the contrast of the first training image, so that V can be in contrast height
Or it is low in the case where can work normally, γ is constant, take based on experience value 50, C be adjacent color pair set.
S1022, the target and background that the first training image is partitioned into using iteration Energy minimization.
In this step, Energy minimization is initialized by K-means algorithm, will belong to target and back respectively
The pixel cluster of scape is K class, and the mean value and covariance of K class are estimated to obtain by belonging to the rgb value of the pixel of target and background.
S1023, according to the target and background of the first training image, establish a figure, pass through max-flow (max-flow) calculate
Method handles weight figure, and is split with minimal cut (min-cut) algorithm.
In this step, weight figure is handled by maximum-flow algorithm, and is split with minimal cut algorithm, such as following formula:
S1024, iteration optimization mixed Gauss model and segmentation result, obtain first object image.
Step S1021~S1023 is repeated, until convergence, each iteration can alternatively optimize mixed Gauss model and divide
It cuts as a result, border matting can also be used to segmentation after iteration optimization mixed Gauss model and segmentation result
Boundary carries out the post-processings such as smooth, to obtain the first object image of more preferable effect.
S103, stylized neural network is constructed, first object image and the second training image is inputted into stylized nerve net
Network is trained stylized neural network.
In this step, stylized neural network is trained based on Tensorflow platform, is based on arbitrary image wind
Lattice fast transferring (Arbitrary Style Transfer in Real-time), in conjunction with the second training image to first object
Image carries out stylization, realizes the task of arbitrary image stylization, and by attempting to adjust ginseng, multiple training pattern ties experiment
Fruit has carried out further optimization.
Stylized neural network is broadly divided into two parts: generating network and calculates loss network;Generating network is one
Feedforward network, later period are used to carry out style switching network;What calculating loss network was constrained when being for training.
As shown in figure 3, step S103 is specifically included:
S1031, generation network is established.
In this step, it generates network to be made of encoder sub-network, AdaIN sub-network and decoder sub-network, encode
Device sub-network utilizes several layers of (until relu4_1) before pre-training VGG19 network.
S1032, by first object image and the second training image input coding device sub-network, obtain encoder sub-network
Output, the input by the output of encoder sub-network as AdaIN sub-network, obtains the output of AdaIN sub-network, by AdaIN
Input of the output of sub-network as decoder sub-network exports stylized image by decoder sub-network.
In this step, the specific processing of encoder sub-network, AdaIN sub-network and decoder sub-network is as follows:
1) encoder sub-network: encoder sub-network has an encoder, is denoted as f, and first object image is content graph
Picture, is denoted as c, and the second training image is style image, is denoted as s, by content images c and style image s input coding device sub-network,
Content images c and style image s are encoded in feature space, the corresponding fisrt feature mapping of output content images c
(feature map) f (c) second feature corresponding with style image s maps (feature map) f (s).
2) AdaIN sub-network: fisrt feature mapping f (c) that encoder sub-network is exported and second feature mapping f (s)
AdaIN sub-network is inputted, output target signature maps (feature map) t;Specifically, content images c is normalized,
The mesh in the every channel style image s is matched by being aligned mean value and the variance of the target signature mapping in every channel of content images c
Mark the mean value and variance of Feature Mapping:
T=AdaIN (f (c), f (s))
Wherein,
Wherein, σ is standard deviation, and μ is average value.
3) decoder sub-network: decoder sub-network has a decoder, is denoted as g, the mesh that AdaIN sub-network is exported
Feature Mapping t input decoder sub-network is marked, image space is transformed by decoder g, is exported stylization image T (c, s):
T (c, s)=g (t)
Decoder sub-network is generally used to be needed with the symmetrical network structure of encoder sub-network, its weight parameter information
Be trained, start can random some initiation parameters, parameter constantly can be updated by gradient decline so that entire
Loss function is smaller, network is gradually restrained.
S1033, network is lost using encoder sub-network as calculating, is schemed according to the stylization of decoder sub-network output
The output and style image of picture, AdaIN sub-network calculate loss.
Specifically, according to stylized image g (t), target signature map t and style image s, calculate loss L, loss L by
Content loss and style loss composition, as follows:
L=Lc+λLs
Wherein, LcFor content loss, LsFor style loss, λ is weight, and the weight selected when training is 2.0;
Content loss LcThe Euclidean distance between t and stylization image g (t) is mapped for target signature, as follows:
Lc=| | f (g (t))-t | |2
Style loses LsIt is as follows:
Wherein, σ is standard deviation, and μ is average value,For a wherein layer network for pre-training VGG19 network, identical power is selected
Relu1_1, relu2_1, relu3_1 and relu4_1 of weight.
S1034, it is transmitted using the optimal way of level fine tuning by being calculated from the last layer using back-propagation algorithm
Gradient is successively transmitted, updates all parameters.
Above-mentioned steps S101~S103 is the training stage, that is, scratches the building stage of figure network and stylized neural network, connect
Step S104~the S105 to get off is to apply (test) stage.It is appreciated that above-mentioned steps S101~S103 is set in a calculating
Standby (such as computer) is completed, and can enter step application (test) stage of S104~S105 on the computing device, can also
Other calculating equipment are shared with the stingy figure network and stylized neural network that go out this calculating device build, in others
Calculate the application stage that S104~S105 is entered step in equipment.
S104, interactively entering according to user are carried out adaptive using the target and background of stingy figure network handles processing image
It should divide, obtain the second target image.
The realization of this step is similar to above-mentioned steps S102, is that step S1022, user can pass through in place of the main distinction
Direct frame selects target to obtain an initial map-T, and the pixel outside box is all used as background pixel TB, T in boxUPicture
The whole pixel as " may be target " of element, user can guide the target of frame choosing with paintbrush, with to target and background into
Row adaptivenon-uniform sampling.
To TBInterior each pixel n, the label α of initialized pixel nn=0, it is background pixel;
To TUInterior each pixel n, the label α of initialized pixel nn=3, the i.e. pixel as " possible target ";
The mixed Gauss model of target and background is estimated by this pixel.Then with k-mean algorithm respectively belonging to mesh
The pixel cluster of mark and background is K class, i.e., K Gauss model in mixed Gauss model.
According to obtained initial map-T to the Gaussian component in each pixel distributive mixing Gauss model, maximum probability
That is exactly the kth of pixel nnA Gaussian component:
The classification done using each pixel in step S1021 for Gaussian component passes through pixel sample as pixel samples collection
This rgb value estimates its mean parameter and covariance, and the weight of Gaussian component is by belonging to the number of pixels of the Gaussian component
It is determined with the ratio of total number of pixels.
In this step, image to be processed is as shown in figure 5, user interactively enters image to be processed as shown in fig. 6, scratching figure net
The image of network output is as shown in Figure 7.
S105, stylized processing is carried out to the second target image using the stylized neural network after training, obtained final
Stylized image.
Specifically, by the wind after the reference style image chosen in the second target image and the second training image input training
It formats neural network, by the final stylized image of stylized neural network output after training, final stylization image is ginseng
Examine the stylized effect in style image.
In this step, with reference to style image as shown in fig. 7, the image of stylized neural network output is as shown in Figure 8.
The interactive image stylization processing method of the present embodiment can according to the shape of target face and illumination particularity from
It adaptively carries out target (prospect) to extract, so as to formulate in target face the stylization in region;Meanwhile this method can letter
Change the manual operation of facial image fusion, improves the operating efficiency and ease for use of face fusion edit tool.
It will be understood by those skilled in the art that realizing that all or part of the steps in the method for above-described embodiment can pass through
Program is completed to instruct relevant hardware, and corresponding program can store in computer readable storage medium.
It should be noted that this is not although describing the method operation of above-described embodiment in the accompanying drawings with particular order
It is required that hint must execute these operations in this particular order, could be real or have to carry out shown in whole operation
Existing desired result.On the contrary, the step of describing can change and execute sequence.Additionally or alternatively, it is convenient to omit certain steps,
Multiple steps are merged into a step to execute, and/or a step is decomposed into execution of multiple steps.
Embodiment 2:
As shown in figure 9, present embodiments providing a kind of interactive image stylization processing system, which includes that image obtains
Modulus block 901 scratches figure network struction module 902, stylized neural network building module 903, scratches at module 904 and stylization
Module 905 is managed, the concrete function of modules is as follows:
Described image obtains module 901, for obtaining multiple first training images and multiple second training images;Wherein,
First training image is the image comprising target and background, and second training image is style image.
The stingy figure network struction module 902, for constructing stingy figure network, to the target and background of the first training image into
Row adaptivenon-uniform sampling obtains first object image.
The stylization neural network constructs module 903, for constructing stylized neural network, by first object image and
Second training image inputs stylized neural network, is trained to stylized neural network.
The stingy module 904 utilizes the target of stingy figure network handles processing image for interactively entering according to user
Adaptivenon-uniform sampling is carried out with background, obtains the second target image.
It is described stylization processing module 905, for using training after stylized neural network to the second target image into
Row stylization processing obtains final stylized image.
As shown in Figure 10, scratching figure network struction module 902 includes:
Modeling unit 9021 is used in RGB color, using the full covariance mixed Gaussian mould of K Gaussian component
Type respectively models the target and background of the first training image.
First cutting unit 9022, for using iteration Energy minimization be partitioned into the first training image target and
Background;Wherein, the Energy minimization is initialized by K-means algorithm, will belong to target and background respectively
Pixel cluster is K class, and the mean value and covariance of K class are estimated to obtain by belonging to the rgb value of the pixel of target and background.
Second cutting unit 9023 establishes a figure, passes through maximum for the target and background according to the first training image
Flow algorithm handles weight figure, and is split with minimal cut algorithm.
Iteration unit 9024 is used for iteration optimization mixed Gauss model and segmentation result, obtains first object image.
As shown in figure 11, stylized neural network building module 903 includes:
Unit 9031 is established, generates network for establishing;Wherein, the generation network is by encoder sub-network, AdaIN
Network and decoder sub-network composition, encoder sub-network utilize several layer buildings before pre-training VGG19 network.
Processing unit 9032, for being compiled first object image and the second training image input coding device sub-network
The output of code device sub-network, the input by the output of encoder sub-network as AdaIN sub-network, obtains AdaIN sub-network
Output, the input by the output of AdaIN sub-network as decoder sub-network export stylized image by decoder sub-network.
Costing bio disturbance unit 9033, for losing network for encoder sub-network as calculating, according to decoder sub-network
The stylized image of output, the output of AdaIN sub-network and the second training image calculate loss.
Optimize unit 9034, for use back-propagation algorithm, using level finely tune optimal way, by from last
Layer calculates transmitting gradient, successively transmits, updates all parameters.
The specific implementation of modules may refer to above-described embodiment 1 in the present embodiment, and this is no longer going to repeat them;It needs
Illustrate, device provided in this embodiment only the example of the division of the above functional modules, in practical applications,
It can according to need and be completed by different functional modules above-mentioned function distribution, i.e., internal structure is divided into different functions
Module, to complete all or part of the functions described above.
Embodiment 3:
A kind of calculating equipment is present embodiments provided, which can be computer, as shown in figure 12 comprising logical
Processor 1202, memory, input unit 1203, display 1204 and the network interface 1205 of the connection of system bus 1201 are crossed,
The processor is calculated for offer and control ability, the memory include non-volatile memory medium 1206 and built-in storage
1207, which is stored with operating system, computer program and database, which is
The operation of operating system and computer program in non-volatile memory medium provides environment, and processor 1202 executes memory and deposits
When the computer program of storage, the interactive image stylization processing method of above-described embodiment 1 is realized, as follows:
Obtain multiple first training images and multiple second training images;Wherein, first training image is to include mesh
The image of mark and background, second training image are style image;
It constructs and scratches figure network, adaptivenon-uniform sampling is carried out to the target and background of the first training image, obtains first object figure
Picture;
Stylized neural network is constructed, first object image and the second training image are inputted into stylized neural network, it is right
Stylized neural network is trained;
According to interactively entering for user, the target and background using stingy figure network handles processing image is adaptively divided
It cuts, obtains the second target image;
Stylized processing is carried out to the second target image using the stylized neural network after training, obtains final stylization
Image.
Embodiment 4:
A kind of storage medium is present embodiments provided, which is computer readable storage medium, is stored with meter
Calculation machine program when described program is executed by processor, when processor executes the computer program of memory storage, realizes above-mentioned reality
The interactive image stylization processing method of example 1 is applied, as follows:
Obtain multiple first training images and multiple second training images;Wherein, first training image is to include mesh
The image of mark and background, second training image are style image;
It constructs and scratches figure network, adaptivenon-uniform sampling is carried out to the target and background of the first training image, obtains first object figure
Picture;
Stylized neural network is constructed, first object image and the second training image are inputted into stylized neural network, it is right
Stylized neural network is trained;
According to interactively entering for user, the target and background using stingy figure network handles processing image is adaptively divided
It cuts, obtains the second target image;
Stylized processing is carried out to the second target image using the stylized neural network after training, obtains final stylization
Image.
Storage medium described in the present embodiment can be disk, CD, computer storage, read-only memory (ROM,
Read-Only Memory), random access memory (RAM, Random Access Memory), USB flash disk, mobile hard disk etc. be situated between
Matter.
In conclusion the present invention scratches figure network and stylized neural network by constructing, it can be defeated according to the interaction of user
Enter, carries out adaptivenon-uniform sampling using the target and background of stingy figure network handles processing image, obtain the target that user specifies region
Image recycles stylized neural network to carry out stylized processing to the target image, to simulate different art forms
Style is drawn, enhances the form of expression of visual information in image, promotes the attraction of image, is led in multimedia, art teaching etc.
Domain has practical value.
The above, only the invention patent preferred embodiment, but the scope of protection of the patent of the present invention is not limited to
This, anyone skilled in the art is in the range disclosed in the invention patent, according to the present invention the skill of patent
Art scheme and its inventive concept are subject to equivalent substitution or change, belong to the scope of protection of the patent of the present invention.
Claims (10)
1. a kind of interactive image stylization processing method, which is characterized in that the described method includes:
Obtain multiple first training images and multiple second training images;Wherein, first training image be comprising target and
The image of background, second training image are style image;
It constructs and scratches figure network, adaptivenon-uniform sampling is carried out to the target and background of the first training image, obtains first object image;
Stylized neural network is constructed, first object image and the second training image are inputted into stylized neural network, to style
Change neural network to be trained;
According to interactively entering for user, adaptivenon-uniform sampling is carried out using the target and background of stingy figure network handles processing image, is obtained
To the second target image;
Stylized processing is carried out to the second target image using the stylized neural network after training, obtains final stylization figure
Picture.
2. interactive image stylization processing method according to claim 1, which is characterized in that described construct scratches figure net
Network carries out adaptivenon-uniform sampling to the target and background of the first training image, obtains first object image, specifically include:
In RGB color, using the full covariance mixed Gauss model of K Gaussian component respectively to the first training image
Target and background is modeled;
The target and background of the first training image is partitioned into using iteration Energy minimization;Wherein, the energy minimizes
Algorithm is initialized by K-means algorithm, respectively by belong to target and background pixel cluster be K class, the mean value of K class and
Covariance is estimated to obtain by belonging to the rgb value of the pixel of target and background;
According to the target and background of the first training image, a figure is established, weight figure is handled by maximum-flow algorithm, and with minimum
Algorithm is cut to be split;
Iteration optimization mixed Gauss model and segmentation result obtain first object image.
3. interactive image stylization processing method according to claim 2, which is characterized in that first training image
Gibbs energy it is as follows:
E (α, k, θ, z)=U (α, k, θ, z)+V (α, z)
Wherein:
α is the label of pixel, and 0 is background, and 1 is target, and 2 be possible background, and 3 be possible target;K=1...K indicates K-
The classification of means algorithm cluster;θ is the weight of Gaussian component;Z is the image data of the first training image;
U is the area item of energy function, indicate in the image data of the first training image, a pixel be classified as target or
The punishment of person's background:
According to gauss hybrid models, negative logarithm is taken to obtain the probability that some pixel belongs to target or background:
Wherein, π is the weight that single Gaussian component contributes probability, and μ is the mean vector of each Gaussian component, and ∑ is covariance
Matrix;
V is boundary energy item, indicates discontinuous punishment between neighborhood territory pixel m and n, which uses in rgb space
It is as follows in the similitude for measuring neighborhood territory pixel m and n:
Wherein, β=(2 < (zm-zn)2>)-1, β determines by the contrast of the first training image, and γ is constant, and C is adjacent color pair
Set.
4. according to the described in any item interactive image stylization processing methods of claim 2-3, which is characterized in that the iteration
After optimization mixed Gauss model and segmentation result, further includes: carried out using boundary of the border matting to segmentation smooth.
5. interactive image stylization processing method according to claim 1, which is characterized in that the building stylization mind
Through network, first object image and the second training image are inputted into stylized neural network, stylized neural network is instructed
Practice, specifically include:
It establishes and generates network;Wherein, the generation network is by encoder sub-network, AdaIN sub-network and decoder sub-network group
At encoder sub-network utilizes several layer buildings before pre-training VGG19 network;
By first object image and the second training image input coding device sub-network, the output of encoder sub-network is obtained, will be compiled
Input of the output of code device sub-network as AdaIN sub-network, obtains the output of AdaIN sub-network, by the defeated of AdaIN sub-network
Input as decoder sub-network out exports stylized image by decoder sub-network;
Using encoder sub-network as loss network is calculated, according to the stylized image of decoder sub-network output, AdaIN subnet
The output of network and the second training image calculate loss;
Using back-propagation algorithm, the optimal way finely tuned using level is successively passed by calculating transmitting gradient from the last layer
It passs, updates all parameters.
6. interactive image stylization processing method according to claim 5, which is characterized in that described by first object figure
Picture and the second training image input coding device sub-network, obtain the output of encoder sub-network, by the output of encoder sub-network
As the input of AdaIN sub-network, the output of AdaIN sub-network is obtained, using the output of AdaIN sub-network as decoder subnet
The input of network exports stylized image by decoder sub-network, specifically includes:
By first object image and the second training image input coding device sub-network, in feature space to first object image and
Two training images are encoded, the corresponding fisrt feature mapping corresponding with the second training image second of output first object image
Feature Mapping;
The fisrt feature mapping and second feature mapping input AdaIN sub-network, output target that encoder sub-network is exported are special
Sign mapping;
The target signature that AdaIN sub-network is exported maps input decoder sub-network, is transformed into figure by decoder sub-network
Image space exports stylized image.
7. interactive image stylization processing method according to claim 5, which is characterized in that the loss is damaged by content
Style of becoming estranged loss composition, as follows:
L=Lc+2Ls
Wherein, LcFor content loss, LsFor style loss, λ is weight;
The content loss be AdaIN sub-network output and decoder sub-network output stylized image between Euclidean away from
From as follows:
Lc=| | f (g (t))-t | |2
Wherein, LcFor content loss, t is the output of AdaIN sub-network, and g (t) is the stylized image of decoder sub-network output;
The style loss is as follows:
Wherein, σ is standard deviation, and μ is average value,For a wherein layer network for pre-training VGG19 network, s is the second training figure
The image data of picture.
8. a kind of interactive image stylization processing system, which is characterized in that the system comprises:
Image collection module, for obtaining multiple first training images and multiple second training images;Wherein, first training
Image is the image comprising target and background, and second training image is style image;
Figure network struction module is scratched adaptively to divide the target and background of the first training image for constructing stingy figure network
It cuts, obtains first object image;
Stylized neural network constructs module, and for constructing stylized neural network, first object image and the second training are schemed
As inputting stylized neural network, stylized neural network is trained;
Module is scratched, for interactively entering according to user, is carried out using the target and background of stingy figure network handles processing image
Adaptivenon-uniform sampling obtains the second target image;
Stylized processing module, for being carried out at stylization using the stylized neural network after training to the second target image
Reason obtains final stylized image.
9. a kind of calculating equipment, including processor and for the memory of storage processor executable program, which is characterized in that
When the processor executes the program of memory storage, the described in any item interactive image stylizations of claim 1-7 are realized
Processing method.
10. a kind of storage medium, is stored with program, which is characterized in that when described program is executed by processor, realize claim
The described in any item interactive image stylization processing methods of 1-7.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910396504.2A CN110222722A (en) | 2019-05-14 | 2019-05-14 | Interactive image stylization processing method, calculates equipment and storage medium at system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910396504.2A CN110222722A (en) | 2019-05-14 | 2019-05-14 | Interactive image stylization processing method, calculates equipment and storage medium at system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110222722A true CN110222722A (en) | 2019-09-10 |
Family
ID=67820970
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910396504.2A Pending CN110222722A (en) | 2019-05-14 | 2019-05-14 | Interactive image stylization processing method, calculates equipment and storage medium at system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110222722A (en) |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110689478A (en) * | 2019-09-25 | 2020-01-14 | 北京字节跳动网络技术有限公司 | Image stylization processing method and device, electronic equipment and readable medium |
CN111127309A (en) * | 2019-12-12 | 2020-05-08 | 杭州格像科技有限公司 | Portrait style transfer model training method, portrait style transfer method and device |
CN111340905A (en) * | 2020-02-13 | 2020-06-26 | 北京百度网讯科技有限公司 | Image stylization method, apparatus, device, and medium |
CN111951154A (en) * | 2020-08-14 | 2020-11-17 | 中国工商银行股份有限公司 | Method and device for generating picture containing background and medium |
CN112102461A (en) * | 2020-11-03 | 2020-12-18 | 北京智源人工智能研究院 | Face rendering method and device, electronic equipment and storage medium |
CN112465064A (en) * | 2020-12-14 | 2021-03-09 | 合肥工业大学 | Image identification method, device and equipment based on deep course learning |
CN112561779A (en) * | 2019-09-26 | 2021-03-26 | 北京字节跳动网络技术有限公司 | Image stylization processing method, device, equipment and storage medium |
CN112561778A (en) * | 2019-09-26 | 2021-03-26 | 北京字节跳动网络技术有限公司 | Image stylization processing method, device, equipment and storage medium |
CN112734769A (en) * | 2020-12-31 | 2021-04-30 | 山东大学 | Medical image segmentation and quantitative analysis method based on interactive information guided deep learning method, computer device and storage medium |
CN112766079A (en) * | 2020-12-31 | 2021-05-07 | 北京航空航天大学 | Unsupervised image-to-image translation method based on content style separation |
WO2021109876A1 (en) * | 2019-12-02 | 2021-06-10 | Oppo广东移动通信有限公司 | Image processing method, apparatus and device, and storage medium |
CN113160033A (en) * | 2020-12-28 | 2021-07-23 | 武汉纺织大学 | Garment style migration system and method |
CN113240599A (en) * | 2021-05-10 | 2021-08-10 | Oppo广东移动通信有限公司 | Image toning method and device, computer-readable storage medium and electronic equipment |
CN113469876A (en) * | 2021-07-28 | 2021-10-01 | 北京达佳互联信息技术有限公司 | Image style migration model training method, image processing method, device and equipment |
CN113542759A (en) * | 2020-04-15 | 2021-10-22 | 辉达公司 | Generating antagonistic neural network assisted video reconstruction |
CN113763232A (en) * | 2020-08-10 | 2021-12-07 | 北京沃东天骏信息技术有限公司 | Image processing method, device, equipment and computer readable storage medium |
CN113763233A (en) * | 2021-08-04 | 2021-12-07 | 深圳盈天下视觉科技有限公司 | Image processing method, server and photographing device |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104820990A (en) * | 2015-05-15 | 2015-08-05 | 北京理工大学 | Interactive-type image-cutting system |
CN106548208A (en) * | 2016-10-28 | 2017-03-29 | 杭州慕锐科技有限公司 | A kind of quick, intelligent stylizing method of photograph image |
CN107463622A (en) * | 2017-07-06 | 2017-12-12 | 西南交通大学 | A kind of automatic Symbolic method for keeping landmark shape facility |
CN107945204A (en) * | 2017-10-27 | 2018-04-20 | 西安电子科技大学 | A kind of Pixel-level portrait based on generation confrontation network scratches drawing method |
CN108629747A (en) * | 2018-04-25 | 2018-10-09 | 腾讯科技(深圳)有限公司 | Image enchancing method, device, electronic equipment and storage medium |
CN108961349A (en) * | 2018-06-29 | 2018-12-07 | 广东工业大学 | A kind of generation method, device, equipment and the storage medium of stylization image |
CN109325903A (en) * | 2017-07-31 | 2019-02-12 | 北京大学 | The method and device that image stylization is rebuild |
CN109697690A (en) * | 2018-11-01 | 2019-04-30 | 北京达佳互联信息技术有限公司 | Image Style Transfer method and system |
CN109712068A (en) * | 2018-12-21 | 2019-05-03 | 云南大学 | Image Style Transfer and analogy method for cucurbit pyrography |
-
2019
- 2019-05-14 CN CN201910396504.2A patent/CN110222722A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104820990A (en) * | 2015-05-15 | 2015-08-05 | 北京理工大学 | Interactive-type image-cutting system |
CN106548208A (en) * | 2016-10-28 | 2017-03-29 | 杭州慕锐科技有限公司 | A kind of quick, intelligent stylizing method of photograph image |
CN107463622A (en) * | 2017-07-06 | 2017-12-12 | 西南交通大学 | A kind of automatic Symbolic method for keeping landmark shape facility |
CN109325903A (en) * | 2017-07-31 | 2019-02-12 | 北京大学 | The method and device that image stylization is rebuild |
CN107945204A (en) * | 2017-10-27 | 2018-04-20 | 西安电子科技大学 | A kind of Pixel-level portrait based on generation confrontation network scratches drawing method |
CN108629747A (en) * | 2018-04-25 | 2018-10-09 | 腾讯科技(深圳)有限公司 | Image enchancing method, device, electronic equipment and storage medium |
CN108961349A (en) * | 2018-06-29 | 2018-12-07 | 广东工业大学 | A kind of generation method, device, equipment and the storage medium of stylization image |
CN109697690A (en) * | 2018-11-01 | 2019-04-30 | 北京达佳互联信息技术有限公司 | Image Style Transfer method and system |
CN109712068A (en) * | 2018-12-21 | 2019-05-03 | 云南大学 | Image Style Transfer and analogy method for cucurbit pyrography |
Non-Patent Citations (7)
Title |
---|
LINGYU LIANG等: "Facial Skin Beautification Using Adaptive Region-Aware Masks", 《IEEE TRANSACTIONS ON CYBERNETICS》 * |
XUN HUANG等: "Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV)》 * |
丁红 等: "基于快速收敛Grabcut的目标提取算法", 《计算机工程与设计》 * |
吴凯琳: "计算机绘制低多边形风格肖像", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
栾奕欣: "基于深度学习的图像风格化处理", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
梁凌宇: "人脸图像的自适应美化与渲染研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
陈梓炜: "基于深度特征插值的图像变换方法研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (27)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110689478B (en) * | 2019-09-25 | 2023-12-01 | 北京字节跳动网络技术有限公司 | Image stylization processing method and device, electronic equipment and readable medium |
CN110689478A (en) * | 2019-09-25 | 2020-01-14 | 北京字节跳动网络技术有限公司 | Image stylization processing method and device, electronic equipment and readable medium |
CN112561779B (en) * | 2019-09-26 | 2023-09-29 | 北京字节跳动网络技术有限公司 | Image stylization processing method, device, equipment and storage medium |
CN112561779A (en) * | 2019-09-26 | 2021-03-26 | 北京字节跳动网络技术有限公司 | Image stylization processing method, device, equipment and storage medium |
CN112561778A (en) * | 2019-09-26 | 2021-03-26 | 北京字节跳动网络技术有限公司 | Image stylization processing method, device, equipment and storage medium |
WO2021109876A1 (en) * | 2019-12-02 | 2021-06-10 | Oppo广东移动通信有限公司 | Image processing method, apparatus and device, and storage medium |
CN111127309A (en) * | 2019-12-12 | 2020-05-08 | 杭州格像科技有限公司 | Portrait style transfer model training method, portrait style transfer method and device |
CN111127309B (en) * | 2019-12-12 | 2023-08-11 | 杭州格像科技有限公司 | Portrait style migration model training method, portrait style migration method and device |
CN111340905A (en) * | 2020-02-13 | 2020-06-26 | 北京百度网讯科技有限公司 | Image stylization method, apparatus, device, and medium |
CN111340905B (en) * | 2020-02-13 | 2023-08-04 | 北京百度网讯科技有限公司 | Image stylization method, device, equipment and medium |
CN113542759B (en) * | 2020-04-15 | 2024-05-10 | 辉达公司 | Generating an antagonistic neural network assisted video reconstruction |
CN113542759A (en) * | 2020-04-15 | 2021-10-22 | 辉达公司 | Generating antagonistic neural network assisted video reconstruction |
CN113763232A (en) * | 2020-08-10 | 2021-12-07 | 北京沃东天骏信息技术有限公司 | Image processing method, device, equipment and computer readable storage medium |
CN111951154B (en) * | 2020-08-14 | 2023-11-21 | 中国工商银行股份有限公司 | Picture generation method and device containing background and medium |
CN111951154A (en) * | 2020-08-14 | 2020-11-17 | 中国工商银行股份有限公司 | Method and device for generating picture containing background and medium |
CN112102461A (en) * | 2020-11-03 | 2020-12-18 | 北京智源人工智能研究院 | Face rendering method and device, electronic equipment and storage medium |
CN112102461B (en) * | 2020-11-03 | 2021-04-09 | 北京智源人工智能研究院 | Face rendering method and device, electronic equipment and storage medium |
CN112465064A (en) * | 2020-12-14 | 2021-03-09 | 合肥工业大学 | Image identification method, device and equipment based on deep course learning |
CN113160033A (en) * | 2020-12-28 | 2021-07-23 | 武汉纺织大学 | Garment style migration system and method |
CN112766079B (en) * | 2020-12-31 | 2023-05-26 | 北京航空航天大学 | Unsupervised image-to-image translation method based on content style separation |
CN112734769B (en) * | 2020-12-31 | 2022-11-04 | 山东大学 | Medical image segmentation and quantitative analysis method based on interactive information guided deep learning method, computer device and storage medium |
CN112766079A (en) * | 2020-12-31 | 2021-05-07 | 北京航空航天大学 | Unsupervised image-to-image translation method based on content style separation |
CN112734769A (en) * | 2020-12-31 | 2021-04-30 | 山东大学 | Medical image segmentation and quantitative analysis method based on interactive information guided deep learning method, computer device and storage medium |
CN113240599A (en) * | 2021-05-10 | 2021-08-10 | Oppo广东移动通信有限公司 | Image toning method and device, computer-readable storage medium and electronic equipment |
CN113469876A (en) * | 2021-07-28 | 2021-10-01 | 北京达佳互联信息技术有限公司 | Image style migration model training method, image processing method, device and equipment |
CN113469876B (en) * | 2021-07-28 | 2024-01-09 | 北京达佳互联信息技术有限公司 | Image style migration model training method, image processing method, device and equipment |
CN113763233A (en) * | 2021-08-04 | 2021-12-07 | 深圳盈天下视觉科技有限公司 | Image processing method, server and photographing device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110222722A (en) | Interactive image stylization processing method, calculates equipment and storage medium at system | |
Li et al. | A closed-form solution to photorealistic image stylization | |
CN110378985B (en) | Animation drawing auxiliary creation method based on GAN | |
CN109816009A (en) | Multi-tag image classification method, device and equipment based on picture scroll product | |
CN107845072B (en) | Image generating method, device, storage medium and terminal device | |
CN107993238A (en) | A kind of head-and-shoulder area image partition method and device based on attention model | |
CN109902798A (en) | The training method and device of deep neural network | |
CN110443239A (en) | The recognition methods of character image and its device | |
CN102184221A (en) | Real-time video abstract generation method based on user preferences | |
CN106778852A (en) | A kind of picture material recognition methods for correcting erroneous judgement | |
CN113255813B (en) | Multi-style image generation method based on feature fusion | |
CN110097616B (en) | Combined drawing method and device, terminal equipment and readable storage medium | |
CN111178312B (en) | Face expression recognition method based on multi-task feature learning network | |
CN114663685B (en) | Pedestrian re-recognition model training method, device and equipment | |
CN113408537B (en) | Remote sensing image domain adaptive semantic segmentation method | |
CN108875693A (en) | A kind of image processing method, device, electronic equipment and its storage medium | |
CN110782448A (en) | Rendered image evaluation method and device | |
CN110516734A (en) | A kind of image matching method, device, equipment and storage medium | |
CN109978074A (en) | Image aesthetic feeling and emotion joint classification method and system based on depth multi-task learning | |
CN114758180A (en) | Knowledge distillation-based light flower recognition method | |
CN113838158B (en) | Image and video reconstruction method and device, terminal equipment and storage medium | |
CN113554653A (en) | Semantic segmentation method for long-tail distribution of point cloud data based on mutual information calibration | |
CN115018729A (en) | White box image enhancement method for content | |
CN115082800A (en) | Image segmentation method | |
US11734389B2 (en) | Method for generating human-computer interactive abstract image |
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: 20190910 |
|
RJ01 | Rejection of invention patent application after publication |