CN112837212B - Image arbitrary style migration method based on manifold alignment - Google Patents

Image arbitrary style migration method based on manifold alignment Download PDF

Info

Publication number
CN112837212B
CN112837212B CN202110118940.0A CN202110118940A CN112837212B CN 112837212 B CN112837212 B CN 112837212B CN 202110118940 A CN202110118940 A CN 202110118940A CN 112837212 B CN112837212 B CN 112837212B
Authority
CN
China
Prior art keywords
style
content
manifold
style migration
migration
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
CN202110118940.0A
Other languages
Chinese (zh)
Other versions
CN112837212A (en
Inventor
霍静
金世印
李文斌
高阳
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.)
Nanjing University
Original Assignee
Nanjing 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 Nanjing University filed Critical Nanjing University
Priority to CN202110118940.0A priority Critical patent/CN112837212B/en
Publication of CN112837212A publication Critical patent/CN112837212A/en
Application granted granted Critical
Publication of CN112837212B publication Critical patent/CN112837212B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G06T3/04—
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • 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
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components

Abstract

The invention relates to an arbitrary style migration method of an image based on manifold alignment, which constructs a style migration network and comprises an encoder, a decoder and a manifold alignment module; the method comprises the steps that a content graph and a style graph are input into an encoder to obtain content characteristics and style characteristics, a manifold alignment module is used for projecting a content manifold to a style manifold space, the converted characteristics are input into a decoder to be decoded to obtain a corresponding stylized image, and the stylized image retains the content of the content graph and the style of the style image; and applying orthogonal constraint on the manifold projection matrix, and calculating the similarity between corresponding framed areas to realize the autonomous selection of the neighbor of the appointed area. The invention can realize style migration of any style; and allows user editing to achieve style migration between user-specified semantic regions, realistic style migration can be achieved by applying orthogonal constraints to the projection matrix.

Description

Image arbitrary style migration method based on manifold alignment
Technical Field
The invention belongs to the field of computer application, and particularly relates to an arbitrary style migration method for images based on manifold alignment.
Background
The style migration is to migrate the style of a certain artistic work to a designated content graph through a certain method, so that the generated image not only keeps the content of the content graph but also has the style of the artistic work. Style migration is an artistic style of creating new artistic images by converting the artistic style of the images, and is popular among the public because the generated images are strong in artistic sense.
Most current methods are based on the artistic style texture of an image and can be described by using global statistics of deep feature images generated in a convolutional neural network, such as using a Gram matrix and a covariance matrix as descriptions of the texture of the image style. Such global statistics are capturing style textures from the entire image and incorporating them into the content image without detail regarding whether specific content information is semantically matched. But for most images, it actually contains many different semantic parts, then the style migration based on global statistics is not sufficient to migrate style textures of many semantics based on semantic matching of content and style images. Therefore, the existing method can make the overall style texture of the migration result look like a style map, but the migration result can not well maintain the local semantic structure of the content image.
It is apparent that the assumption that global statistics of deep features in a neural network are used to represent artistic style texture information of an image may not be suitable in many cases because an image may contain multiple semantic objects, and a single global statistic cannot fully capture feature distribution of different semantics. Some researchers have proposed style migration methods based on local semantic alignment that find semantically similar or corresponding regions between content images and style images, which achieve better results in terms of preservation of content structure and details of style texture migration for similar semantic regions than simple global statistics-based methods. However, these methods either involve multiple stages or have many penalty terms to trade off, making the overall algorithm very difficult to adjust to optimum, and thus failing to achieve both effect and efficiency, resulting in a reduction in the overall effect of style migration. Generally, style migration has the following difficulties: 1) The stylized mode has rich diversity, and the style migration corresponding to the semantics is a proper stylized mode, and how to realize the style migration at the semantic level is a hot spot problem of current research. 2) The current style migration algorithm cannot achieve user control, namely, a user cannot control a corresponding region according to own requirements to obtain a specified style. 3) At present, most of style migration algorithms can only process one scene in the style migration of realism or non-realism, and how to realize an algorithm to process two scenes in the style migration of realism and non-realism at the same time is a worth discussing problem.
Disclosure of Invention
The invention aims to: the invention provides an arbitrary style migration method of an image based on manifold alignment aiming at the task of style migration, which converts a style migration problem into a manifold alignment problem through manifold assumption, solves through closed solution, realizes style migration of arbitrary images and simultaneously supports user editing and realistic style migration.
The technical scheme is as follows: the invention discloses a manifold alignment-based image arbitrary style migration method, which specifically comprises the following steps:
(1) Constructing a style migration network and training; the style migration network comprises an encoder, a decoder and a manifold alignment module;
(2) Non-realistic style migration popularity alignment: extracting content characteristics and style characteristics from the content graph and the style graph respectively by using a trained encoder, regarding the content characteristics and the style characteristics as two different manifold distributions, calculating a similarity matrix between the content manifold and the style manifold, extracting a K neighbor matrix through the similarity matrix, solving a projection matrix from the content manifold to the style manifold, and projecting the content characteristics into a style space by using the projection matrix;
(3) Realistic style migration manifold alignment: if the user wants to perform the style migration of the sense of reality, adding an orthogonal constraint to the projection matrix when calculating the projection matrix in the step (2), ensuring that the content features are not destroyed after projection, and keeping the structure of the content graph to realize the style migration of the sense of reality;
(4) User editing: if the user wants to perform style migration between the designated areas, corresponding areas need to be framed on the content graph and the style graph, and when the similarity graph between the content features and the style features is calculated in the step (2), similarity calculation between the corresponding framed areas is performed, so that autonomous selection of neighbors of the designated areas is realized;
(5) And inputting the converted characteristics obtained through the non-photorealistic style migration manifold alignment or the photorealistic style migration manifold alignment module into a trained decoder to reconstruct the stylized picture.
Further, the implementation process of the step (2) is as follows:
the non-realistic style migration optimization objective is as follows:
where P is the projection matrix to be solved, A cs Is F c And F is equal to s Neighbor relation matrix of F c 、F s Content features and style features, respectively, W c 、H c 、W s 、H s Respectively F c 、F s N is the number of neighbor pairs formed whenWhen (I)>And (3) withIs optimized to be closer and closer;
the closed-form solution of the objective function is found as follows:
wherein ,
further, an adaptive weight cross-layer link module is added in the sense style migration popularity alignment in the step (3), so that the sense of reality of the image is enhanced.
Further, the implementation process of the step (3) is as follows:
the realistic style migration optimization objective is as follows:
s.t.P T P＝I
the solution is as follows:
P＝UV T
where P is the projection matrix to be solved, A cs Is F c And F is equal to s Neighbor relation matrix of F c 、F s Content features and style features, respectively, W c 、H c 、W s 、H s Respectively F c 、F s N is the number of constituent neighbor pairs.
Further, the implementation process of the step (4) is as follows:
order theRepresenting some corresponding semantic regions specified by the user on the content and style graphs, where m represents the number of corresponding semantic regions specified by the user, neighbor matrix to be obtained by the corresponding semantic regions +.>And the original neighbor matrix->Performing AND operation to obtain final neighbor matrix, wherein +.>The definition is as follows:
the beneficial effects are that: compared with the prior art, the invention has the beneficial effects that: 1. the invention not only realizes the style migration of any style; semantic-level style migration can also be realized; 2. the invention allows the user to edit so as to realize style migration among the user-specified semantic regions, and has interactivity which is not available in other algorithms; 3. the invention can realize the migration of the sense of realism style by applying orthogonal constraint to the projection matrix, and the prior art can only deal with single situations of sense of realism or non-sense of realism.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of a non-realistic style migration network structure according to the present invention, wherein (a) is a single-layer style migration structure and (b) is a multi-layer style migration structure;
FIG. 3 is a schematic diagram of a realistic style migration network structure according to the present invention, wherein (a) is a network structure schematic diagram, and (b) is an operation diagram when linking is performed;
FIG. 4 is a diagram of an exemplary non-realistic style migration result of the present invention;
FIG. 5 is an exemplary diagram of a realism style migration result of the present invention;
fig. 6 is a diagram showing an example of the result of user editing according to the present invention.
Detailed Description
The invention is described in further detail below with reference to the accompanying drawings.
The invention provides an arbitrary style migration method of an image based on manifold alignment, a constructed style migration network is composed of an encoder, a decoder and a manifold alignment module, a content image and a style image are input into the encoder to obtain content characteristics and style characteristics, the content characteristics and the style characteristics are distributed from different manifolds, the manifold alignment module can project the content manifold into a style manifold space, the converted characteristics are input into the decoder to be decoded to obtain a corresponding stylized image, and the style of the stylized image is reserved while the content of the content image is reserved. In order to achieve stronger interactivity, the method also allows the user to select the appointed areas in the content map and the style map frame respectively so as to achieve style migration between the appointed areas. In order to realize the migration of the sense of reality style, the method can apply orthogonal constraint on the manifold projection matrix, and a brand new self-adaptive weight cross-layer connection module is provided to further enhance the stylization effect. As shown in fig. 1, the method specifically comprises the following steps:
step 1: constructing a style migration network and training; the constructed style migration network includes an encoder, a decoder, and a manifold alignment module.
The method comprises the steps of obtaining a content drawing data set and a style drawing data set in advance, and dividing a training set and a testing set respectively; respectively inputting the content graph and the style graph into an encoder to respectively extract content characteristics and style characteristics; content map/style map reconstruction: inputting the characteristics of the content map/style map into a decoder to reconstruct the content map and the style map respectively; and reconstructing the self-encoder based on the divided content image and style image data set training image, and storing model parameters after the self-encoder converges.
Step 2: non-realistic style migration popularity alignment: and respectively extracting content characteristics and style characteristics from the content graph and the style graph by using a trained encoder, regarding the content characteristics and the style characteristics as two different manifold distributions, calculating a similarity matrix between the content manifold and the style manifold, extracting a K neighbor matrix through the similarity matrix, solving a projection matrix from the content manifold to the style manifold, and projecting the content characteristics into a style space by using the projection matrix.
The non-photorealistic style migration network is shown in fig. 2, and specifically includes: an encoder, a non-realistic manifold alignment module (Manifold Alignment, MA), a decoder. The invention adopts 5 different layers of VGG19 network which are commonly used in the style migration field as 5 encoders with different scales, namely Relu5_1, relu4_1, relu3_1, relu2_1 and Relu1_1, and respectively constructs decoders with structures symmetrical to 5 encoders, namely, decoder5, decoder4, decoder3, decoder2 and Decoder1.
In a single-layer style migration structure, as shown in the figure2 (a), giving a content map I c And style sheet I s Extracting corresponding features by an Encoder and />Wherein C is the number of channels of the feature, W c 、W s Is the width of the feature, H c 、H s Is a feature high. We denote the content and style with subscripts c and s, respectively. Thus F can be set c (F s ) Regarded as W c ×H c (W s ×H s ) A set of C-dimensional vectors. F without any treatment c And F is equal to s Typically of different distribution, the manifold alignment module uses a mapping relationship to characterize the content F c Mapping to style characteristics F s Is defined in the space of the mold.
Specifically, learn a mapping matrixProjection of content features into a style feature space is achieved with P, namely:
F cs ＝P T F c
firstly, the following steps:
wherein ,represent phi j (F s ) K-nearest neighbor of->Represent phi i (F c ) In (2) using cosine distance to perform distance measurement to find k-nearest neighbor, thus when phi i (F c ) Is phi j (F s ) K-nearest neighbor or phi of (1) j (F s ) Is phi i (F c ) K neighbor time->Otherwise is 0, actually A cs Define F c And F is equal to s Is a neighbor relation matrix of (a).
The semantic similarity between the content features and the style features is measured by using the neighbor relation matrix, the optimization aim is to make the content features and the style features with the semantic similarity more approximate after projection, and the objective function is as follows:
where P is the projection matrix to be solved, A cs Is F c And F is equal to s Neighbor relation matrix of F c 、F s Content features and style features, respectively, W c 、H c 、W s 、H s Respectively F c 、F s N is the number of neighbor pairs formed whenWhen (I)>And (3) withIs optimized to be closer and closer.
A closed-form solution of the objective function is obtained as follows:
wherein ,
and obtaining the converted characteristics, and inputting the converted characteristics into a trained decoder to reconstruct the stylized picture.
The above is a flow and a specific principle of a single-layer structure of a non-realistic style migration network, and in order to further enhance the stylized effect, a multi-layer style migration structure is designed, as shown in fig. 2 (b):
the multi-layer structure is based on a single-layer structure, and the stylized structure image output by the single-layer structure is taken as a content image of the next layer and is stylized again with the original style image. Such as: the encoder Relu5_1 forms a first layer structure with the non-realistic manifold alignment module MA and the Decoder5, receives the original content image I c And original style graph I s Obtaining a stylized result I of the first layer 5 After that I 5 Input to the next layer (encoder Relu4_1, MA, decoder 4) together with the original style graph to obtain stylized result I 4 This is repeated until the final stylized result of the last layer output is obtained.
Step 3: realistic style migration manifold alignment: if the user needs to perform the style migration of realism, an orthogonal constraint needs to be added to the projection matrix when the projection matrix is calculated, so that the content features are not seriously damaged after projection, and the structure of the content graph is well maintained to realize the style migration of realism.
The structure of the realistic style migration network is shown in fig. 3 (a), and specifically includes an encoder, an adaptive weight cross-layer linking module (Adaptative Weight Skip Connection, AWSC), a realistic manifold alignment module (Manifold Alignment, MA), and a decoder. Unlike non-realistic style migration, realistic style migration is to keep the structure of a content map as undamaged as possible when migrating the style of a style map onto the content map. The realistic style migration network is similar to the non-realistic style migration network, but has several points that the difference is that, firstly, the manifold alignment module applies orthogonal constraint to the projection matrix in the realistic style migration manifold alignment module, and after the orthogonal constraint is added, the self-similarity of the content characteristics after projection is not changed, so that the content structure of the image after projection is not changed greatly; secondly, in order to further enhance the sense of reality, an adaptive weight cross-layer link module (AWSC) is added on the basis of a traditional self-encoder (AutoEncoder), and a sense of reality style migration manifold alignment module and an AWSC module are specifically described below.
(1) And the realistic style migration manifold alignment module is used for:
the realism style migration objective function after adding the orthogonal constraint is as follows:
s.t.P T P＝I
where P is the projection matrix to be solved, A cs Is F c And F is equal to s Is a matrix of neighbor relation of (c),is a local feature at the ith position in the projected features, F c 、F s Content features and style features, respectively, W c 、H c 、W s 、H s Respectively F c 、F s N is the number of neighbor pairs that are formed, and the objective function is solved as follows:
(2) Adaptive weight cross-layer link module (AWSC)
Firstly, the converted features of the content graph and the style graph Relu5_1 layer are obtained through a realistic style migration manifold alignment module, the converted features are sent to the corresponding previous layers of the Decoder to be reconstructed, the obtained features and the features obtained through conversion in Relu4_1 are subjected to self-adaptive weight linking, and the steps are repeated until the features of the Relu1_1 layer are linked, so that high-level and low-level semantic information is combined.
The operation of the link is specifically shown in fig. 3 (b), and features after the conversion of the present layerF connect The channel mean variance is aligned to the feature F obtained after high-level conversion and through partial hierarchical reconstruction of the decoder out Thereafter, the mean and variance F will be aligned connect And F is equal to out Adding element levels to obtain an output characteristic F' out 。
Step 4: user editing: if the user wants to perform style migration between the designated areas, the corresponding areas need to be framed on the content graph and the style graph, and when the similarity graph between the content features and the style features is calculated in the step (2), similarity calculation between the corresponding framed areas is performed, so that autonomous selection of neighbors of the designated areas is realized.
The method of the invention is easily expanded into a mode which can be edited by a user, and the user can respectively draw a plurality of strokes on the content graph and the style graph to represent the semantic consistency, and only needs to change the neighbor matrix A cs Such user-specified constraints may be added. Specifically, let theRepresenting some corresponding semantic regions specified by the user on the content and style graphs, where m represents the number of corresponding semantic regions specified by the user, neighbor matrices to be derived by these corresponding semantic regions +.>And the original neighbor matrix->Performing an AND operation to obtain a final neighbor matrix, wherein +.>The definition is as follows:
fig. 4 shows a non-realistic style migration effect, in which line 1 is a content graph, line 2 is a style graph, and line 3 is a migration result, and it can be seen that the resulting graph retains the content structure of the content graph, and also well migrates the style of the style graph in the past. The resulting picture is very similar to a style sheet, both in terms of overall color and texture detail.
Fig. 5 shows a realistic style migration effect, in which, line 1 is a content graph, line 2 is a style graph, and line 3 is a migration result, it can be seen that the obtained migration result well retains the content structure of the content graph and has no obvious deformation, and has a very strong sense of reality.
Fig. 6 shows the user editing effect, columns 1 and 2 are respectively a content graph and a style graph, columns 3 and 4 are corresponding areas drawn on the content graph and the style graph, and columns 5 and 6 are migration results before and after user editing, so that after user editing, style migration between user-specified areas can be realized, for example, the color of sky becomes a specified color.

Claims (4)

1. An arbitrary style migration method of an image based on manifold alignment is characterized by comprising the following steps:
(1) Constructing a style migration network and training; the style migration network comprises an encoder, a decoder and a manifold alignment module;
(2) Non-realistic style migration popularity alignment: extracting content characteristics and style characteristics from the content graph and the style graph respectively by using a trained encoder, regarding the content characteristics and the style characteristics as two different manifold distributions, calculating a similarity matrix between the content manifold and the style manifold, extracting a K neighbor matrix through the similarity matrix, solving a projection matrix from the content manifold to the style manifold, and projecting the content characteristics into a style space by using the projection matrix;
(3) Realistic style migration manifold alignment: if the user wants to perform the style migration of the sense of reality, adding an orthogonal constraint to the projection matrix when calculating the projection matrix in the step (2), ensuring that the content features are not destroyed after projection, and keeping the structure of the content graph to realize the style migration of the sense of reality;
(4) User editing: if the user wants to perform style migration between the designated areas, the corresponding areas need to be framed on the content graph and the style graph, and when the similarity graph between the content characteristics and the style characteristics is calculated in the step (2), similarity calculation between the corresponding framed areas is performed, so that autonomous selection of the neighbors of the designated areas is realized;
(5) Inputting the converted characteristics obtained by the non-realistic style migration manifold alignment or the realistic style migration manifold alignment module into a trained decoder to reconstruct the stylized picture;
the implementation process of the step (3) is as follows:
the realistic style migration optimization objective is as follows:
s.t.P T P＝I
the solution is as follows:
P＝UV T
where P is the projection matrix to be solved, A cs Is F c And F is equal to s Is a matrix of neighbor relation of (c),is a local feature at the ith position in the projected features, F c 、F s Content features and style features, respectively, W c 、H c 、W s 、H s Respectively F c 、F s N is the number of constituent neighbor pairs, +.>And the normalized neighbor relation matrix.
2. The manifold alignment-based image arbitrary style migration method of claim 1, wherein the step (2) is implemented as follows:
the non-realistic style migration optimization objective is as follows:
where P is the projection matrix to be solved, A cs Is F c And F is equal to s Neighbor relation matrix of F c 、F s Content features and style features, respectively, W c 、H c 、W s 、H s Respectively F c 、F s N is the number of neighbor pairs formed whenWhen (I)>And->Is optimized to be closer and closer;
the closed-form solution of the objective function is found as follows:
wherein ,
3. the manifold alignment-based image arbitrary style migration method of claim 1, wherein an adaptive weight cross-layer link module is further added in the sense of realism style migration popular alignment in the step (3), so as to enhance the sense of realism of the image.
4. The manifold alignment-based image arbitrary style migration method of claim 1, wherein the step (4) is implemented as follows:
order theRepresenting some corresponding semantic regions specified by the user on the content and style graphs, where m represents the number of corresponding semantic regions specified by the user, neighbor matrix to be obtained by the corresponding semantic regions +.>And the original neighbor matrix->Performing AND operation to obtain final neighbor matrix, wherein +.>The definition is as follows:
CN202110118940.0A 2021-01-28 2021-01-28 Image arbitrary style migration method based on manifold alignment Active CN112837212B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110118940.0A CN112837212B (en) 2021-01-28 2021-01-28 Image arbitrary style migration method based on manifold alignment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110118940.0A CN112837212B (en) 2021-01-28 2021-01-28 Image arbitrary style migration method based on manifold alignment

Publications (2)

Publication Number Publication Date
CN112837212A CN112837212A (en) 2021-05-25
CN112837212B true CN112837212B (en) 2023-09-26

Family

ID=75932219

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110118940.0A Active CN112837212B (en) 2021-01-28 2021-01-28 Image arbitrary style migration method based on manifold alignment

Country Status (1)

Country Link
CN (1) CN112837212B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113989102B (en) * 2021-10-19 2023-01-06 复旦大学 Rapid style migration method with high shape-preserving property
CN113888400B (en) * 2021-11-04 2024-04-26 大连理工大学 Image style migration method and device

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109409287A (en) * 2018-10-25 2019-03-01 山东大学 A kind of transfer learning method by macro sheet feelings to micro- expression
CN111325681A (en) * 2020-01-20 2020-06-23 南京邮电大学 Image style migration method combining meta-learning mechanism and feature fusion
CN111476805A (en) * 2020-05-22 2020-07-31 南京大学 Cross-source unsupervised domain adaptive segmentation model based on multiple constraints

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10748324B2 (en) * 2018-11-08 2020-08-18 Adobe Inc. Generating stylized-stroke images from source images utilizing style-transfer-neural networks with non-photorealistic-rendering

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109409287A (en) * 2018-10-25 2019-03-01 山东大学 A kind of transfer learning method by macro sheet feelings to micro- expression
CN111325681A (en) * 2020-01-20 2020-06-23 南京邮电大学 Image style migration method combining meta-learning mechanism and feature fusion
CN111476805A (en) * 2020-05-22 2020-07-31 南京大学 Cross-source unsupervised domain adaptive segmentation model based on multiple constraints

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Manifold Alignment for Semantically Aligned Style Transfer;jing huo 等;《arxiv：2005.10777》;4321-4330 *
相关对齐的总变分风格迁移新模型;谢斌;汪宁;范有伟;;中国图象图形学报(第02期);31-44 *

Also Published As

Publication number Publication date
CN112837212A (en) 2021-05-25

Similar Documents

Publication Publication Date Title
Zhang et al. Semantic image inpainting with progressive generative networks
CN112837212B (en) Image arbitrary style migration method based on manifold alignment
CN111696028A (en) Method and device for processing cartoon of real scene image, computer equipment and storage medium
CN111242844B (en) Image processing method, device, server and storage medium
CN113901894A (en) Video generation method, device, server and storage medium
CN111862294A (en) ArcGAN network-based automatic coloring network structure and method for hand-drawn 3D building
CN114820341A (en) Image blind denoising method and system based on enhanced transform
CN114493991A (en) Attention circulation countermeasure network-based style migration system, method and device
CN113255813B (en) Multi-style image generation method based on feature fusion
CN111986075A (en) Style migration method for target edge clarification
WO2023151529A1 (en) Facial image processing method and related device
CN112381716A (en) Image enhancement method based on generation type countermeasure network
CN115829876A (en) Real degraded image blind restoration method based on cross attention mechanism
CN116958534A (en) Image processing method, training method of image processing model and related device
CN113298704B (en) Skin color segmentation and beautification method by utilizing graph migration under broadcast television news
CN112884636B (en) Style migration method for automatically generating stylized video
CN114841859A (en) Single-image super-resolution reconstruction method based on lightweight neural network and Transformer
Mun et al. Texture preserving photo style transfer network
CN117522697A (en) Face image generation method, face image generation system and model training method
CN117095071A (en) Picture or video generation method, system and storage medium based on main body model
CN117315241A (en) Scene image semantic segmentation method based on transformer structure
CN116167920A (en) Image compression and reconstruction method based on super-resolution and priori knowledge
CN116245968A (en) Method for generating HDR image based on LDR image of transducer
CN113344771B (en) Multifunctional image style migration method based on deep learning
CN114331894A (en) Face image restoration method based on potential feature reconstruction and mask perception

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