CN110298898A - Change the method and its algorithm structure of automobile image body color - Google Patents
Change the method and its algorithm structure of automobile image body color Download PDFInfo
- Publication number
- CN110298898A CN110298898A CN201910465289.7A CN201910465289A CN110298898A CN 110298898 A CN110298898 A CN 110298898A CN 201910465289 A CN201910465289 A CN 201910465289A CN 110298898 A CN110298898 A CN 110298898A
- Authority
- CN
- China
- Prior art keywords
- image
- regional area
- pixel
- production
- area
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/001—Texturing; Colouring; Generation of texture or colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of method for changing image local area, belongs to image processing technology.The described method includes: obtaining the first image and pixel vectors, the regional area in the first image is determined;The regional area for generating and there is pixel characteristic corresponding with the pixel vectors is mapped in the regional area using the pixel vectors.The present invention can generate the auto graph of true body color, increase the diversity of advertisement material, improve the accuracy that model changes colour, and the present invention is also equipped with the ability of vehicle body segmentation, and makes model in colourshifting process, increase the uniformity changed colour.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of method, one kind for changing image local area
For changing algorithm structure, a kind of training method for changing image local area algorithm, Yi Zhongyong of image local area
Equipment and a kind of computer readable storage medium in change image local area.
Background technique
With the high speed development of internet, Internet advertising becomes one of current most important advertisement form, due to it
Quickly, efficiently, the ideal feature of effect, favor of the Internet advertising by each medium-sized and small enterprises client.In order to improve advertisement
The pattern of clicking rate, Internet advertising also emerges one after another, and specifically includes various form such as text, picture, video, ad content
Rich, advertisement pattern diversity has great importance for the clicking rate for improving advertisement.Since the energy of advertiser has
Limit, the material that uses is limited when launching advertisement, so needing to help advertiser to improve the more of advertisement material using the means of technology
Sample.
It is more more and more intense for increasing advertisement material, reduction advertiser's workload, the diversity requirement of raising advertisement dispensing.
There are no the appearance of relevant algorithm currently on the market.It is a kind of deep learning model that production, which fights network, is complicated in recent years
One of the method for unsupervised learning most prospect in distribution.Model passes through (at least) two modules in frame: production model
The mutual Game Learning of (Generative Model) and discrimination model (Discriminative Model) generates fairly good
Output.The automobile that traditional production confrontation network (GAN, Generative Adversarial Networks) model generates
Change colour picture, and availability is low, uneven color and effect are undesirable.The advertisement figure of the present invention this class of hanging down for automobile industry
Piece proposes method and algorithm structure for changing body of a motor car color.
Summary of the invention
The purpose of the embodiment of the present invention is that providing the method and its algorithm structure of change automobile image body color.
To achieve the goals above, the embodiment of the present invention provides a kind of method for changing image local area, this method packet
It includes:
S1 the first image and pixel vectors) are obtained, determine the regional area in the first image;
S2 generation) is mapped in the regional area using the pixel vectors has picture corresponding with the pixel vectors
The regional area of plain feature.
Specifically, step S1) in determine the first image in regional area, comprising:
S101 the object in the first image) is determined;
S102 the edge detection and figure about the object and the background of the first image) are carried out to the first image
As segmentation, the image with boundary characteristic is obtained;
S103) according to the relative position in the region where the object in the first image, determining has side described
Regional area corresponding with the relative position in the image of boundary's feature.
Specifically, step S1) in determine the first image in regional area, comprising:
The production that the first image is input to the decoder with edge detection and function of image segmentation is fought into net
Network fights network by the production and determines the regional area in the first image.
Specifically, step S2) include:
S201) utilize the pixel vectors in conjunction with the neuron of the decoder of production confrontation network in the part
The regional area with pixel characteristic corresponding with the pixel vectors is generated in region, and is formed with the first image
Second image of background and the regional area with the pixel characteristic;
S202) classified by the classifier that the production fights network to the regional area of second image,
It obtains classification results and obtains pixel vectors corresponding with the regional area of second image according to the classification results;
S203) calculate corresponding with the regional area of second image pixel vectors and step S1) pixel vectors
Pixel vectors cross entropy is obtained using the first image and second image in conjunction with the arbiter of production confrontation network
Loss must be fought, and the pixel vectors cross entropy is updated to the current pixel that the production fights the classifier of network
Vector cross entropy loses the current confrontation that the confrontation loss is updated to the arbiter of the production confrontation network.
Specifically, further include:
S3) acquisition participates in step S2) output data of the neuron of the decoders of map making processes, and according to institute
State the masking-out area that output data combination prediction model obtains the regional area with pixel characteristic, calculate the masking-out area and
The regional area cross entropy in the mark masking-out area of the regional area in the first image, and more by the regional area cross entropy
Regional area cross entropy newly current for the prediction model.
Specifically, can carry out parallel and can also connect the above method to continue with the above method, further includes:
S3 the second image of background and the regional area with pixel characteristic with the first image) is formed,
And obtain original pixel vector corresponding with the regional area of the first image;
S4 generation) is re-mapped in the regional area with pixel characteristic using the original pixel vector has and institute
State the regional area of the corresponding pixel characteristic of original pixel vector, and formed background with second image and with it is described
The third image of the regional area of the corresponding pixel characteristic of original pixel vector;
S5 the circulation consistency loss of the first image and the third image) is calculated, and by the circulation consistency
Loss is updated to step S1) or step S4) map making processes current circulation consistency loss.
Specifically, can carry out parallel and can also connect the above method to continue with the above method, further includes:
S3 original pixel vector corresponding with the regional area of the first image) is obtained, is existed using the original pixel vector
The regional area for generating and there is pixel characteristic corresponding with the original pixel vector is re-mapped in the regional area, and forms tool
There is the regional area of pixel characteristic corresponding with the original pixel vector and the background with the first image the 4th figure
Picture;
S4 the reconstruction loss of the first image and the 4th image) is calculated, and reconstruction loss is updated to walk
Rapid S1) or step S3) map making processes current reconstruction loss.
Specifically, step S2) include:
Fighting network by production in the regional area using the pixel vectors and generate has and the pixel
The regional area of the corresponding pixel characteristic of vector.
Specifically, can carry out parallel and can also connect the above method to continue with the above method, further includes:
S3 the 5th image in the adjacent frame domain of the first image) is obtained, the first image is updated to described
5th image, go to step S1).
The embodiment of the present invention provides a kind of for changing the algorithm structure of image local area, which includes:
Production fights network, for will the first image that inputted and input pixel vectors the first image office
Mapping generates the regional area with pixel characteristic corresponding with the pixel vectors in portion region.
Optionally, the production model of the production confrontation network includes: encoder, hidden layer space vector and decoding
Device;
The input layer of the encoder is used to receive the data of the first image;
The hidden layer space vector is used to the data that the output layer of the encoder exports being transferred to the decoder
Input layer;
The hidden layer of the decoder is used to receive the data of the input layer output of the decoder and acquires and merge described
The data of the hidden layer of encoder;
Wherein, the hidden layer of the decoder is also used to receive the pixel vectors and exports and has and the pixel vectors pair
The regional area for the pixel characteristic answered;
The output layer of the decoder is used to export the regional area with pixel characteristic corresponding with the pixel vectors
And with the first image background the second image data.
Optionally, neuron relative data direction of transfer declines in dimension in the encoder;
The relatively described data direction of transfer of neuron rises in dimension in the decoder;
The hidden layers numbers of the encoder and the decoder are equal, and enabling the number of plies is N, sequence by the data direction of transfer
Sequence and N are positive integer;
I-th layer of neuron is used to acquire and merge (N-i in the hidden layer of the encoder in the hidden layer of the decoder
+ 1) data of the neuron of layer and the synchronous reception pixel vectors, i is positive integer and i is less than or equal to N.
Optionally, the encoder includes at least one layer of convolutional layer and at least two layers empty convolutional layer;
The decoder includes at least two layers empty convolutional layer and at least one layer of warp lamination.
Optionally, the production confrontation network has production model, and the production model includes encoder reconciliation
Code device;
The encoder includes at least one layer of convolutional layer and at least two layers empty convolutional layer;
The decoder includes at least two layers empty convolutional layer and at least one layer of warp lamination.
Optionally, the production model of the production confrontation network has the decoding of edge detection and function of image segmentation
Device;
The decoder is used to determine the regional area in the first image.
Optionally, the decoder includes: input layer, hidden layer, prediction model and output layer;
The prediction model is used to receive the data of the hidden layer output of the decoder and is also used to according to the decoder
The data of hidden layer output obtain the masking-out area of the regional area with pixel characteristic;
The prediction model is also used to calculate the mark masking-out of the masking-out area and the regional area in the first image
The regional area cross entropy is simultaneously updated to current regional area cross entropy by the regional area cross entropy in area.
Optionally, the production model of production confrontation network be used to form background with the first image and
Second image of the regional area with the pixel characteristic;
The discrimination model of the production confrontation network includes: to differentiate feature extraction convolutional layer, classifier and arbiter, institute
It states and differentiates that feature extraction convolutional layer receives the data of second image, the classifier receives the differentiation feature extraction convolution
The data and the synchronous data for receiving the differentiation feature extraction convolutional layer output of the arbiter of layer output;
The classifier is for classifying to the regional area of second image and obtaining classification results;
The arbiter is for differentiating the first image and second image and obtaining differentiation result;
The production confrontation network is also used to obtain the regional area with second image according to the classification results
Corresponding pixel vectors;
Production confrontation network be also used to calculate pixel vectors corresponding with the regional area of second image and
The pixel vectors cross entropy of the pixel vectors of input;
The production confrontation network is also used to obtain confrontation loss according to the differentiation result;
Production confrontation network is also used to for the pixel vectors cross entropy being updated to the current of the classifier
Pixel vectors cross entropy loses the current confrontation that the confrontation loss is updated to the arbiter.
Optionally, the production confrontation network is also used to be formed the background with the first image and has the picture
Second image of the regional area of plain feature;
The production confrontation network is also used to receive original pixel vector corresponding with the regional area of the first image;
The production confrontation network is also used to by using second image and the original pixel vector as current new
Input obtain the background with second image and the partial zones with pixel characteristic corresponding with the original pixel vector
The third image in domain;
The production confrontation network is also used to calculate the circulation consistency damage of the first image and the third image
It loses and the circulation consistency loss is updated to current circulation consistency and lose;
Production confrontation network is also used to by using the first image and the original pixel vector as currently
Input obtains the back regional area with pixel characteristic corresponding with the original pixel vector and with the first image
4th image of scape;
The reconstruction that production confrontation network is also used to calculate the first image and the 4th image lose and incite somebody to action
The reconstruction loss is updated to current reconstruction loss.
The embodiment of the present invention provides a kind of for changing the training method of image local area algorithm, the training method packet
It includes:
S1 mark sample data sets) are obtained, wherein the mark sample data sets include having different pixels feature
Object different images, the regional area of image corresponding with the object where the object have mark masking-out area and and also
With mark pixel vectors corresponding with pixel characteristic;
S2) building production fights network, and production confrontation network includes production model and discrimination model, described
Production model, which has, to be updated to present input data for history output data and obtains the feature and history input of partial data
The consistent circulatory function currently exported of the feature of data and the circulatory function recycle consistency for building and lose and rebuild
Loss, the production model also have edge detection and function of image segmentation, solution for constructing regional area cross entropy
Code device, and the discrimination model has the arbiter for constructing confrontation loss and the classification for constructing pixel vectors cross entropy
Device;
S3 the mark sample data sets) are input to the production confrontation network and start to train the production
Fight network, circulation consistency loss, reconstruction loss, the regional area cross entropy, the confrontation lose and
At least one in the pixel vectors cross entropy terminates to train when meeting preset condition, or in circulation consistency loss, institute
It states and rebuilds any at least two in loss, the regional area cross entropy, confrontation loss and the pixel vectors cross entropy
Meet the time sequencing for reaching preset threshold and terminate to train when meeting preset condition, obtains the production confrontation net of pre-training
Network.
In another aspect, the embodiment of the present invention provides a kind of equipment for changing image local area, comprising:
At least one processor;
Memory is connect at least one described processor;
Wherein, the memory is stored with the instruction that can be executed by least one described processor, described at least one
The instruction that device is stored by executing the memory is managed, at least one described processor passes through the finger for executing the memory storage
It enables and realizes method above-mentioned.
Another aspect, the embodiment of the present invention provide a kind of computer readable storage medium, are stored with computer instruction, work as institute
When stating computer instruction and running on computers, so that computer executes method above-mentioned.
Corresponding above content, the present invention are carried out by pixel vectors (its pixel characteristic characterizes colouring information) in regional area
Calculate, mapping generate have pixel characteristic regional area, with the prior art image processing in using shearing, covering or filling side
Formula is different, and common GAN model can not carry out mapping generation in regional area;
The present invention provides a kind of embodiments for determining regional area;
Production of the present invention fights network discrimination model, has classifier, can significantly improve the partial zones that mapping generates
The accuracy of the pixel characteristic in domain, and real-time synchronization learns together with arbiter, so that there is the present invention mapping to generate pixel spy
The accuracy of sign;
Production of the present invention fights network production model, has new decoder, can determine regional area in image
Relative position, and the masking-out area with regional area boundary characteristic can be exported, the masking-out area with there is mark boundary characteristic
Masking-out area calculated, can by cross entropy realize to determine regional area position learning process accuracy quantization, from
And the present invention has the accuracy for determining regional area position;
Production of the present invention fights network, has circulation consistency loss, can be used in the accuracy for determining regional area
The quantization of the accuracy of the pixel characteristic of the regional area generated with mapping, so that partial zones can be continuously improved significantly in the present invention
The accuracy of the pixel characteristic for the regional area that the accuracy and mapping that domain position determines generate;
Production of the present invention fights network, has and rebuilds loss, can be further used for determining the accuracy of regional area and
The quantization of the accuracy of the pixel characteristic of the regional area generated is mapped, thus the standard that there is the present invention mapping to generate generation area
The accuracy of the pixel characteristic for the regional area that true property and mapping generate;
The present invention provides a kind of implementations of mapping;
The present invention is capable of handling video image frame and dynamic image frame;
Production of the present invention fights network production model, constructs new decoder, which can acquire encoder
Middle data characteristics and the hidden layers numbers for possessing the number of plies identical as encoder, and the output layer of the decoder there are also it is parallel,
For the output layer in masking-out area, which can be constructed by prediction model;
The present invention constructs new production confrontation network, and is followed by the structure that the production fights network by setting
The ability progress of loss, regional area cross entropy, confrontation loss and pixel vectors cross entropy to network is rebuild in the loss of ring consistency
Training process is realized in description.
The other feature and advantage of the embodiment of the present invention will the following detailed description will be given in the detailed implementation section.
Detailed description of the invention
Attached drawing is to further understand for providing to the embodiment of the present invention, and constitute part of specification, under
The specific embodiment in face is used to explain the present invention embodiment together, but does not constitute the limitation to the embodiment of the present invention.Attached
In figure:
Fig. 1 is the change body color process schematic of the embodiment of the present invention;
Fig. 2 is the first production model schematic of the embodiment of the present invention;
Fig. 3 is the discrimination model schematic diagram of the embodiment of the present invention;
Fig. 4 is second of production model schematic of the embodiment of the present invention;
Fig. 5 is the model training process schematic of the embodiment of the present invention.
Specific embodiment
It is described in detail below in conjunction with specific embodiment of the attached drawing to the embodiment of the present invention.It should be understood that this
Locate described specific embodiment and be merely to illustrate and explain the present invention embodiment, is not intended to restrict the invention embodiment.
Embodiment 1
The embodiment of the present invention provides a kind of method for changing image local area, this method comprises: S1) obtain the first image
And pixel vectors, determine the regional area in the first image;S2 it) is reflected in the regional area using the pixel vectors
Penetrate the regional area for generating and there is pixel characteristic corresponding with the pixel vectors;
Such as Fig. 1, can enable regional area is vehicle body position, the first image is with background BG (containing two dotted line quadrangles in Fig. 1
The box of star) and automobile image x with body color aa(region in Fig. 1 containing oblique line is front windshield or vehicle window glass
Glass), the pixel characteristic of pixel vectors is that (order is color b) to color characteristic.
Specifically, step S1) in determine the first image in regional area, comprising:
S101 the object in the first image) is determined;
S102 the edge detection and figure about the object and the background of the first image) are carried out to the first image
As segmentation, the image with boundary characteristic is obtained;
S103) according to the relative position in the region where the object in the first image, determining has side described
Regional area corresponding with the relative position in the image of boundary's feature.
Specifically, step S1) in determine the first image in regional area, comprising:
The production that the first image is input to the decoder with edge detection and function of image segmentation is fought into net
Network fights network by the production and determines the regional area in the first image.
Specifically, step S2) include:
S201) utilize the pixel vectors in conjunction with the neuron of the decoder of production confrontation network in the part
The regional area with pixel characteristic corresponding with the pixel vectors is generated in region, and is formed with the first image
Second image of background and the regional area with the pixel characteristic;
S202) classified by the classifier that the production fights network to the regional area of second image,
It obtains classification results and obtains pixel vectors corresponding with the regional area of second image according to the classification results;
S203) calculate corresponding with the regional area of second image pixel vectors and step S1) pixel vectors
Pixel vectors cross entropy is obtained using the first image and second image in conjunction with the arbiter of production confrontation network
Loss must be fought, and the pixel vectors cross entropy is updated to the current pixel that the production fights the classifier of network
Vector cross entropy loses the current confrontation that the confrontation loss is updated to the arbiter of the production confrontation network;
Such as Fig. 5, the first image passes through production network encoder GencAfter become a hidden layer space vector Z, hidden layer space
Vector Z and color b are input to production network decoder G togetherdecThe automobile image for having color b is obtained afterwards, and the second image is
Automobile image with the first image background and with body color bAutomobile image xaAnd automobile imageBy differentiating mould
Type calculates current confrontation loss, meanwhile, automobile image xaAnd automobile imageColor classification, energy are also carried out by classifier C
Pixel vectors (color) cross entropy is obtained (in Fig. 5)。
Specifically, further include:
S3) acquisition participates in step S2) output data of the neuron of the decoders of map making processes, and according to institute
State the masking-out area that output data combination prediction model obtains the regional area with pixel characteristic, calculate the masking-out area and
The regional area cross entropy in the mark masking-out area of the regional area in the first image, and more by the regional area cross entropy
Regional area cross entropy newly current for the prediction model.
The ability of production network positioning vehicle body position in order to further increase, further promotes the convergence of model, into
The uniformity that one step increase is changed colour, production model generate automobile imageAfterwards, increase prediction model, prediction model can pass through
Decoder is constructed, hidden layer output position increases two-tier network (can be followed successively by warp lamination and convolutional layer), finally will
Prediction result (masking-out area) and the body of a motor car result manually marked (mark masking-out area) calculate vehicle body position cross entropy (partial zones
Domain cross entropy), the present invention has the segmentation ability of Automobile sound and background, can preferably position body of a motor car position, can be more
Good finds the position for needing to modify color.
Specifically, can carry out parallel and can also connect the above method to continue with the above method, further includes:
S3 the second image of background and the regional area with pixel characteristic with the first image) is formed,
And obtain original pixel vector corresponding with the regional area of the first image;
S4 generation) is re-mapped in the regional area with pixel characteristic using the original pixel vector has and institute
State the regional area of the corresponding pixel characteristic of original pixel vector, and formed background with second image and with it is described
The third image of the regional area of the corresponding pixel characteristic of original pixel vector;
S5 the circulation consistency loss of the first image and the third image) is calculated, and by the circulation consistency
Loss is updated to step S1) or step S4) map making processes current circulation consistency loss;
Production confrontation network after pre-training of the present invention also has circulation, automobile imageThis can also be passed through with color a
The production model of invention generates the automobile image for having color a againL1 canonical formula then can be used in (third image)
(two pictures are subtracted each other and taken absolute value by location of pixels by L1 canonical formula) calculates automobile imageWith automobile image xaCirculation
Consistency loss, the embodiment can preferably guarantee auto graph while modifying body color, background area it is constant
Property, it changes colour more acurrate.
Specifically, can carry out parallel and can also connect the above method to continue with the above method, further includes:
S3 original pixel vector corresponding with the regional area of the first image) is obtained, is existed using the original pixel vector
The regional area for generating and there is pixel characteristic corresponding with the original pixel vector is re-mapped in the regional area, and forms tool
There is the regional area of pixel characteristic corresponding with the original pixel vector and the background with the first image the 4th figure
Picture;
S4 the reconstruction loss of the first image and the 4th image) is calculated, and reconstruction loss is updated to walk
Rapid S1) or step S3) map making processes current reconstruction loss;
In the production confrontation network after pre-training of the present invention, hidden layer space vector Z is also input to life with color a together
Accepted way of doing sth network decoder GdecThe automobile image for having color a is obtained afterwards(the 4th image), the automobile image of generationAnd vapour
Vehicle image xaIt is calculated with L1 canonical formula and rebuilds loss.
Specifically, step S2) include:
Fighting network by production in the regional area using the pixel vectors and generate has and the pixel
The regional area of the corresponding pixel characteristic of vector.
Specifically, can carry out parallel and can also connect the above method to continue with the above method, further includes:
S3 the 5th image in the adjacent frame domain of the first image) is obtained, the first image is updated to described
5th image, go to step S1).
The present embodiment proposes new the production prototype network result for being more suitable for color alteration of automobile scene and discrimination model net
Network structure generates auto graph true to nature, increases the diversity of advertisement material;
It increases circulation consistency loss and improves the accuracy that model changes colour;
It introduces pixel vectors cross entropy, vehicle body segmentation loss (regional area cross entropy) and rebuilds loss, have vehicle body
The ability of segmentation, and make model in colourshifting process, increase the uniformity changed colour.
Embodiment 2
A kind of for changing the algorithm structure of image local area such as Fig. 1, which includes: production confrontation net
Network, for the pixel vectors of the first image inputted and input to be mapped generation in the regional area of the first image and had
The regional area of pixel characteristic corresponding with the pixel vectors.
Optionally, the production model of the production confrontation network includes: encoder, hidden layer space vector and decoding
Device;The input layer of the encoder is used to receive the data of the first image;The hidden layer space vector is used for the volume
The data of the output layer output of code device are transferred to the input layer of the decoder;The hidden layer of the decoder is for receiving the solution
Code device input layer output data and acquire and merge the encoder hidden layer data;Wherein, the decoder is hidden
Layer is also used to receive the pixel vectors and exports the regional area with pixel characteristic corresponding with the pixel vectors;It is described
The output layer of decoder be used for export the regional area with pixel characteristic corresponding with the pixel vectors and with described in
The data of second image of the background of the first image.
Optionally, neuron relative data direction of transfer declines in dimension in the encoder;It is neural in the decoder
The relatively described data direction of transfer of member rises in dimension;The hidden layers numbers of the encoder and the decoder are equal, enable the number of plies
For N, sequence sorts by the data direction of transfer and N is positive integer;I-th layer of neuron is used in the hidden layer of the decoder
It acquires and merges in the hidden layer of the encoder data of the neuron of (N-i+1) layer and synchronous receive the pixel vectors, i
For positive integer and i is less than or equal to N.
Optionally, the encoder includes at least one layer of convolutional layer and at least two layers empty convolutional layer;The decoder packet
Include at least two layers empty convolutional layer and at least one layer of warp lamination;
Since the region area that body of a motor car needs to change color is larger, if will cause using traditional convolutional network
Vehicle body receptive field is not big enough or the case where missing some tiny component (such as reflective mirrors) on vehicle body of causing to change colour, and in turn results in
The non-uniform situation of color alteration of automobile, so last two layers of the convolutional network of encoder described in the present embodiment and the decoder are initial
The network structure that two layers of convolutional network uses empty convolution does not reduce generation while increasing convolutional network receptive field
The resolution ratio for exporting image, so that model changes colour, effect is more evenly;
If Fig. 2-Fig. 4, Conv are convolutional layer, k, s and r are respectively core size, acquire size and sample rate, for example, <
Conv, k4s2r1 >, represent this layer be core having a size of 4, acquisition having a size of 2 and sample rate as 1 convolutional layer;As r in the present embodiment
When being 2 or 4, which is empty convolution, and empty convolution can increase the sense in current empty convolutional layer of upper layer output image
By open country;Deconv is warp lamination, and warp lamination is also a kind of convolutional layer, but warp lamination output image is with respect to the warp lamination
Input picture expands in pixel;
Preferably, production model includes: encoder, hidden layer space vector Z and decoder, and decoder acquires encoder
Export the data of image;
Preferably, encoder includes: encoder input layer<Conv, k4s2r1>, encoder hidden layer (encoder hidden layer packet
Include: first coding hidden layer<Conv, k4s2r1>, second coding hidden layer<Conv, k4s2r1>and third encode hidden layer<Conv,
K3s1r2>) and encoder output layer<Conv, k3s1r4>;
Preferably, decoder includes: input layer<Conv, k3s1r4>, hidden layer (hidden layer include: the first decoding hidden layer<
Conv, k3s1r2>, second decoding hidden layer<Deconv, k4s2r1>and third decode hidden layer<Deconv, k4s2r1>) and output
Layer<Deconv, k4s2r1>;
Preferably, the first decoding hidden layer<Conv, k3s1r2>acquisition third coding hidden layer<Conv, k3s1r2>output figure
The data of picture and pixel vectors are received simultaneously, second decoding hidden layer<Deconv, the coding hidden layer<Conv of k4s2r1>acquisition second,
K4s2r1>output image data and receive pixel vectors simultaneously, third decode hidden layer<Deconv, k4s2r1>acquisition first
Encode hidden layer<Conv, k4s2r1>output image data and receive pixel vectors simultaneously;
Preferably, output layer<Deconv of decoder, k4s2r1>acquisition encoder input layer<Conv, k4s2r1>it is defeated
The data of image and pixel vectors are received simultaneously out, can further promote the accuracy for position of changing colour;
Preferably, discrimination model includes: to differentiate input layer<Conv, k4s2r1>, differentiate that hidden layer (differentiates that hidden layer includes: the
One differentiate hidden layer<Conv, k4s2r1>, second differentiate hidden layer<Conv, k4s2r1>and third differentiate hidden layer<Conv, k3s1r2>)
With differentiate output layer (differentiate that output layer includes: empty convolutional layer<Conv, k3s1r4>, receive the cavity convolutional layer<Conv,
K3s1r4>output image arbiter<Fc, gan>, receive the cavity convolutional layer<Conv, the classification of k3s1r4>output image
Device<Fc, att>).
Optionally, the production model of the production confrontation network has the decoding of edge detection and function of image segmentation
Device;The decoder is used to determine the regional area in the first image.
Optionally, the decoder includes: input layer, hidden layer, prediction model and output layer;The prediction model is for connecing
It receives the data of the hidden layer output of the decoder and is also used to obtain the tool according to the data that the hidden layer of the decoder exports
There is the masking-out area (Mask prediction in Fig. 5) of the regional area of pixel characteristic;The prediction model is also used to calculate described
The regional area cross entropy in the mark masking-out area of the regional area in masking-out area and the first image is (in Fig. 5
Crossentropy) and by the regional area cross entropy it is updated to current regional area cross entropy;
Preferably, prediction model includes for receiving third decoding hidden layer<Deconv, k4s2r1>output image data
Warp lamination<Deconv, k4s2r1>and receive warp lamination<Deconv, the convolutional layers of k4s2r1>output image data<
Conv, k1s1r1>, the masking-out area of the vehicle body of convolutional layer<Conv of prediction model, k1s1r1>can export prediction (or be vehicle body
Exposure mask).
Embodiment 3
A kind of for changing the training method of image local area algorithm such as Fig. 5, which includes:
S1 mark sample data sets) are obtained, wherein the mark sample data sets include having different pixels feature
Object different images, the regional area of image corresponding with the object where the object have mark masking-out area and and also
With mark pixel vectors corresponding with pixel characteristic;
S2) building production fights network, and production confrontation network includes production model and discrimination model, described
Production model, which has, to be updated to present input data for history output data and obtains the feature and history input of partial data
The consistent circulatory function currently exported of the feature of data and the circulatory function recycle consistency for building and lose and rebuild
Loss, the production model also have edge detection and function of image segmentation, solution for constructing regional area cross entropy
Code device, and the discrimination model has the arbiter for constructing confrontation loss and the classification for constructing pixel vectors cross entropy
Device;
S3 the mark sample data sets) are input to the production confrontation network and start to train the production
Fight network, circulation consistency loss, reconstruction loss, the regional area cross entropy, the confrontation lose and
At least one in the pixel vectors cross entropy terminates to train when meeting preset condition, or in circulation consistency loss, institute
It states and rebuilds any at least two in loss, the regional area cross entropy, confrontation loss and the pixel vectors cross entropy
Meet the time sequencing for reaching preset threshold and terminate to train when meeting preset condition, obtains the production confrontation net of pre-training
Network;
1) the training data preparation stage
I. collect auto graph, and mark the color of each picture automobile, in the present embodiment altogether use it is orange, brown, white,
Purple, green, red, champagne, ash, Huang, black, blue color in totally 11;
Ii. the masking-out area of the vehicle body of auto graph in i is marked;
2) model training stage
Model training stage is carried out in the way of production model, discriminative model alternately training.
I. according to the process of Fig. 5, training picture, native color attribute are input in production network, color of object attribute
It is random to generate, the picture generated;
Ii. firstly generate formula model parameter it is constant, update discriminative model parameter, secondly discriminative model parameter constant,
Update production model parameter.
Wherein, in the training process, total lossesAre as follows:
It is lost for confrontation,To rebuild loss,For circulation consistency loss, λ, μ are constant coefficient, and F is linear letter
Number relationship in the training process, such as is related to parameter to circulation consistency loss and is trained, it may be assumed that
It is differentiated using total losses to loss corresponding to the model (production model or discrimination model) being currently trained to
The current trained degree of the model parameter being currently trained to can be obtained, so as to complete by reaching preset threshold condition
At current training, preset threshold condition can be arranged according to model using purpose, for example, picture material in later period use process
There is no other colors except sample, then loss reduction can be taken as threshold condition, but in other embodiments, and loss is most
Small usual presence and sample overfitting problem.
The optional embodiment of the embodiment of the present invention is described in detail in conjunction with attached drawing above, still, the embodiment of the present invention is simultaneously
The detail being not limited in above embodiment can be to of the invention real in the range of the technology design of the embodiment of the present invention
The technical solution for applying example carries out a variety of simple variants, these simple variants belong to the protection scope of the embodiment of the present invention.
It is further to note that specific technical features described in the above specific embodiments, in not lance
In the case where shield, it can be combined in any appropriate way.In order to avoid unnecessary repetition, the embodiment of the present invention pair
No further explanation will be given for various combinations of possible ways.
It will be appreciated by those skilled in the art that implementing the method for the above embodiments is that can pass through
Program is completed to instruct relevant hardware, which is stored in a storage medium, including some instructions are used so that single
Piece machine, chip or processor (processor) execute all or part of the steps of each embodiment the method for the application.And it is preceding
The storage medium stated includes: USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory
The various media that can store program code such as (RAM, Random Access Memory), magnetic or disk.
In addition, any combination can also be carried out between a variety of different embodiments of the embodiment of the present invention, as long as it is not
The thought of the embodiment of the present invention is violated, equally should be considered as disclosure of that of the embodiment of the present invention.
Claims (20)
1. a kind of method for changing image local area, which is characterized in that this method comprises:
S1 the first image and pixel vectors) are obtained, determine the regional area in the first image;
S2 generation) is mapped in the regional area using the pixel vectors has pixel corresponding with the pixel vectors special
The regional area of sign.
2. it is according to claim 1 change image local area method, which is characterized in that step S1) in determine described in
Regional area in first image, comprising:
S101 the object in the first image) is determined;
S102 the edge detection and image point about the object and the background of the first image) are carried out to the first image
It cuts, obtains the image with boundary characteristic;
S103) according to the relative position in the region where the object in the first image, determine has boundary special described
Regional area corresponding with the relative position in the image of sign.
3. it is according to claim 1 change image local area method, which is characterized in that step S1) in determine described in
Regional area in first image, comprising:
The production that the first image is input to the decoder with edge detection and function of image segmentation is fought into network, is led to
It crosses production confrontation network and determines regional area in the first image.
4. the method for change image local area according to claim 3, which is characterized in that step S2) include:
S201) utilize the pixel vectors in conjunction with the neuron of the decoder of production confrontation network in the regional area
It is interior to generate the regional area with pixel characteristic corresponding with the pixel vectors, and form the background with the first image
With the second image of the regional area with the pixel characteristic;
S202) classified by the classifier that the production fights network to the regional area of second image, obtained
Classification results simultaneously obtain pixel vectors corresponding with the regional area of second image according to the classification results;
S203 corresponding with the regional area of second image pixel vectors and step S1) are calculated) pixel vectors pixel
Vector cross entropy fights the arbiter acquisition pair of network using the first image and second image in conjunction with the production
Damage-retardation is lost, and the pixel vectors cross entropy is updated to the current pixel vectors that the production fights the classifier of network
Cross entropy loses the current confrontation that the confrontation loss is updated to the arbiter of the production confrontation network.
5. the method for change image local area according to claim 3, which is characterized in that further include:
S3) acquisition participates in step S2) output data of the neuron of the decoders of map making processes, and according to described defeated
Data combination prediction model obtains the masking-out area of the regional area with pixel characteristic out, calculates the masking-out area and described
The regional area cross entropy in the mark masking-out area of the regional area in the first image, and the regional area cross entropy is updated to
The current regional area cross entropy of the prediction model.
6. the method for change image local area according to claim 1, which is characterized in that further include:
S3 the second image of background and the regional area with pixel characteristic with the first image) is formed, and is obtained
Take original pixel vector corresponding with the regional area of the first image;
S4 generation) is re-mapped in the regional area with pixel characteristic using the original pixel vector has and the original
The regional area of the corresponding pixel characteristic of pixel vectors, and form the background with second image and have and the preimage
The third image of the regional area of the corresponding pixel characteristic of plain vector;
S5 the circulation consistency loss of the first image and the third image) is calculated, and the circulation consistency is lost
Be updated to step S1) or step S4) map making processes current circulation consistency loss.
7. the method for change image local area according to claim 1, which is characterized in that further include:
S3 original pixel vector corresponding with the regional area of the first image) is obtained, using the original pixel vector described
Re-mapped in regional area generate have pixel characteristic corresponding with the original pixel vector regional area, and formed have with
The regional area of the corresponding pixel characteristic of the original pixel vector and the background with the first image the 4th image;
S4 the reconstruction loss of the first image and the 4th image) is calculated, and reconstruction loss is updated to step S1)
Or step S3) map making processes current reconstruction loss.
8. the method for change image local area according to claim 1, which is characterized in that step S2) include:
Fighting network by production in the regional area using the pixel vectors and generate has and the pixel vectors
The regional area of corresponding pixel characteristic.
9. the method for change image local area according to claim 1, which is characterized in that further include:
S3 the 5th image in the adjacent frame domain of the first image) is obtained, the first image is updated to the described 5th
Image, go to step S1).
10. a kind of for changing the algorithm structure of image local area, which is characterized in that the algorithm structure includes:
Production fights network, for will the first image that inputted and input pixel vectors in the partial zones of the first image
Mapping generates the regional area with pixel characteristic corresponding with the pixel vectors in domain.
11. according to claim 10 for changing the algorithm structure of image local area, which is characterized in that
The production model of the production confrontation network includes: encoder, hidden layer space vector and decoder;
The input layer of the encoder is used to receive the data of the first image;
The hidden layer space vector is used to for the data that the output layer of the encoder exports being transferred to the input of the decoder
Layer;
The hidden layer of the decoder is used to receive the data of the input layer output of the decoder and acquires and merge the coding
The data of the hidden layer of device;
Wherein, the hidden layer of the decoder is also used to receive the pixel vectors and exports with corresponding with the pixel vectors
The regional area of pixel characteristic;
The output layer of the decoder be used for export have pixel characteristic corresponding with the pixel vectors regional area and
The data of second image of the background with the first image.
12. according to claim 11 for changing the algorithm structure of image local area, which is characterized in that
Neuron relative data direction of transfer declines in dimension in the encoder;
The relatively described data direction of transfer of neuron rises in dimension in the decoder;
The hidden layers numbers of the encoder and the decoder are equal, and the number of plies is enabled to sort for N, sequence by the data direction of transfer
And N is positive integer;
I-th layer of neuron is used to acquire and merge in the hidden layer of the encoder (N-i+1) in the hidden layer of the decoder
The data of the neuron of layer and the synchronous reception pixel vectors, i is positive integer and i is less than or equal to N.
13. according to claim 11 or 12 for changing the algorithm structure of image local area, which is characterized in that
The encoder includes at least one layer of convolutional layer and at least two layers empty convolutional layer;
The decoder includes at least two layers empty convolutional layer and at least one layer of warp lamination.
14. according to claim 10 for changing the algorithm structure of image local area, which is characterized in that
The production model of the production confrontation network has the decoder of edge detection and function of image segmentation;
The decoder is used to determine the regional area in the first image.
15. according to claim 14 for changing the algorithm structure of image local area, which is characterized in that
The decoder includes: input layer, hidden layer, prediction model and output layer;
The prediction model is used to receive the data of the hidden layer output of the decoder and is also used to according to the hidden of the decoder
The data of floor output obtain the masking-out area of the regional area with pixel characteristic;
The prediction model is also used to calculate the mark masking-out area of the masking-out area and the regional area in the first image
The regional area cross entropy is simultaneously updated to current regional area cross entropy by regional area cross entropy.
16. according to claim 10 for changing the algorithm structure of image local area, which is characterized in that
The production model of the production confrontation network is used to form the background with the first image and has the picture
Second image of the regional area of plain feature;
The discrimination model of the production confrontation network includes: to differentiate feature extraction convolutional layer, classifier and arbiter, described to sentence
It is defeated that other feature extraction convolutional layer receives the data of second image, the classifier reception differentiation feature extraction convolutional layer
Data and the arbiter out are synchronous to receive the data for differentiating the output of feature extraction convolutional layer;
The classifier is for classifying to the regional area of second image and obtaining classification results;
The arbiter is for differentiating the first image and second image and obtaining differentiation result;
The production confrontation network is also used to be obtained according to the classification results corresponding with the regional area of second image
Pixel vectors;
The production confrontation network is also used to calculate pixel vectors corresponding with the regional area of second image and input
Pixel vectors pixel vectors cross entropy;
The production confrontation network is also used to obtain confrontation loss according to the differentiation result;
The production confrontation network is also used to for the pixel vectors cross entropy being updated to the current pixel of the classifier
Vector cross entropy loses the current confrontation that the confrontation loss is updated to the arbiter.
17. according to claim 10 for changing the algorithm structure of image local area, which is characterized in that
The production confrontation network is also used to be formed the background with the first image and the office with the pixel characteristic
Second image in portion region;
The production confrontation network is also used to receive original pixel vector corresponding with the regional area of the first image;
Production confrontation network is also used to by using second image and the original pixel vector as currently new defeated
Enter to obtain the background with second image and the regional area with pixel characteristic corresponding with the original pixel vector
Third image;
The production confrontation network is also used to calculate the first image and the circulation consistency of the third image is lost, simultaneously
The circulation consistency loss is updated to current circulation consistency loss;
The production confrontation network is also used to by using the first image and the original pixel vector as current input
Obtain regional area with pixel characteristic corresponding with the original pixel vector and there is the background of the first image
4th image;
The production confrontation network is also used to calculate the reconstruction loss of the first image and the 4th image and will be described
It rebuilds loss and is updated to current reconstruction loss.
18. a kind of for changing the training method of image local area algorithm, which is characterized in that the training method includes:
S1 mark sample data sets) are obtained, wherein the mark sample data sets include pair with different pixels feature
The different images of elephant, the regional area of image corresponding with the object has mark masking-out area and also has where the object
Mark pixel vectors corresponding with pixel characteristic;
S2) building production fights network, and the production confrontation network includes production model and discrimination model, the generation
Formula model, which has, to be updated to present input data for history output data and obtains the feature and history input data of partial data
The consistent circulatory function currently exported of feature and the circulatory function for construct circulation consistency loss and rebuild loss,
The production model also has edge detection and function of image segmentation, decoder for constructing regional area cross entropy,
And the discrimination model has the arbiter for constructing confrontation loss and the classifier for constructing pixel vectors cross entropy;
S3 the mark sample data sets) are input to the production confrontation network and start that the production is trained to fight
Network, circulation consistency loss, reconstruction loss, the regional area cross entropy, the confrontation loss and it is described
At least one in pixel vectors cross entropy terminates to train when meeting preset condition, or the circulation consistency loss, it is described heavy
Build any at least two satisfaction in loss, the regional area cross entropy, confrontation loss and the pixel vectors cross entropy
Reach the time sequencing of preset threshold and terminate to train when meeting preset condition, obtains the production confrontation network of pre-training.
19. a kind of equipment for changing image local area characterized by comprising
At least one processor;
Memory is connect at least one described processor;
Wherein, the memory is stored with the instruction that can be executed by least one described processor, at least one described processor
Instruction by executing the memory storage is realized described in any one of claim 1 to 9, claim 18 claim
Method.
20. a kind of computer readable storage medium, is stored with computer instruction, when the computer instruction is run on computers
When, so that computer perform claim requires method described in any one of 1 to 9, claim 18 claim.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910465289.7A CN110298898B (en) | 2019-05-30 | 2019-05-30 | Method for changing color of automobile image body and algorithm structure thereof |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910465289.7A CN110298898B (en) | 2019-05-30 | 2019-05-30 | Method for changing color of automobile image body and algorithm structure thereof |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110298898A true CN110298898A (en) | 2019-10-01 |
CN110298898B CN110298898B (en) | 2023-05-23 |
Family
ID=68027368
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910465289.7A Active CN110298898B (en) | 2019-05-30 | 2019-05-30 | Method for changing color of automobile image body and algorithm structure thereof |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110298898B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110913193A (en) * | 2019-11-25 | 2020-03-24 | 深圳市创凯智能股份有限公司 | Image processing method, device, apparatus and computer readable storage medium |
CN110956505A (en) * | 2019-12-04 | 2020-04-03 | 腾讯科技(深圳)有限公司 | Advertisement inventory estimation method and related device |
CN111429544A (en) * | 2020-03-23 | 2020-07-17 | 北京字节跳动网络技术有限公司 | Vehicle body color processing method and device and electronic equipment |
CN111462264A (en) * | 2020-03-17 | 2020-07-28 | 中国科学院深圳先进技术研究院 | Medical image reconstruction method, medical image reconstruction network training method and device |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102254324A (en) * | 2011-07-18 | 2011-11-23 | 中兴通讯股份有限公司 | Local image translating method and terminal with touch screen |
CN104091320A (en) * | 2014-07-16 | 2014-10-08 | 武汉大学 | Noise human face super-resolution reconstruction method based on data-driven local feature conversion |
CN104200233A (en) * | 2014-06-24 | 2014-12-10 | 南京航空航天大学 | Clothes classification and identification method based on Weber local descriptor |
CN107330956A (en) * | 2017-07-03 | 2017-11-07 | 广东工业大学 | A kind of unsupervised painting methods of caricature manual draw and device |
US20170372193A1 (en) * | 2016-06-23 | 2017-12-28 | Siemens Healthcare Gmbh | Image Correction Using A Deep Generative Machine-Learning Model |
CN109635874A (en) * | 2018-12-19 | 2019-04-16 | 中国人民解放军军事科学院评估论证研究中心 | A kind of Image Blind evaluation method and device of characteristics dictionary mapping |
-
2019
- 2019-05-30 CN CN201910465289.7A patent/CN110298898B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102254324A (en) * | 2011-07-18 | 2011-11-23 | 中兴通讯股份有限公司 | Local image translating method and terminal with touch screen |
CN104200233A (en) * | 2014-06-24 | 2014-12-10 | 南京航空航天大学 | Clothes classification and identification method based on Weber local descriptor |
CN104091320A (en) * | 2014-07-16 | 2014-10-08 | 武汉大学 | Noise human face super-resolution reconstruction method based on data-driven local feature conversion |
US20170372193A1 (en) * | 2016-06-23 | 2017-12-28 | Siemens Healthcare Gmbh | Image Correction Using A Deep Generative Machine-Learning Model |
CN107330956A (en) * | 2017-07-03 | 2017-11-07 | 广东工业大学 | A kind of unsupervised painting methods of caricature manual draw and device |
CN109635874A (en) * | 2018-12-19 | 2019-04-16 | 中国人民解放军军事科学院评估论证研究中心 | A kind of Image Blind evaluation method and device of characteristics dictionary mapping |
Non-Patent Citations (2)
Title |
---|
LEIZHANG等: "Two-stage image denoising by principal component analysis with local pixel grouping", 《PATTERN RECOGNITION》 * |
陈文兵等: "基于条件生成式对抗网络的数据增强方法", 《计算机应用》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110913193A (en) * | 2019-11-25 | 2020-03-24 | 深圳市创凯智能股份有限公司 | Image processing method, device, apparatus and computer readable storage medium |
CN110956505A (en) * | 2019-12-04 | 2020-04-03 | 腾讯科技(深圳)有限公司 | Advertisement inventory estimation method and related device |
CN111462264A (en) * | 2020-03-17 | 2020-07-28 | 中国科学院深圳先进技术研究院 | Medical image reconstruction method, medical image reconstruction network training method and device |
CN111462264B (en) * | 2020-03-17 | 2023-06-06 | 中国科学院深圳先进技术研究院 | Medical image reconstruction method, medical image reconstruction network training method and device |
CN111429544A (en) * | 2020-03-23 | 2020-07-17 | 北京字节跳动网络技术有限公司 | Vehicle body color processing method and device and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
CN110298898B (en) | 2023-05-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110298898A (en) | Change the method and its algorithm structure of automobile image body color | |
Tan et al. | ArtGAN: Artwork synthesis with conditional categorical GANs | |
Baldassarre et al. | Deep koalarization: Image colorization using cnns and inception-resnet-v2 | |
CN108875935B (en) | Natural image target material visual characteristic mapping method based on generation countermeasure network | |
CN110378985B (en) | Animation drawing auxiliary creation method based on GAN | |
CN109671018A (en) | A kind of image conversion method and system based on production confrontation network and ResNets technology | |
CN108961350B (en) | Wind painting migration method based on saliency matching | |
CN108537743A (en) | A kind of face-image Enhancement Method based on generation confrontation network | |
CN108830913B (en) | Semantic level line draft coloring method based on user color guidance | |
Wang et al. | A survey on face data augmentation | |
CN109903236A (en) | Facial image restorative procedure and device based on VAE-GAN to similar block search | |
CN113807265B (en) | Diversified human face image synthesis method and system | |
Li et al. | Globally and locally semantic colorization via exemplar-based broad-GAN | |
CN109359527B (en) | Hair region extraction method and system based on neural network | |
CN110263865A (en) | A kind of semi-supervised multi-modal multi-class image interpretation method | |
CN110222837A (en) | A kind of the network structure ArcGAN and method of the picture training based on CycleGAN | |
CN108564120A (en) | Feature Points Extraction based on deep neural network | |
CN110895795A (en) | Improved semantic image inpainting model method | |
CN107977605A (en) | Ocular Boundary characteristic extraction method and device based on deep learning | |
CN116129013A (en) | Method, device and storage medium for generating virtual person animation video | |
Oza et al. | Semi-supervised image-to-image translation | |
CN107833198B (en) | Outdoor scene relighting method based on large-scale decomposition | |
Oza et al. | Automatic image colorization using ensemble of deep convolutional neural networks | |
Coutinho et al. | On the challenges of generating pixel art character sprites using GANs | |
CN115018729A (en) | White box image enhancement method for content |
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 |