CN110232652A - Image processing engine processing method, the image processing method for terminal, terminal - Google Patents
Image processing engine processing method, the image processing method for terminal, terminal Download PDFInfo
- Publication number
- CN110232652A CN110232652A CN201910445967.3A CN201910445967A CN110232652A CN 110232652 A CN110232652 A CN 110232652A CN 201910445967 A CN201910445967 A CN 201910445967A CN 110232652 A CN110232652 A CN 110232652A
- Authority
- CN
- China
- Prior art keywords
- image processing
- image
- loss function
- intelligent terminal
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T1/00—General purpose image data processing
- G06T1/20—Processor architectures; Processor configuration, e.g. pipelining
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/94—Hardware or software architectures specially adapted for image or video understanding
- G06V10/955—Hardware or software architectures specially adapted for image or video understanding using specific electronic processors
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image processing engine processing methods, the image processing method for terminal, terminal, comprising: training generates at least two image processing models;Loss function and network parameter that multiple images processing model generates are obtained, the logical relation between loss function and overall loss function, network parameter is established, overall loss function is optimized;Image processing model network parameter is adjusted by overall loss function, the present invention uses transfer of learning technology, image processing module is designed to the intelligentized intelligent terminal image processing engine of a height, it can be with hoisting module reusing degree and portability, the image procossing intelligent engine based on AI technology integrates natural language processing simultaneously, it can be camera, picture library, social application etc. increases new function, underlying algorithm is provided to support, convenient for developing personalized function, increase product function attraction, promote the frequency used the product by the user, increase the interest that user uses mobile phone, improve user experience.
Description
Technical field
The present invention relates to intelligent terminal technical field, it is specifically related to a kind of image processing engine processing method, for eventually
The image processing method at end, terminal.
Background technique
In recent years, with the fast development of deep learning, depth learning technology is more wide in the application of field of image processing
It is general, and some beneficial outcomes, such as natural language processing are obtained in fields such as image procossing, machine vision, natural language processings
Field, the BERT that google is proposed;Image generates field, tall and handsome up to face generating algorithm (StyleGAN) disclosed in company;Machine
Device visual field, the extensive use and improvement of various convolutional neural networks variants.
It can be trained for carrying out at transformation picture based on all kinds of generations confrontation network (GAN) for generating confrontation mechanism
Reason, including generations of high definition picture, artistic style conversion, Character generation diagram piece, image restoration and image are converted, such as satellite mapping and
Topographic map mutually turns, and graph mutually turns with sketch map, and the Transformer mechanism based on attention (Attention) can train
For the task of picture talk, i.e., a given picture generates the description and understanding of scene shown to picture, but current intelligence
Picture library, photograph album application still use traditional image processing method, the image processing effect that can be provided in energy terminal such as mobile phone
It is limited, more intelligent and personalized method is needed to improve and be promoted the treatment effect of image, for camera, picture library, social product
Deng use.
Summary of the invention
It is an object of the invention to overcome, the image processing effect that intelligent terminal can provide in the prior art is limited, at image
The technical problem for managing effect difference provides a kind of image processing engine processing method, the image processing method for terminal, terminal.
To achieve the above object, The technical solution adopted by the invention is as follows: a kind of processing method of AI image processing engine,
The described method includes:
At least two image processing models are generated based on transfer learning technique drill;
Loss function and network parameter that multiple images processing model generates are obtained, loss function and overall loss letter are established
Logical relation between number, network parameter, optimizes overall loss function;
When overall loss function meets preset condition, by overall loss function to the network parameter of image processing model
It is adjusted.
Further, overall loss function is optimized and includes:
Calculate loss function and network parameter that each image processing model generates;
Feature extraction is carried out to input picture and output image using convolutional network model, convolutional network model is calculated and generates
Loss function and network parameter;
Overall loss function is calculated, wherein overall loss function is the loss function and volume that multiple images handle that model generates
The total and/or weighted value for the loss function that product network model generates;
When overall loss function minimization, joined according to the network that overall loss function adjusts each image processing model
Number.
Further, described image processing model includes confrontation network model and Transformer model.
Further, the confrontation network model includes Style Transfer network model, high-definition picture generation network mould
Type, image completion restore one of network model or a variety of.
A kind of image processing method for intelligent terminal, which comprises
Obtain the image information of pretreatment image;
The image information of pretreatment image is sent to AI image processing engine to handle image information;
Image after exporting optimization processing.
Further, the AI image processing engine migrates to intelligent terminal.
Further, the AI image processing engine migrates to before intelligent terminal to image in the AI image processing engine
Processing model one of is compressed, cut or a variety of optimization processings.
Further, the AI image processing engine migrate to intelligent terminal include will treated AI image processing engine
Middle image processing model and OpenCV module are integrated, are compiled as a dynamic link library for calling.
Further, the compression includes optimizing down-sampled, weight trimming, avoiding full articulamentum, reduce number of channels, subtract
One of small convolution kernel size, discretization weight are a variety of.
A kind of intelligent terminal, the intelligent terminal include:
Module is obtained, for obtaining the image information of pretreatment image;
Calling module, for calling AI image processing engine to concentrate at pretreatment image the Intelligent terminal data
Reason;
Output module, for exporting the image after optimization processing.
Further, the intelligent terminal further includes OpenCV module, integrated collector, and the integrated collector is used
Dynamic link library is compiled as in integrating image processing model in the AI image processing engine and the OpenCV module.
Further, the intelligent terminal includes one of mobile phone, Pad, palm PC or a variety of.
By the above-mentioned description of this invention it is found that compared with prior art, a kind of image processing engine provided by the invention
Processing method, the image processing method for terminal, terminal generate at least two image procossings based on transfer learning technique drill
Model;Obtain loss function and network parameter that multiple images processing model generates, establish loss function and overall loss function,
Logical relation between network parameter optimizes overall loss function;When overall loss function meets preset condition, lead to
It crosses overall loss function to be adjusted the network parameter of image processing model, using transfer of learning technology, by image procossing mould
Block is designed to the intelligentized intelligent terminal image processing engine of a height, can be with hoisting module reusing degree and portability, together
When the image procossing intelligent engine based on AI technology integrate natural language processing, computer vision and field of image processing at
Ripe algorithm can increase new function for camera, picture library, social application etc., provide underlying algorithm support, convenient for developing personalized function
Can, increase product function attraction, promote the frequency used the product by the user, increases the interest that user uses mobile phone, improve
User experience.
Detailed description of the invention
Fig. 1 is the processing method flow chart of AI image processing engine of the present invention;
Fig. 2 is image processing model optimized flow chart of the present invention;
Fig. 3 is image processing model optimized flow chart of the present invention;
Fig. 4 is the image processing method flow chart that the embodiment of the present invention one is used for intelligent terminal;
Fig. 5 is one intelligent terminal structural block diagram of the embodiment of the present invention.
Fig. 6 is the image processing method flow chart that the embodiment of the present invention two is used for intelligent terminal;
Fig. 7 is two intelligent terminal structural block diagram of the embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.Intelligent terminal can be mobile phone, Pad, any one in palm, this implementation
Example, which takes the mobile phone as an example, to be illustrated:
Specific embodiment one:
As shown in Figure 1, a kind of processing method of AI image processing engine, the described method comprises the following steps:
S101: generating at least two image processing models based on transfer learning technique drill,
Wherein, image processing model includes confrontation network model and Transformer model, the confrontation network model packet
Include Style Transfer network model, high-definition picture generates network model, image completion restores one of network model or more
Kind, image understanding and verbal description function may be implemented in Transformer model;
S102: loss function and network parameter that multiple images processing model generates are obtained, loss function and totality are established
Logical relation between loss function, network parameter optimizes overall loss function,
Specifically, calculating loss function and network parameter that each image processing model generates;Using convolutional network model
Feature extraction is carried out to input picture and output image, calculates loss function and network parameter that convolutional network model generates;Meter
Overall loss function is calculated, wherein overall loss function is the loss function and convolutional network model that multiple images handle that model generates
The total and/or weighted value of the loss function of generation;
S103: when overall loss function meets preset condition, by overall loss function to the net of image processing model
Network parameter is adjusted,
Specifically, adjusting each image processing model according to overall loss function when overall loss function minimization
Network parameter.
In the present embodiment, in order on mobile phone realize Style Transfer, high-definition picture generate, image completion it is multiple
Former and image understanding and verbal description function, the present embodiment are trained this four network models, and specific embodiment is as follows:
As shown in Fig. 2, training data is expressed as (X, Y), using CycleGAN network model to Style Transfer network model
Be trained, wherein network parameter is W1, loss function Loss_Style, Loss_Style be expressed as Loss_Style (W1 |
X, Y);
Network model is generated to high-definition picture using PGGAN network model to be trained, wherein network parameter is W2,
Loss function is Loss_Resolution, and Loss_Resolution is expressed as Loss_Resolution (W2 | X, Y);
It restores network model to image completion using SRGAN network model to be trained, wherein network parameter is W3, loss
Function is Loss_Restore, and Loss_Restore is expressed as Loss_Restore (W3 | X, Y);
Transformer model is trained using BERT or GPT network model, wherein network parameter is W4, loss
Function is Loss_Transoformer, and Loss_Transoformer is expressed as Loss_Transoformer (W4 | X, Y);
Feature extraction is carried out to input picture using convolutional network MODEL C NN-A, vector FeatureA is denoted as, using convolution
Network model CNN-B carries out feature extraction to the image of output, is denoted as vector FeatureB, loss function Loss_
Difference=F (FeatureA, Feature | W5), wherein F is based on similar Euclidean distance between two features
Or the loss function of cosine similarity measurement, W5 are the splicing vector of two convolutional network parameters of CNN-A and CNN-B, this damage
Function Loss_Difference is lost for ensuring input picture and carrying out high-resolution generation and image restoration, completion etc.
Output picture material after conversion is able to maintain higher consistency between higher level feature, is unlikely to serious distortion, such as A article
Will still appear as A article, only article A details after network is converted has promotion, wherein convolutional network MODEL C NN-A,
AlexNet, LeNet etc. can be used in CNN-B;
The loss function of each model is added up, an overall loss function is designed, to overall loss function Loss_
Total is optimized, so that Loss_Total is minimized, specifically, Loss_Total=Loss_Style+Loss_
Resolution+Loss_Restore+Loss_Transoformer+Loss_Differenc e, is expressed as Loss_Total
(W1, W2, W3, W4 | X, Y),
When optimization, random initializtion network parameter is the value N (0, I) for meeting normal distribution;
As shown in figure 3, calculating the loss function Loss_Style, Loss_ of each network model when propagated forward
Resolution, Loss_Restore, Loss_Transoformer, Loss_Difference finally calculate Loss_Total;
Using back-propagation algorithm, gradient dW1, dW2, the dW3 of relatively each network parameter of Loss_Total function are calculated,
DW4, dW5 update each network model, so that:
W1:=W1+a1*dW1
W2:=W1+a2*dW1
W3:=W3+a3*dW3
W4:=W4+a4*dW4
W5:=W5+a5*dW5
Wherein, a1, a2, a3, a4, a5 are learning rate, and value range is [1e-6,1e-2].
W1, W2, W3 are updated, W4 after W5 parameter, calculates the value of Loss_Total, then calculates the gradient of each parameter again, weight
Multiple above procedure Loss_Total terminates iteration until no longer declining, and combines overall loss function Loss_ in the process
Variation tendency in the training process of Total and each network losses function assesses network overall performance, when reaching default the number of iterations
Shi Xunlian is completed.
After having trained, AI image processing engine is designed as a stand-alone program and is run on Cloud Server, and in cloud
Restfull api interface is arranged on server to call for mobile phone, mobile phone there must be data connection and support http agreement at this time, need
Higher data connection is wanted, calls AI image processing engine to handle pretreatment image by Restfull api interface.
As shown in figure 4, the present embodiment also provides a kind of image processing method for intelligent terminal, the method includes with
Lower step:
S201: the image information of pretreatment image is obtained;
S202: the image information of pretreatment image is sent to AI image processing engine, image information is handled;
S203: the image after output optimization processing.
Specific embodiment is as follows: obtaining the image information of pretreatment image, can be camera on mobile phone, picture library, social activity
Image on product, mobile phone and Cloud Server establish data connection, are called by the Restfull api interface on Cloud Server
AI image processing engine handles image information, and exports the image after optimization processing.
As shown in figure 5, the present embodiment also provides a kind of intelligent terminal, the intelligent terminal includes obtaining module 201, calling
Module 202, output module 203,
The image information that module 201 obtains pretreatment image is obtained, wherein pre-processed image information can be phase on mobile phone
Machine, picture library, the image on social product;
Calling module 202 calls AI image processing engine to concentrate at pretreatment image the Intelligent terminal data
Reason;
Output module 203 is for exporting the image after optimization processing.
Specific embodiment two:
Although the function of smart phone is stronger and stronger, their computing capability, battery life and available disk space
It is still extremely limited, therefore this implementation and the difference of embodiment one are that the present embodiment is in AI image processing engine at image
Image processing model is handled after the completion of reason model training, so that AI image processing engine can be applied to intelligent terminal
On.It is specific as shown in figure 4,
As shown in fig. 6, a kind of image processing method for intelligent terminal, the described method comprises the following steps:
S301: to image processing model in AI image processing engine one of compressed, cut or a variety of optimizations at
Reason;
S302: migrating to intelligent terminal for AI image processing engine,
Specifically, will treated that image processing model is integrated with OpenCV module, be compiled as a dynamic link library with
For calling
S303: the image information of pretreatment image is obtained;
S304: the image information of pretreatment image is sent to AI image processing engine, image information is handled;
S305: the image after output optimization processing.
Specific embodiment is as follows: compression can be trimmed using down-sampled, weight is optimized, avoid full articulamentum, reduction logical
One of road quantity, reduction convolution kernel size, discretization weight are a variety of, and cutting can be respectively on the intelligent terminal
IMAGENET and MSCOCO data set on accurate adjustment;
Optimize down-sampled: the pondization operation of layer and fixed quantity for fixed quantity, neural network may show
Big difference, this is because the characterization of data and calculation amount size, which depend on these ponds, operates in where completion, if pond
Change the relatively early completion of operation, the dimension of data can be reduced.Dimension is fewer, and the processing speed of network is faster, if the pond in network
Operation completion is later, then most information can be retained, calculating speed is slower, therefore is evenly arranged in neural network
Down-sampled is a kind of effective structure, and good balance can be kept between accuracy and speed;
Weight trimming: in a trained neural network, activation of some weights for some neuron elements
It is worth most important, and other weight has substantially no effect on result.Nevertheless, we are still to the weight less important to these
Some calculating are done, trimming (pruning) is the process for deleting minimum strength connection completely, we can skip this in this way
It is a little to calculate.This can reduce accuracy but network can be allowed faster more to simplify, and leave out in the case where not influencing accuracy as far as possible to the greatest extent
Connection more than possible, such as leave out most weak connection to save the calculating time and space;
Avoid connecting entirely: full articulamentum is the most common part in neural network, they can usually play great role.So
And since each neuron is connected with all neurons of preceding layer, they need to store and update a large amount of ginsengs
Number, this is all very unfavorable to speed and disk space.Convolutional layer is using the layer of locally coherence in input, and therefore, each is refreshing
Through member be no longer connected with all neurons of preceding layer, this facilitate network reduced while keeping high degree of accuracy connection/
Connection/weight quantity of the quantity of weight, full articulamentum will use few connection or the non-layer energy connected entirely far more than convolutional layer
The volume of model is reduced, while keeping its high accuracy, speed can be improved in this method, while reducing disk usage amount;
Reduce number of channels, reduce convolution kernel size: the convolutional layer for possessing a large amount of channels can make network extract relevant information,
But corresponding cost is also paid, rejecting some characteristic patterns is the straightforward procedure for saving space, acceleration model, passes through diminution
Convolution kernel size, convolution reduce the perception of local mode, can reduce calculating cost;
Discretization weight: in order to save neural network in disk, needing to record the value of each weight in network, this meaning
Taste need to save a floating number for each parameter, while also implying that the consumption of a large amount of disk spaces, for example,
A floating number occupies 4 bytes, i.e., 32 in C.One has the network (such as Google-Net or VGG-16) of more than one hundred million parameters
Can readily take up up to a hundred Mbytes of space, and such consumption be in a mobile device it is unacceptable, in order to minimize
The amount of network storage can reduce the precision of weight by discretization weight, in this process, change the expression of number
So that it is no longer indicated occurrence, but limits the subset that it is numerical value.It only needs to store the primary value Jing Guo discretization in this way, so
It is mapped them into the weight of network afterwards, the storage of discretization weight indexes rather than floating point values, reduces calculate cost in this way;
Cutting can accurate adjustment on IMAGENET the and MSCOCO data set on mobile phone respectively;
Performance of the performance indicator of image processing model after optimization processing compared to the image processing model that training generates
Index decline is not higher than 5%, and judgement meets scheduled performance indicator;
Image processing model after processing is migrated into intelligent terminal, specifically, will treated image processing model with
OpenCV module is integrated, is compiled as a dynamic link library for calling;
The image information for obtaining pretreatment image can be camera on mobile phone, picture library, the image on social product, directly
AI image processing engine handles image information on calling mobile phone, and exports the image after optimization processing.
As shown in fig. 7, the present embodiment also provides a kind of intelligent terminal, the intelligent terminal includes that the intelligent terminal includes
OpenCV module 301, integrated collector 302, obtain module 303, calling module 304, output module 305,
The integrated collector 302 is by image processing model in the AI image processing engine and the OpenCV module
301 integrated are compiled as dynamic link library;
The image information that module 303 obtains pretreatment image is obtained, wherein pre-processed image information can be phase on mobile phone
Machine, picture library, the image on social product;
Calling module 304 calls AI image processing engine to concentrate at pretreatment image the Intelligent terminal data
Reason;
Output module 305 is for exporting the image after optimization processing
By the above-mentioned description of this invention it is found that compared with prior art, a kind of image processing engine provided by the invention
Processing method, the image processing method for terminal, terminal generate at least two image procossings based on transfer learning technique drill
Model;Obtain loss function and network parameter that multiple images processing model generates, establish loss function and overall loss function,
Logical relation between network parameter optimizes overall loss function;When overall loss function meets preset condition, lead to
It crosses overall loss function to be adjusted the network parameter of image processing model, using transfer of learning technology, by image procossing mould
Block is designed to the intelligentized intelligent terminal image processing engine of a height, can be with hoisting module reusing degree and portability, together
When the image procossing intelligent engine based on AI technology integrate natural language processing, computer vision and field of image processing at
Ripe algorithm can increase new function for camera, picture library, social application etc., provide underlying algorithm support, convenient for developing personalized function
Can, increase product function attraction, promote the frequency used the product by the user, increases the interest that user uses mobile phone, improve
User experience.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, of the invention
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (12)
1. a kind of processing method of AI image processing engine, which is characterized in that the described method includes:
At least two image processing models are generated based on transfer learning technique drill;
Obtain loss function and network parameter that multiple images processing model generates, establish loss function and overall loss function,
Logical relation between network parameter optimizes overall loss function;
When overall loss function meets preset condition, carried out by network parameter of the overall loss function to image processing model
Adjustment.
2. the processing method of AI image processing engine according to claim 1, it is characterised in that: to overall loss function into
Row optimizes
Calculate loss function and network parameter that each image processing model generates;
Feature extraction is carried out to input picture and output image using convolutional network model, calculates the damage that convolutional network model generates
Lose function and network parameter;
Overall loss function is calculated, wherein overall loss function is the loss function and convolution net that multiple images handle that model generates
The total and/or weighted value for the loss function that network model generates;
When overall loss function minimization, the network parameter of each image processing model is adjusted according to overall loss function.
3. the processing method of AI image processing engine according to claim 1, it is characterised in that: described image handles model
Including confrontation network model and Transformer model.
4. the processing method of AI image processing engine according to claim 3, it is characterised in that: the confrontation network model
Generate network model including Style Transfer network model, high-definition picture, image completion restores one of network model or
It is a variety of.
5. a kind of image processing method for intelligent terminal, which is characterized in that the described method includes:
Obtain the image information of pretreatment image;
The image information of pretreatment image is sent to AI image processing engine to handle image information;
Image after exporting optimization processing.
6. being used for the image processing method of intelligent terminal according to claim 5, it is characterised in that: the AI image procossing draws
It holds up and migrates to intelligent terminal.
7. being used for the image processing method of intelligent terminal according to claim 6, it is characterised in that: the AI image procossing draws
It holds up and image processing model in the AI image processing engine one of is compressed, cut before migrating to intelligent terminal or more
Kind optimization processing.
8. the image processing method according to claim 6 or 7 for intelligent terminal, it is characterised in that: the AI image
Processing engine migrate to intelligent terminal include will image processing model and OpenCV module in treated AI image processing engine
It is integrated, a dynamic link library is compiled as calling.
9. the image processing method according to claim 5 for intelligent terminal, it is characterised in that: the compression includes excellent
One of change down-sampled, weight trimming, avoid full articulamentum, reduce number of channels, reduce convolution kernel size, discretization weight
Or it is a variety of.
10. a kind of intelligent terminal, it is characterised in that: the intelligent terminal includes
Module is obtained, for obtaining the image information of pretreatment image;
Calling module, for calling AI image processing engine to concentrate pretreatment image to handle the Intelligent terminal data;
Output module, for exporting the image after optimization processing.
11. intelligent terminal according to claim 10, it is characterised in that: the intelligent terminal further include OpenCV module,
Integrated collector, the integrated collector be used for by image processing model in the AI image processing engine with it is described
OpenCV module is integrated to be compiled as dynamic link library.
12. intelligent terminal described in 0-11 any one according to claim 1, it is characterised in that: the intelligent terminal includes hand
One of machine, Pad, palm PC are a variety of.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910445967.3A CN110232652A (en) | 2019-05-27 | 2019-05-27 | Image processing engine processing method, the image processing method for terminal, terminal |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910445967.3A CN110232652A (en) | 2019-05-27 | 2019-05-27 | Image processing engine processing method, the image processing method for terminal, terminal |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110232652A true CN110232652A (en) | 2019-09-13 |
Family
ID=67858156
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910445967.3A Pending CN110232652A (en) | 2019-05-27 | 2019-05-27 | Image processing engine processing method, the image processing method for terminal, terminal |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110232652A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110968671A (en) * | 2019-12-03 | 2020-04-07 | 北京声智科技有限公司 | Intent determination method and device based on Bert |
CN111986075A (en) * | 2020-08-12 | 2020-11-24 | 兰州交通大学 | Style migration method for target edge clarification |
CN112232594A (en) * | 2020-11-06 | 2021-01-15 | 中国人民解放军国防科技大学 | Multi-satellite emergency observation task planning method and device and computer equipment |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2017077121A1 (en) * | 2015-11-06 | 2017-05-11 | Thomson Licensing | Method for transfer of a style of a reference visual object to another visual object, and corresponding electronic device, computer readable program products and computer readable storage medium |
CN108647583A (en) * | 2018-04-19 | 2018-10-12 | 浙江大承机器人科技有限公司 | A kind of face recognition algorithms training method based on multiple target study |
CN109508654A (en) * | 2018-10-26 | 2019-03-22 | 中国地质大学(武汉) | Merge the human face analysis method and system of multitask and multiple dimensioned convolutional neural networks |
CN109523463A (en) * | 2018-11-20 | 2019-03-26 | 中山大学 | A kind of face aging method generating confrontation network based on condition |
CN109544442A (en) * | 2018-11-12 | 2019-03-29 | 南京邮电大学 | The image local Style Transfer method of production confrontation network based on dual confrontation |
CN109711529A (en) * | 2018-11-13 | 2019-05-03 | 中山大学 | A kind of cross-cutting federal learning model and method based on value iterative network |
-
2019
- 2019-05-27 CN CN201910445967.3A patent/CN110232652A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2017077121A1 (en) * | 2015-11-06 | 2017-05-11 | Thomson Licensing | Method for transfer of a style of a reference visual object to another visual object, and corresponding electronic device, computer readable program products and computer readable storage medium |
CN108647583A (en) * | 2018-04-19 | 2018-10-12 | 浙江大承机器人科技有限公司 | A kind of face recognition algorithms training method based on multiple target study |
CN109508654A (en) * | 2018-10-26 | 2019-03-22 | 中国地质大学(武汉) | Merge the human face analysis method and system of multitask and multiple dimensioned convolutional neural networks |
CN109544442A (en) * | 2018-11-12 | 2019-03-29 | 南京邮电大学 | The image local Style Transfer method of production confrontation network based on dual confrontation |
CN109711529A (en) * | 2018-11-13 | 2019-05-03 | 中山大学 | A kind of cross-cutting federal learning model and method based on value iterative network |
CN109523463A (en) * | 2018-11-20 | 2019-03-26 | 中山大学 | A kind of face aging method generating confrontation network based on condition |
Non-Patent Citations (3)
Title |
---|
JULIEN DESPOIS: "入门 | 极致的优化:智能手机是如何处理大型神经网络的", 《机器之心》 * |
何时夕: "android中的深度学习——快速风格迁移", 《终端研发部》 * |
裔隽: "《Python机器学习实战》", 31 January 2018 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110968671A (en) * | 2019-12-03 | 2020-04-07 | 北京声智科技有限公司 | Intent determination method and device based on Bert |
CN111986075A (en) * | 2020-08-12 | 2020-11-24 | 兰州交通大学 | Style migration method for target edge clarification |
CN112232594A (en) * | 2020-11-06 | 2021-01-15 | 中国人民解放军国防科技大学 | Multi-satellite emergency observation task planning method and device and computer equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110378468A (en) | A kind of neural network accelerator quantified based on structuring beta pruning and low bit | |
CN109949214A (en) | A kind of image Style Transfer method and system | |
CN105261066B (en) | A kind of three-dimensional geographic information system real-time rendering multithreading distribution and control method | |
CN109102064A (en) | A kind of high-precision neural network quantization compression method | |
CN109919319A (en) | Deeply learning method and equipment based on multiple history best Q networks | |
CN111784602A (en) | Method for generating countermeasure network for image restoration | |
CN109360097A (en) | Prediction of Stock Index method, apparatus, equipment and storage medium based on deep learning | |
CN108764466A (en) | Convolutional neural networks hardware based on field programmable gate array and its accelerated method | |
JP7287397B2 (en) | Information processing method, information processing apparatus, and information processing program | |
CN110222029A (en) | A kind of big data multidimensional analysis computational efficiency method for improving and system | |
CN109934336A (en) | Neural network dynamic based on optimum structure search accelerates platform designing method and neural network dynamic to accelerate platform | |
CN110232652A (en) | Image processing engine processing method, the image processing method for terminal, terminal | |
CN109117742A (en) | Gestures detection model treatment method, apparatus, equipment and storage medium | |
CN113010312B (en) | Super-parameter tuning method, device and storage medium | |
CN114154646A (en) | Efficiency optimization method for federal learning in mobile edge network | |
CN110222835A (en) | A kind of convolutional neural networks hardware system and operation method based on zero value detection | |
CN101833691A (en) | Realizing method of least square support vector machine serial structure based on EPGA (Filed Programmable Gate Array) | |
CN118096992A (en) | Shadow display method and device, storage medium and electronic device | |
JP2023076820A (en) | Image processing method, device, electronic device, and storage medium | |
CN107507255A (en) | Picture compression quality factor acquisition methods, system, equipment and storage medium | |
CN103995886A (en) | Multidimensional product design knowledge pushing frame and construction method | |
CN110120083A (en) | Water quality forecast is distributed drawing drawing method, device and electronic equipment | |
CN109002885A (en) | A kind of convolutional neural networks pond unit and pond calculation method | |
CN113850365A (en) | Method, device, equipment and storage medium for compressing and transplanting convolutional neural network | |
CN109255004A (en) | The online interaction formula method for visualizing of accuracy guarantee towards geographical spatial data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190913 |
|
RJ01 | Rejection of invention patent application after publication |