CN110232392A - Vision optimization method, optimization system, computer equipment and readable storage medium storing program for executing - Google Patents
Vision optimization method, optimization system, computer equipment and readable storage medium storing program for executing Download PDFInfo
- Publication number
- CN110232392A CN110232392A CN201810178131.7A CN201810178131A CN110232392A CN 110232392 A CN110232392 A CN 110232392A CN 201810178131 A CN201810178131 A CN 201810178131A CN 110232392 A CN110232392 A CN 110232392A
- Authority
- CN
- China
- Prior art keywords
- default
- loss function
- model
- perception loss
- weight
- 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
- 238000005457 optimization Methods 0.000 title claims abstract description 105
- 238000000034 method Methods 0.000 title claims abstract description 49
- 238000003860 storage Methods 0.000 title claims abstract description 10
- 230000006870 function Effects 0.000 claims abstract description 189
- 230000008447 perception Effects 0.000 claims abstract description 156
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 64
- 238000004590 computer program Methods 0.000 claims description 17
- 230000008859 change Effects 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims 2
- 230000008569 process Effects 0.000 description 15
- 238000011156 evaluation Methods 0.000 description 12
- 238000010586 diagram Methods 0.000 description 8
- 230000004069 differentiation Effects 0.000 description 8
- 238000004064 recycling Methods 0.000 description 6
- 230000008901 benefit Effects 0.000 description 4
- 230000011218 segmentation Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention proposes a kind of vision optimization method, vision optimization system, computer equipment and computer readable storage mediums.Wherein, vision optimization method includes: the perception loss function for obtaining the default convolutional neural networks and default problem of completing pre-training;Multiple models corresponding with the default number of plies are obtained according to the corresponding perception loss function training of the number of plies is preset in the perception loss function and default convolutional neural networks of default problem;Multiple models are assessed according to pre-set level;The weight of each model in multiple models is obtained according to assessment result and default index weight rule;The perception loss function of optimization is obtained according to the corresponding perception loss function of the weight of each model and the default number of plies.The present invention makes full use of and incorporates the good network of pre-training from low layer to high level, from part to whole semantic information, while having theoretical guarantee, realizes the similitude for accurately fine instructing two pictures.
Description
Technical field
The present invention relates to vision optimisation technique fields, in particular to a kind of empty vision optimization method, vision optimization system
System, computer equipment, computer readable storage medium.
Background technique
Vision optimization in, J.Johnson et al. propose perception loss function be a kind of two pictures of description whether language
Similar good method in justice.This method compares training set by choosing the good network (by taking vgg network as an example) of a pre-training
Middle generation removes rain figure piece and really goes rain figure piece in vgg (one kind of the good convolutional neural networks of existing pre-training) a certain layer of network
On mean square error.But influence of which layer due to choosing vgg on earth to result is very big.
SATOSHI IIZUKA et al. constructs local resolving device and global resolving device by fighting in production in network
Mode judges whether two pictures are similar from part and whole two angles.This optimization is compared to the net good using pre-training
Network can be more likely to over-fitting training set, to influence generalization ability.
YijunLi et al. is constructing semanteme just by emphasizing that two identical picture semantic segmentation results should be similar
Then change function.Although the good network of pre-training is utilized in this method, but identical semantic segmentation result is possible to correspond to
Different semantic information, it is possible that the similitude of two pictures can not be instructed fine.
Summary of the invention
The present invention is directed to solve at least one of the technical problems existing in the prior art or related technologies.
For this purpose, first aspect of the present invention is to propose a kind of vision optimization method.
The second aspect of the invention is to propose a kind of vision optimization system.
The third aspect of the invention is to propose a kind of computer equipment.
The fourth aspect of the invention is to propose a kind of computer readable storage medium.
In view of this, according to an aspect of the present invention, it proposes a kind of vision optimization methods, comprising: obtain and complete in advance
The perception loss function of trained default convolutional neural networks and default problem;According to the perception loss function of default problem and
The corresponding perception loss function training of the number of plies is preset in default convolutional neural networks obtains multiple models corresponding with the default number of plies;
Multiple models are assessed according to pre-set level;Each model in multiple models is obtained according to assessment result and default index weight rule
Weight;The perception loss function of optimization is obtained according to the corresponding perception loss function of the weight of each model and the default number of plies.
Vision optimization method provided by the invention, obtains the good convolutional neural networks of some pre-training and some is preset
The perception loss function of particular problem passes through the perception loss function of particular problem and the net good based on pre-training on training set
The perception loss function of the default number of plies of network trains multiple models, chooses some default differentiation evaluation index and assesses multiple moulds
Type, the result and index weight algorithmic rule for recycling assessment to obtain obtain the weight of each model, with the weight to default layer
The corresponding perception loss function of number carries out budget, obtains the perception loss function of particular problem and based on index weight algorithmic rule
Optimization perceive loss function, made full use of in the process for the perception loss function for obtaining final optimization and incorporate pre-training
Good network from part to whole semantic information, while having theoretical guarantee from low layer to high level, realizes accurately fine
Instruct the similitude of two pictures.
Above-mentioned vision optimization method according to the present invention, can also have following technical characteristic:
In the above-mentioned technical solutions, it is preferable that according to the perception loss function and default convolutional neural networks of default problem
In preset the corresponding perception loss function of the number of plies trained the step of obtaining multiple models corresponding with the default number of plies, specifically include:
The perception loss function of the default number of plies is obtained according to the first preset formula:
The perception loss function of multiple models is obtained according to the second preset formula:
Pi'=L+Pi
Wherein, PiTo preset the corresponding perception loss function of i-th layer of convolutional neural networks, Vi() is from default convolution mind
I-th layer of transformation for being input to default convolutional neural networks through network, training set raining set are (Xj,Yj), size (Vi
(Xj)) it is Vi(Xj) size, # () be training set size, Pi' be i-th of model perception loss function, L is pre- puts up a question
The perception loss function of topic.
In the technical scheme, the default number of plies in some preset trained network is chosen, can be the pre-training
A part of number of plies in good network, is also possible to convolutional layer whole in the good network of the pre-training, according in training set training
The number of plies of selection obtains the loss function of each model according to the first preset formula and the second preset formula, utilizes loss function
Train multiple models.
In any of the above-described technical solution, it is preferable that the step of assessing multiple models according to pre-set level specifically includes:
O is obtained according to third preset formulai(Xj) and YjSimilitude:
Wherein, Oi(Xj) be i-th of model output, E () is pre-set level, and training set raining set is
(Xj,Yj),For Oi(Xj) and YjSimilitude.
In the technical scheme, specific evaluation index E () is chosen to assess each model, is preset using third
Formula obtains output and the Y of each modeljSimilitude, realize the assessment of each model obtained to training.
In any of the above-described technical solution, it is preferable that obtain multiple moulds according to assessment result and default index weight rule
It in type the step of the weight of each model, specifically includes: obtaining the weight of each model according to the 4th preset formula:
Wherein, wiFor the weight of i-th of model, vgg is default convolutional neural networks.
In the technical scheme, the assessment result based on each model obtained using the 4th preset formulaUtilize
Four preset formulas (index weight algorithm) obtain the weight of each model, and default convolutional neural networks can be selected according to circumstances,
For example choose vgg (one kind of the good convolutional neural networks of existing pre-training).The weight of each model obtained is to obtain finally
Optimization perception loss function provides foundation.
In any of the above-described technical solution, it is preferable that damaged according to the weight of each model and the corresponding perception of the default number of plies
The step of function obtains the perception loss function of optimization is lost, is specifically included: being damaged according to the perception that the 5th preset formula obtains optimization
Lose function:
Wherein, L' is the perception loss function of optimization, and L is the perception loss function of default problem, and vgg is default convolution mind
Through network, wiFor the weight of i-th of model, vgg is default convolutional neural networks.
In the technical scheme, pass through the corresponding model of each layer of the weight of each model network good to pre-training
Loss function is weighted and averaged, with L'=∑i∈vggwi(L+Pi)=L+ ∑i∈vggwiPiAs the vision by index weight algorithm
The loss function of optimization.The loss function of vision optimization is according to the perception loss function of particular problem and based on index weight
What algorithmic rule obtained, it is made full use of in the process for the perception loss function for obtaining final optimization and to incorporate pre-training good
Network from part to whole semantic information, while having theoretical guarantee, realize and accurately fine instruct from low layer to high level
The similitude of two pictures.
According to the second aspect of the invention, a kind of vision optimization system is proposed, comprising: acquiring unit, for obtaining
Complete the default convolutional neural networks of pre-training and the perception loss function of default problem;Training unit, for according to default
The number of plies corresponding perception loss function training is preset in the perception loss function and default convolutional neural networks of problem to obtain and in advance
If the corresponding multiple models of the number of plies;Assessment unit, for assessing multiple models according to pre-set level;Weight unit is used for basis
Assessment result and default index weight rule obtain the weight of each model in multiple models;Optimize unit, for according to each
The weight of model and the corresponding perception loss function of the default number of plies obtain the perception loss function of optimization.
Vision optimization system provided by the invention, obtains the good convolutional neural networks of some pre-training and some is preset
The perception loss function of particular problem passes through the perception loss function of particular problem and the net good based on pre-training on training set
The perception loss function of the default number of plies of network trains multiple models, chooses some default differentiation evaluation index and assesses multiple moulds
Type, the result and index weight algorithmic rule for recycling assessment to obtain obtain the weight of each model, with the weight to default layer
The corresponding perception loss function of number carries out budget, obtains the perception loss function of particular problem and based on index weight algorithmic rule
Optimization perceive loss function, made full use of in the process for the perception loss function for obtaining final optimization and incorporate pre-training
Good network from part to whole semantic information, while having theoretical guarantee from low layer to high level, realizes accurately fine
Instruct the similitude of two pictures.
Above-mentioned vision optimization system according to the present invention, can also have following technical characteristic:
In the above-mentioned technical solutions, it is preferable that training unit specifically includes: the first computing unit, for pre- according to first
If formula obtains the perception loss function of the default number of plies:
Second computing unit, for obtaining the perception loss function of multiple models according to the second preset formula:
Pi'=L+Pi
Wherein, PiTo preset the corresponding perception loss function of i-th layer of convolutional neural networks, Vi() is from default convolution mind
I-th layer of transformation for being input to default convolutional neural networks through network, training set raining set are (Xj,Yj), size (Vi
(Xj)) it is Vi(Xj) size, # () be training set size, Pi' be i-th of model perception loss function, L is pre- puts up a question
The perception loss function of topic.
In the technical scheme, the default number of plies in some preset trained network is chosen, can be the pre-training
A part of number of plies in good network, is also possible to convolutional layer whole in the good network of the pre-training, according in training set training
The number of plies of selection obtains the loss function of each model according to the first preset formula and the second preset formula, utilizes loss function
Train multiple models.
In any of the above-described technical solution, it is preferable that assessment unit specifically includes: third computing unit, for according to the
Three preset formulas obtain Oi(Xj) and YjSimilitude:
Wherein, Oi(Xj) be i-th of model output, E () is pre-set level, and training set raining set is
(Xj,Yj),For Oi(Xj) and YjSimilitude.
In the technical scheme, specific evaluation index E () is chosen to assess each model, is preset using third
Formula obtains output and the Y of each modeljSimilitude, realize the assessment of each model obtained to training.
In any of the above-described technical solution, it is preferable that weight unit specifically includes: the 4th computing unit, for according to the
Four preset formulas obtain the weight of each model:
Wherein, wiFor the weight of i-th of model, vgg is default convolutional neural networks.
In the technical scheme, the assessment result based on each model obtained using the 4th preset formulaUtilize
Four preset formulas (index weight algorithm) obtain the weight of each model, and default convolutional neural networks can be selected according to circumstances,
For example choose vgg (one kind of the good convolutional neural networks of existing pre-training).The weight of each model obtained is to obtain finally
Optimization perception loss function provides foundation.
In any of the above-described technical solution, it is preferable that optimization unit specifically includes: the 5th computing unit, for according to the
Five preset formulas obtain the perception loss function of optimization:
Wherein, L' is the perception loss function of optimization, and L is the perception loss function of default problem, and vgg is default convolution mind
Through network, wiFor the weight of i-th of model, vgg is default convolutional neural networks.
In the technical scheme, pass through the corresponding model of each layer of the weight of each model network good to pre-training
Loss function is weighted and averaged, with L'=∑i∈vggwi(L+Pi)=L+ ∑i∈vggwiPiAs the vision by index weight algorithm
The loss function of optimization.The loss function of vision optimization is according to the perception loss function of particular problem and based on index weight
What algorithmic rule obtained, it is made full use of in the process for the perception loss function for obtaining final optimization and to incorporate pre-training good
Network from part to whole semantic information, while having theoretical guarantee, realize and accurately fine instruct from low layer to high level
The similitude of two pictures.
According to the third aspect of the present invention, the present invention provides a kind of computer equipment, including memory, processor and
The computer program that can be run on a memory and on a processor is stored, processor is realized when executing computer program as above-mentioned
The step of vision optimization method of any one.
A kind of computer equipment provided by the invention, processor are realized when executing computer program: obtaining some pre-training
The perception loss function of good convolutional neural networks and some preset particular problem, passes through particular problem on training set
Perception loss function and the perception loss function of the default number of plies based on the good network of pre-training train multiple models, choose certain
A default differentiation evaluation index assesses multiple models, and the result and index weight algorithmic rule for recycling assessment to obtain obtain each
The weight of model carries out budget with weight perception loss function corresponding to the default number of plies, obtains the perception damage of particular problem
It loses function and perceives loss function with the optimization based on index weight algorithmic rule, in the perception loss function for obtaining final optimization
Process make full use of and incorporate the good network of pre-training from low layer to high level, from part to whole semantic information, simultaneously
It is ensured with theory, realizes the similitude for accurately fine instructing two pictures.
According to the fourth aspect of the present invention, it the present invention provides a kind of computer readable storage medium, is stored thereon with
Computer program, when computer program is executed by processor the step of the realization such as vision optimization method of any of the above-described.
A kind of computer readable storage medium provided by the invention is stored thereon with computer program, computer program quilt
Processor is realized when executing: obtaining the good convolutional neural networks of some pre-training and the perception damage of some preset particular problem
Function is lost, passes through the sense of the perception loss function and the default number of plies based on the good network of pre-training of particular problem on training set
Know that loss function trains multiple models, chooses some default differentiation evaluation index and assess multiple models, assessment is recycled to obtain
Result and index weight algorithmic rule obtain the weight of each model, lose letter with corresponding to the default number of plies perception of the weight
Number carries out budget, the perception loss function and the optimization perception loss letter based on index weight algorithmic rule for obtaining particular problem
Number, the process for the perception loss function for obtaining final optimization make full use of and incorporate the good network of pre-training from low layer to
High level from part to whole semantic information, while having theoretical guarantee, realizes the phase for accurately fine instructing two pictures
Like property.
Additional aspect and advantage of the invention will become obviously in following description section, or practice through the invention
Recognize.
Detailed description of the invention
Above-mentioned and/or additional aspect of the invention and advantage will become from the description of the embodiment in conjunction with the following figures
Obviously and it is readily appreciated that, in which:
Fig. 1 shows the flow diagram of the vision optimization method of one embodiment of the present of invention;
Fig. 2 shows the vision optimization methods of a specific embodiment of the invention to apply the process signal in particular problem
Figure;
Fig. 3 shows the schematic block diagram of the vision optimization system of one embodiment of the present of invention;
Fig. 4 shows the schematic block diagram of the vision optimization system of another embodiment of the invention;
Fig. 5 shows the schematic block diagram of the computer equipment of one embodiment of the present of invention.
Specific embodiment
It is with reference to the accompanying drawing and specific real in order to be more clearly understood that aforementioned aspect of the present invention, feature and advantage
Applying mode, the present invention is further described in detail.It should be noted that in the absence of conflict, the implementation of the application
Feature in example and embodiment can be combined with each other.
In the following description, numerous specific details are set forth in order to facilitate a full understanding of the present invention, still, the present invention may be used also
To be implemented using other than the one described here other modes, therefore, protection scope of the present invention is not limited to following public affairs
The limitation for the specific embodiment opened.
The embodiment of first aspect present invention proposes that a kind of vision optimization method, Fig. 1 show an implementation of the invention
The flow diagram of the vision optimization method of example:
Step 102, it obtains and completes the default convolutional neural networks of pre-training and the perception loss function of default problem;
Step 104, corresponding according to the number of plies is preset in the perception loss function of default problem and default convolutional neural networks
Perception loss function training obtains multiple models corresponding with the default number of plies;
Step 106, multiple models are assessed according to pre-set level;
Step 108, the weight of each model in multiple models is obtained according to assessment result and default index weight rule;
Step 110, the perception of optimization is obtained according to the corresponding perception loss function of the weight of each model and the default number of plies
Loss function.
Vision optimization method provided by the invention, obtains the good convolutional neural networks of some pre-training and some is preset
The perception loss function of particular problem passes through the perception loss function of particular problem and the net good based on pre-training on training set
The perception loss function of the default number of plies of network trains multiple models, chooses some default differentiation evaluation index and assesses multiple moulds
Type, the result and index weight algorithmic rule for recycling assessment to obtain obtain the weight of each model, with the weight to default layer
The corresponding perception loss function of number carries out budget, obtains the perception loss function of particular problem and based on index weight algorithmic rule
Optimization perceive loss function, made full use of in the process for the perception loss function for obtaining final optimization and incorporate pre-training
Good network from part to whole semantic information, while having theoretical guarantee from low layer to high level, realizes accurately fine
Instruct the similitude of two pictures.
In the above embodiment, it is preferable that according in the perception loss function and default convolutional neural networks of default problem
The step of corresponding perception loss function training of the number of plies obtains multiple models corresponding with the default number of plies is preset, is specifically included: root
The perception loss function of the default number of plies is obtained according to the first preset formula:
The perception loss function of multiple models is obtained according to the second preset formula:
Pi'=L+Pi
Wherein, PiTo preset the corresponding perception loss function of i-th layer of convolutional neural networks, Vi() is from default convolution mind
I-th layer of transformation for being input to default convolutional neural networks through network, training set raining set are (Xj,Yj), size (Vi
(Xj)) it is Vi(Xj) size, # () be training set size, Pi' be i-th of model perception loss function, L is pre- puts up a question
The perception loss function of topic.
In this embodiment, the default number of plies in some preset trained network is chosen, it is good to can be the pre-training
Network in a part of number of plies, be also possible to convolutional layer whole in the good network of the pre-training, default convolutional neural networks can
According to circumstances to select, for example choose vgg (one kind of the good convolutional neural networks of existing pre-training).According in training set training
The number of plies of selection obtains the loss function of each model according to the first preset formula and the second preset formula, utilizes loss function
Train multiple models.
In any of the above-described embodiment, it is preferable that the step of assessing multiple models according to pre-set level specifically includes: root
O is obtained according to third preset formulai(Xj) and YjSimilitude:
Wherein, Oi(Xj) be i-th of model output, E () is pre-set level, and training set raining set is
(Xj,Yj),For Oi(Xj) and YjSimilitude.
In this embodiment, specific evaluation index E () is chosen to assess each model, utilizes the default public affairs of third
Formula obtains output and the Y of each modeljSimilitude, realize the assessment of each model obtained to training.
In any of the above-described embodiment, it is preferable that obtain multiple models according to assessment result and default index weight rule
In each model weight the step of, specifically include: obtaining the weight of each model according to the 4th preset formula:
Wherein, wiFor the weight of i-th of model, vgg is default convolutional neural networks.
In this embodiment, the assessment result based on each model obtained using the 4th preset formulaUtilize the 4th
Preset formula (index weight algorithm) obtains the weight of each model, and default convolutional neural networks can be selected according to circumstances, than
Such as choose vgg (one kind of the good convolutional neural networks of existing pre-training).The weight of each model obtained is final excellent to obtain
Allelopathic knows that loss function provides foundation.
In any of the above-described embodiment, it is preferable that lost according to the weight of each model and the corresponding perception of the default number of plies
Function obtains the step of perception loss function of optimization, specifically includes: being lost according to the perception that the 5th preset formula obtains optimization
Function:
Wherein, L' is the perception loss function of optimization, and L is the perception loss function of default problem, and vgg is default convolution mind
Through network, wiFor the weight of i-th of model, vgg is default convolutional neural networks.
In this embodiment, pass through the damage of the corresponding model of each layer of the weight of each model network good to pre-training
It loses function to be weighted and averaged, with L'=∑i∈vggwi(L+Pi)=L+ ∑i∈vggwiPiAs excellent by the vision of index weight algorithm
The loss function of change.The loss function of vision optimization is to be reruned according to the perception loss function of particular problem with based on exponential weight
Method rule obtains, and makes full use of in the process for the perception loss function for obtaining final optimization and incorporates the good net of pre-training
Network from part to whole semantic information, while having theoretical guarantee from low layer to high level, realizes and accurately fine instructs two
The similitude of picture.
Fig. 2 shows the vision optimization methods of a specific embodiment of the invention to apply the process signal in particular problem
Figure.In this particular problem, L is mean square error loss function and confrontation loss function.As shown in Fig. 2, by vision optimization side
Process of the method for particular problem first passes through L+P1, L+P2Train two model M odel1, Model2, then use index weight
Algorithm obtains weight w1,w2, finally obtain the vision optimization loss function L+w based on index weight1P1+w2P2.The specific implementation
The major advantage of example is to make full use of and incorporate from low layer to high level from part to whole semantic information, while having one
Fixed theoretical guarantee.
The embodiment of second aspect of the present invention, proposes a kind of vision optimization system 300, and Fig. 3 shows of the invention one
The schematic block diagram of the vision optimization system 300 of embodiment.As shown in figure 3, vision optimization system 300 include: acquiring unit 302,
Training unit 304, assessment unit 306, weight unit 308 and optimization unit 310.Wherein, acquiring unit 302 is completed for obtaining
The default convolutional neural networks of pre-training and the perception loss function of default problem;Training unit 304 is used to be put up a question according to pre-
The corresponding perception loss function training of the number of plies is preset in the perception loss function and default convolutional neural networks of topic to obtain and preset
The corresponding multiple models of the number of plies;Assessment unit 306 is used to assess multiple models according to pre-set level;Weight unit 308 is used for root
The weight of each model in multiple models is obtained according to assessment result and default index weight rule;Optimize unit 310 and is used for basis
The weight of each model and the corresponding perception loss function of the default number of plies obtain the perception loss function of optimization.
In vision optimization system provided by the invention, obtains the good convolutional neural networks of some pre-training and some is default
Particular problem perception loss function, it is good by the perception loss function of particular problem and based on pre-training on training set
The perception loss function of the default number of plies of network trains multiple models, chooses some default differentiation evaluation index and assesses multiple moulds
Type, the result and index weight algorithmic rule for recycling assessment to obtain obtain the weight of each model, with the weight to default layer
The corresponding perception loss function of number carries out budget, obtains the perception loss function of particular problem and based on index weight algorithmic rule
Optimization perceive loss function, made full use of in the process for the perception loss function for obtaining final optimization and incorporate pre-training
Good network from part to whole semantic information, while having theoretical guarantee from low layer to high level, realizes accurately fine
Instruct the similitude of two pictures.
Fig. 4 shows the schematic block diagram of the vision optimization system 400 of another embodiment of the invention.Wherein, vision is excellent
Change system 400 includes: acquiring unit 402, training unit 404, assessment unit 406, weight unit 408 and optimization unit 410.Its
In, acquiring unit 402 is used to obtain the perception loss function of the default convolutional neural networks and default problem of completing pre-training;
Training unit 404 is used to preset the corresponding sense of the number of plies in perception loss function and default convolutional neural networks according to the problem of presetting
Know that loss function training obtains multiple models corresponding with the default number of plies;Assessment unit 406 is used to be assessed according to pre-set level more
A model;Weight unit 408 is used to obtain each model in multiple models according to assessment result and default index weight rule
Weight;Optimize unit 410 to be used to obtain optimization according to the corresponding perception loss function of the weight of each model and the default number of plies
Perceive loss function.
In the above embodiment, it is preferable that training unit 404 specifically includes: the first computing unit 442 and second calculates list
Member 444.First computing unit 442 is used to obtain the perception loss function of the default number of plies according to the first preset formula:
Second computing unit 444 is used to obtain the perception loss function of multiple models according to the second preset formula:
Pi'=L+Pi
Wherein, PiTo preset the corresponding perception loss function of i-th layer of convolutional neural networks, Vi() is from default convolution mind
I-th layer of transformation for being input to default convolutional neural networks through network, training set raining set are (Xj,Yj), size (Vi
(Xj)) it is Vi(Xj) size, # () be training set size, Pi' be i-th of model perception loss function, L is pre- puts up a question
The perception loss function of topic.
In this embodiment, the default number of plies in some preset trained network is chosen, it is good to can be the pre-training
Network in a part of number of plies, be also possible to convolutional layer whole in the good network of the pre-training, according to training set training select
The number of plies taken is obtained the loss function of each model according to the first preset formula and the second preset formula, is instructed using loss function
Practise multiple models.
In any of the above-described embodiment, it is preferable that assessment unit 406 specifically includes third computing unit 462.Third calculates
Unit 462.For obtaining O according to third preset formulai(Xj) and YjSimilitude:
Wherein, Oi(Xj) be i-th of model output, E () is pre-set level, and training set raining set is
(Xj,Yj),For Oi(Xj) and YjSimilitude.
In this embodiment, specific evaluation index E () is chosen to assess each model, utilizes the default public affairs of third
Formula obtains output and the Y of each modeljSimilitude, realize the assessment of each model obtained to training.
In any of the above-described embodiment, it is preferable that weight unit 408 specifically includes the 4th computing unit 482.4th calculates
Unit 482 is used to obtain the weight of each model according to the 4th preset formula:
Wherein, wiFor the weight of i-th of model, vgg is default convolutional neural networks.
In this embodiment, the assessment result based on each model obtained using the 4th preset formulaUtilize the 4th
Preset formula (index weight algorithm) obtains the weight of each model, and default convolutional neural networks can be selected according to circumstances, than
Such as choose vgg (one kind of the good convolutional neural networks of existing pre-training).The weight of each model obtained is final excellent to obtain
Allelopathic knows that loss function provides foundation.
In any of the above-described embodiment, it is preferable that optimization unit 410 specifically includes the 5th computing unit 412.5th calculates
Unit 412 is used to obtain the perception loss function of optimization according to the 5th preset formula:
Wherein, L' is the perception loss function of optimization, and L is the perception loss function of default problem, and vgg is default convolution mind
Through network, wiFor the weight of i-th of model, vgg is default convolutional neural networks.
In this embodiment, pass through the damage of the corresponding model of each layer of the weight of each model network good to pre-training
It loses function to be weighted and averaged, with L'=∑i∈vggwi(L+Pi)=L+ ∑i∈vggwiPiAs excellent by the vision of index weight algorithm
The loss function of change.The loss function of vision optimization is to be reruned according to the perception loss function of particular problem with based on exponential weight
Method rule obtains, and makes full use of in the process for the perception loss function for obtaining final optimization and incorporates the good net of pre-training
Network from part to whole semantic information, while having theoretical guarantee from low layer to high level, realizes and accurately fine instructs two
The similitude of picture.
The embodiment of third aspect present invention, proposes a kind of computer equipment, and Fig. 5 shows one embodiment of the present of invention
Computer equipment 500 schematic block diagram.Wherein, which includes:
Memory 502, processor 504 and it is stored in the computer journey that can be run on memory 502 and on processor 504
The step of sequence, processor 504 realizes the vision optimization method such as any of the above-described when executing computer program.
A kind of computer equipment 500 provided by the invention, processor 504 are realized when executing computer program: obtaining some
The perception loss function of the good convolutional neural networks of pre-training and some preset particular problem, by specific on training set
The perception loss function of problem and the perception loss function of the default number of plies based on the good network of pre-training train multiple models,
It chooses some default differentiation evaluation index and assesses multiple models, the result and index weight algorithmic rule for recycling assessment to obtain obtain
To the weight of each model, budget is carried out with weight perception loss function corresponding to the default number of plies, obtains particular problem
It perceives loss function and perceives loss function with the optimization based on index weight algorithmic rule, in the perception damage for obtaining final optimization
The process for losing function makes full use of and incorporates the good network of pre-training from low layer to high level, from part to whole semantic letter
Breath, while there is theoretical guarantee, realize the similitude for accurately fine instructing two pictures.
The embodiment of fourth aspect present invention provides a kind of computer readable storage medium, is stored thereon with computer
Program, when computer program is executed by processor the step of the realization such as vision optimization method of any of the above-described.
A kind of computer readable storage medium provided by the invention is stored thereon with computer program, computer program quilt
Processor is realized when executing: obtaining the good convolutional neural networks of some pre-training and the perception damage of some preset particular problem
Function is lost, passes through the sense of the perception loss function and the default number of plies based on the good network of pre-training of particular problem on training set
Know that loss function trains multiple models, chooses some default differentiation evaluation index and assess multiple models, assessment is recycled to obtain
Result and index weight algorithmic rule obtain the weight of each model, lose letter with corresponding to the default number of plies perception of the weight
Number carries out budget, the perception loss function and the optimization perception loss letter based on index weight algorithmic rule for obtaining particular problem
Number, the process for the perception loss function for obtaining final optimization make full use of and incorporate the good network of pre-training from low layer to
High level from part to whole semantic information, while having theoretical guarantee, realizes the phase for accurately fine instructing two pictures
Like property.
In the description of this specification, the description of term " one embodiment ", " some embodiments ", " specific embodiment " etc.
Mean that particular features, structures, materials, or characteristics described in conjunction with this embodiment or example are contained at least one reality of the invention
It applies in example or example.In the present specification, schematic expression of the above terms are not necessarily referring to identical embodiment or reality
Example.Moreover, description particular features, structures, materials, or characteristics can in any one or more of the embodiments or examples with
Suitable mode combines.
The foregoing is only a preferred embodiment of the present invention, is not intended to restrict the invention, for the skill of this field
For art personnel, the invention may be variously modified and varied.All within the spirits and principles of the present invention, made any to repair
Change, equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.
Claims (12)
1. a kind of vision optimization method characterized by comprising
Obtain the perception loss function of the default convolutional neural networks and default problem of completing pre-training;
According to the corresponding perception of the number of plies default in the perception loss function of the default problem and the default convolutional neural networks
Loss function training obtains multiple models corresponding with the default number of plies;
The multiple model is assessed according to pre-set level;
The weight of each model in the multiple model is obtained according to assessment result and default index weight rule;
The perception loss of optimization is obtained according to the corresponding perception loss function of the weight of each model and the default number of plies
Function.
2. vision optimization method according to claim 1, which is characterized in that described to be damaged according to the perception of the default problem
The corresponding perception loss function training of the number of plies is preset in mistake function and the default convolutional neural networks to obtain and the default layer
It the step of number corresponding multiple models, specifically includes:
The perception loss function of the default number of plies is obtained according to the first preset formula:
The perception loss function of the multiple model is obtained according to the second preset formula:
Pi'=L+Pi
Wherein, PiFor the corresponding perception loss function of i-th layer of the default convolutional neural networks, Vi() is from the default volume
I-th layer of the transformation for being input to the default convolutional neural networks of product neural network, training set raining set are (Xj,
Yj), size (Vi(Xj)) it is Vi(Xj) size, # () be the training set size, Pi' lost for the perception of i-th of model
Function, L are the perception loss function of the default problem.
3. vision optimization method according to claim 1, which is characterized in that described the multiple according to pre-set level assessment
The step of model, specifically includes:
O is obtained according to third preset formulai(Xj) and YjSimilitude:
Wherein, Oi(Xj) be i-th of model output, E () is the pre-set level, and training set raining set is
(Xj,Yj),For Oi(Xj) and YjSimilitude.
4. vision optimization method according to claim 3, which is characterized in that described according to assessment result and default exponential weight
Weight-normality then obtains the step of weight of each model in the multiple model, specifically includes:
The weight of each model is obtained according to the 4th preset formula:
Wherein, wiFor the weight of i-th of model, vgg is the default convolutional neural networks.
5. vision optimization method according to any one of claim 1 to 4, which is characterized in that described according to described each
The step of weight of model and the corresponding perception loss function of the default number of plies obtain the perception loss function of optimization, it is specific to wrap
It includes:
The perception loss function of the optimization is obtained according to the 5th preset formula:
Wherein, L' is the perception loss function of the optimization, and L is the perception loss function of the default problem, and vgg is described pre-
If convolutional neural networks, wiFor the weight of i-th of model, vgg is the default convolutional neural networks.
6. a kind of vision optimization system characterized by comprising
Acquiring unit completes the default convolutional neural networks of pre-training and the perception loss function of default problem for obtaining;
Training unit, for presetting layer in the perception loss function and the default convolutional neural networks according to the default problem
The corresponding perception loss functions of number are trained to obtain multiple models corresponding with the default number of plies;
Assessment unit, for assessing the multiple model according to pre-set level;
Weight unit, for obtaining the power of each model in the multiple model according to assessment result and default index weight rule
Weight;
Optimize unit, it is excellent for being obtained according to the weight of each model and the corresponding perception loss function of the default number of plies
The perception loss function of change.
7. vision optimization system according to claim 6, which is characterized in that the training unit specifically includes:
First computing unit, for obtaining the perception loss function of the default number of plies according to the first preset formula:
Second computing unit, for obtaining the perception loss function of the multiple model according to the second preset formula:
Pi'=L+Pi
Wherein, PiFor the corresponding perception loss function of i-th layer of the default convolutional neural networks, Vi() is from the default volume
I-th layer of the transformation for being input to the default convolutional neural networks of product neural network, training set raining set are (Xj,
Yj), size (Vi(Xj)) it is Vi(Xj) size, # () be the training set size, Pi' lost for the perception of i-th of model
Function, L are the perception loss function of the default problem.
8. vision optimization system according to claim 6, which is characterized in that the assessment unit specifically includes:
Third computing unit, for obtaining O according to third preset formulai(Xj) and YjSimilitude:
Wherein, Oi(Xj) be i-th of model output, E () is the pre-set level, and training set raining set is
(Xj,Yj),For Oi(Xj) and YjSimilitude.
9. vision optimization system according to claim 8, which is characterized in that the weight unit specifically includes:
4th computing unit, for obtaining the weight of each model according to the 4th preset formula:
Wherein, wiFor the weight of i-th of model, vgg is the default convolutional neural networks.
10. vision optimization system according to any one of claims 6 to 9, which is characterized in that the optimization unit is specific
Include:
5th computing unit, for obtaining the perception loss function of the optimization according to the 5th preset formula:
Wherein, L' is the perception loss function of the optimization, and L is the perception loss function of the default problem, and vgg is described pre-
If convolutional neural networks, wiFor the weight of i-th of model, vgg is the default convolutional neural networks.
11. a kind of computer equipment, including memory, processor and it is stored on the memory and can be on the processor
The computer program of operation, which is characterized in that the processor realizes such as claim 1 to 5 when executing the computer program
Any one of described in vision optimization method the step of.
12. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
It is realized when being executed by processor as described in any one of claims 1 to 5 the step of vision optimization method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810178131.7A CN110232392B (en) | 2018-03-05 | 2018-03-05 | Visual optimization method, optimization system, computer device and readable storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810178131.7A CN110232392B (en) | 2018-03-05 | 2018-03-05 | Visual optimization method, optimization system, computer device and readable storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110232392A true CN110232392A (en) | 2019-09-13 |
CN110232392B CN110232392B (en) | 2021-08-17 |
Family
ID=67861973
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810178131.7A Expired - Fee Related CN110232392B (en) | 2018-03-05 | 2018-03-05 | Visual optimization method, optimization system, computer device and readable storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110232392B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106845471A (en) * | 2017-02-20 | 2017-06-13 | 深圳市唯特视科技有限公司 | A kind of vision significance Forecasting Methodology based on generation confrontation network |
CN107590786A (en) * | 2017-09-08 | 2018-01-16 | 深圳市唯特视科技有限公司 | A kind of image enchancing method based on confrontation learning network |
CN107705242A (en) * | 2017-07-20 | 2018-02-16 | 广东工业大学 | A kind of image stylization moving method of combination deep learning and depth perception |
US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
-
2018
- 2018-03-05 CN CN201810178131.7A patent/CN110232392B/en not_active Expired - Fee Related
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
CN106845471A (en) * | 2017-02-20 | 2017-06-13 | 深圳市唯特视科技有限公司 | A kind of vision significance Forecasting Methodology based on generation confrontation network |
CN107705242A (en) * | 2017-07-20 | 2018-02-16 | 广东工业大学 | A kind of image stylization moving method of combination deep learning and depth perception |
CN107590786A (en) * | 2017-09-08 | 2018-01-16 | 深圳市唯特视科技有限公司 | A kind of image enchancing method based on confrontation learning network |
Non-Patent Citations (3)
Title |
---|
CHRISTIAN LEDIG ET.AL: "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
JUSTIN JOHNSON ET.AL: "Perceptual Losses for Real-Time Style Transfer and Super-Resolution", 《ARXIV:1603.08155V1 [CS.CV]》 * |
LEON A. GATYS ET.AL: "A Neural Algorithm of Artistic Style", 《ARXIV:1508.06576V2 [CS.CV]》 * |
Also Published As
Publication number | Publication date |
---|---|
CN110232392B (en) | 2021-08-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106156781B (en) | Sort convolutional neural networks construction method and its image processing method and device | |
Olson et al. | Counterfactual state explanations for reinforcement learning agents via generative deep learning | |
WO2018054330A1 (en) | Method and device for data processing and storage medium | |
Stewart et al. | A scenario‐based framework for multicriteria decision analysis in water resources planning | |
Sivasakthi | Classification and prediction based data mining algorithms to predict students' introductory programming performance | |
CN108846314A (en) | A kind of food materials identification system and food materials discrimination method based on deep learning | |
CN115064020B (en) | Intelligent teaching method, system and storage medium based on digital twin technology | |
CN110825884A (en) | Embedded representation processing method and device based on artificial intelligence and electronic equipment | |
Eagle et al. | Interaction Networks: Generating High Level Hints Based on Network Community Clustering. | |
CN109597998B (en) | Visual feature and semantic representation joint embedded image feature construction method | |
CN108090472A (en) | Pedestrian based on multichannel uniformity feature recognition methods and its system again | |
CN109712165A (en) | A kind of similar foreground picture image set dividing method based on convolutional neural networks | |
CN109389151A (en) | A kind of knowledge mapping treating method and apparatus indicating model based on semi-supervised insertion | |
CN109145974A (en) | One kind being based on the matched multi-level image Feature fusion of picture and text | |
Koliba et al. | Complexity theory and systems analysis | |
CN108416535A (en) | The method of patent valve estimating based on deep learning | |
KR20190125029A (en) | Methods and apparatuses for generating text to video based on time series adversarial neural network | |
Runfola et al. | A multi-criteria geographic information systems approach for the measurement of vulnerability to climate change | |
CN111488979A (en) | Method and apparatus for continuously learning neural network for analyzing input data on device | |
Figueiredo et al. | struc2vec: Learning node representations from structural identity | |
Grivokostopoulou et al. | An ontology-based approach for user modelling and personalization in e-learning systems | |
Liu et al. | Self-supervised depth estimation to regularise semantic segmentation in knee arthroscopy | |
Leyshon et al. | Landscape and climate change | |
Jahani et al. | Evaluating the aesthetic quality of the landscape in the environment: A Review of the Concepts and Scientific Developments in the World | |
CN105915883B (en) | Blind reference stereo image quality evaluation method with binocular fusion is learnt based on the limit |
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 | ||
TR01 | Transfer of patent right | ||
TR01 | Transfer of patent right |
Effective date of registration: 20230327 Address after: 100871 No. 5, the Summer Palace Road, Beijing, Haidian District Patentee after: Peking University Address before: 100871 No. 5, the Summer Palace Road, Beijing, Haidian District Patentee before: Peking University Patentee before: PEKING UNIVERSITY FOUNDER GROUP Co.,Ltd. Patentee before: BEIJING FOUNDER ELECTRONICS Co.,Ltd. |
|
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210817 |