CN110135498A - A kind of image-recognizing method based on depth Evolutionary Neural Network - Google Patents
A kind of image-recognizing method based on depth Evolutionary Neural Network Download PDFInfo
- Publication number
- CN110135498A CN110135498A CN201910410375.8A CN201910410375A CN110135498A CN 110135498 A CN110135498 A CN 110135498A CN 201910410375 A CN201910410375 A CN 201910410375A CN 110135498 A CN110135498 A CN 110135498A
- Authority
- CN
- China
- Prior art keywords
- individual
- chromosome
- neural network
- population
- algorithm
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
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 discloses a kind of image-recognizing methods based on depth Evolutionary Neural Network, belong to the fields such as deep learning, evolution algorithm.The present invention will initialize multiple minimum neural network structures as initial population when handling problem of image recognition first, and carry out the judge of environment fitness to all individuals in population by environment fitness function;Secondly individual is selected using parent selection algorithm;Then mutation operation is carried out to selected individual, to generate the population of a new generation;It finally repeats the above steps repeatedly, until mostly the highest individual configurations of environment fitness no longer change between generation, it can be by this individual as final disaggregated model structure.The present invention combines evolution algorithm with neural network, produce depth Evolutionary Neural Network algorithm, and apply it in field of image recognition, solve that Artificial Neural Network Structures are difficult to select, hyper parameter is difficult to the problems such as determining, and improves the efficiency of model buildings.
Description
Technical field
The present invention relates to deep learnings and evolution algorithm technical field, and in particular to one kind is based on depth Evolutionary Neural Network
Image-recognizing method.
Background technique
In recent years, neural network is as field most popular in machine learning field and branch's hair with strongest influence power
Exhibition is very swift and violent, learns studied computer vision, natural language processing, the intelligence that intensified learning is studied in deep learning
Good achievement is all achieved in body.It is simulation biological brain neuronal structure and mechanism, builds artificial nerve network model.
Deep learning is to establish multiple hidden layers between the input layer and output layer of network, complicated neural network is constituted, by right
Sample data optimizes and adjusts the parameter between network, classification problem after solution returns without label or the study for having label.
But still there are many shortcomings for neural network in practical applications.First, the model buildings of neural network are difficult,
An intact nervous network is built in practical application, is first had to the network structure for selecting to be suitble to problem, is selected again after choosing structure
The network number of plies, the number of nodes or other parameters of each layer network are selected, so artificial experimental method or trial and error procedure would generally be used, but people
Work experiment needs neural network abundant and builds experience, and trial and error procedure is not only blindly but also time-consuming.Second, network is in training
It needs to adjust there are also many parameters, as learning rate, weights initialisation method can allow whole if these parameters cannot accurately be arranged
There is the problems such as training time long, poor accuracy or network are not restrained in a network training, and the adjustment to these parameters
Equally take time and effort.Third, when encounter that training data is sparse or data distribution not quietly the case where when, as intensified learning is asked
Topic, the problem that neural network can not be restrained or even can not normally be trained.
It is difficult to build various with hyper parameter in network for the network structure occurred in neural network and is difficult to choose etc. and asks
Topic, numerous domestic and foreign scholars propose many solutions.Such as tall and big open proposes a kind of " pyramid " to determine neural network
The number of interior joint.Sun etc., which is proposed, a kind of to be exported using binary system by error in classification and neural network propagated forward is combined to go
Choose the method for hiding node layer.Zhang Yunong proposes weight with number of nodes by that can determine weight in conjunction with training error
With the model of number of nodes.Yang makes network by the way that random hidden layer node number is adaptively trained and combined to network
Possess better Nonlinear Learning ability.With the new development of Neural Network Optimization, more domestic and foreign scholars begin one's study by
Evolution algorithm and neural network combine the evolution nerve network algorithm of generation.
Evolution algorithm is a kind of Stochastic Optimization Algorithms for simulating biological evolution, has good ability of searching optimum and is not necessarily to
The advantageous property of the gradient information energy autonomous learning of error function.Evolutionary Neural Network is looked for by evolution algorithm, is pair
The intelligence of neural network builds good exploration.On the one hand, Evolutionary Neural Network, which can help to reduce, manually builds network model
Spent energy, and more novel structures in neural network can also be explored;On the other hand, Evolutionary Neural Network can be helped
It helps and manually adjusts the numerous hyper parameters of neural network, making network just has better learning ability;Finally, Evolutionary Neural Network can
To allow neural network preferably to train under the unstable environment of gradient information, so that improving neural network solves various problems
Ability.
Summary of the invention
For the deficiency in the presence of the prior art, the present invention provides a kind of images based on depth Evolutionary Neural Network
Recognition methods solves the problems, such as that Artificial Neural Network Structures are difficult to select, hyper parameter is difficult to determine the low efficiency of model buildings.
To achieve the above object, present invention employs the following technical solutions:
A kind of image-recognizing method based on depth Evolutionary Neural Network, comprising the following steps:
S1. according to problem of image recognition data, choose chromosome solution space and mutation operation parameter sets, mutation operation
Resource selection is the correspondence mutation operation of chromosome solution space selection, obtains the attribute letter of the corresponding level of corresponding chromosome solution space
Breath;
S2. it is based on initialization of population algorithm, obtains the grouping number N in initial population, the calculating of initialization of population algorithm
Formula is as follows:
Wherein, PZ is the size of initial population, and Num is the threshold value of variation, and N is the grouping in the initial population to be calculated
Number, as a result needs to round up;
S3. by all individuals of N group initial population N-1 times mutation operation, initial individuals group is obtained, the value of N is positive
Integer;
S4. the judge of environment fitness will be carried out to all individuals according to environment fitness function, fitness function formula is such as
Under:
Fit (Chromosome)=" Val " _ accuracy+ α log (C)
Wherein Fit is environment fitness function, and Chromosome is the individual of selection, and " Val " _ accuracy is individual
Performance in test data set, C are the complexities of model;When the response functional value of individual is constant, this final image is obtained
The model structure of identification mission;When the variation of the response functional value of individual, step S5 is executed;
S5. parent chromosome is obtained from the individual in initial individuals group using parent chromosome selection algorithm, then with
Machine chooses mutation operation from mutation operation set and is once made a variation, and executes step S4.
Further, taking chromosome solution space includes the network layer in deep learning, and network layer is specially convolutional Neural net
The convolutional layer and pond layer abstracted in network, the full articulamentum abstracted in full Connection Neural Network, active coating, batch normalizing
Change layer and Softmax layers;The attribute information of the corresponding level of chromosome solution space includes the convolution kernel size of convolutional layer, convolution nucleus number
The information such as amount and convolution step-length.Convolution kernel magnitude range is the integer between 3 to 7, and convolution kernel quantitative range is between 10 to 100
Integer, convolution step-length range be 1 to 3 between integer.While level determines, also other data of this layer are carried out
The attribute information of both corresponding levels of chromosome solution space is recorded, such as when one convolutional layer of coding, while also to record convolution kernel
The information such as size, convolution nuclear volume and convolution step-length, these information can completely restore the structure of this in network layer.
Further, mutation operation set includes learning rate, insertion convolutional layer, deletes convolutional layer and change convolution step-length.
Further, the detailed process in step S5 are as follows:
S5-1. first according to tournament algorithm, i.e., Size chromosome is randomly choosed from all chromosome of this generation
As candidates;Then chromosome selection is carried out using roulette wheel selection, i.e., by all individuals according to environment fitness
The floating-point number interval to 0 to 1 is normalized in value;Finally according to each individual adaptation degree proportion, from this space with
Machine chooses a floating number, and the location of this floating number is corresponding individual by the individual as final choice;Finally repeat
This process M times can be obtained M parent chromosome;The formula of roulette algorithm is as follows:
Wherein ciFor the individual for participating in selection, Size is the individual sum for participating in championship, and fit is environment fitness letter
Number;
S5-2. M parent chromosome is chosen from mutation operation set to mutation operation at random and carries out P variation, is obtained
New child chromosome, N is equal to the product of M and P, therefore number of individuals M*P is identical as number of individuals N in parent population in new population.
Compared with the prior art, the invention has the following beneficial effects:
1. in invention, going to improve neural network using evolution algorithm, neural network model can solve in processing image
When identification problem, Artificial Neural Network Structures are difficult to select, hyper parameter is difficult to determine the low efficiency problem of model buildings.
2. in the present invention, using didactic model buildings method, it is possible to reduce the manpower object during model buildings
Power consumption, improves the efficiency of model buildings.
3. in the present invention, optimizing initial population and generating the calculation of many places such as algorithm, environment fitness function, parent selection algorithm
Method details, final classification model can be improved builds effect.
Detailed description of the invention
Fig. 1 is flow diagram of the invention;
Minimum individual schematic diagram in Fig. 2 present invention;
Fig. 3 is Mutation parameter operation diagram used in the present invention;
Fig. 4 is the image recognition data set schematic diagram of MNIST handwritten numeral.
Fig. 5 is the optimal models structure of the image recognition data set of MNIST handwritten numeral
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that described herein, specific examples are only used to explain the present invention, not
For limiting the present invention.
Embodiment 1
Referring to Fig. 1, a kind of image-recognizing method based on depth Evolutionary Neural Network inputs image to be processed first and knows
Other data, these data seek to the problem of building model treatment;Secondly algorithm to start determine chromosome solution space and
The operational set of mutation operation;Then start initialization algorithm, and use environment fitness function to individuals all in population into
Row environment fitness evaluation, and parent chromosome selection is carried out with parent selection algorithm, then these selected parents are contaminated
Colour solid carries out mutation operation to generate new population;Finally after excessive generation circulation, environment fitness in last generation is selected
Highest individual is used as final image identification model.Specific step is as follows:
S1. according to problem of image recognition data, choose chromosome solution space and mutation operation parameter sets, mutation operation
Resource selection is the correspondence mutation operation of chromosome solution space selection, obtains the attribute letter of the corresponding level of corresponding chromosome solution space
Breath;
S2. it is based on initialization of population algorithm, obtains the grouping number N in initial population, the calculating of initialization of population algorithm
Formula is as follows:
Wherein, PZ is the size of initial population, and Num is the threshold value of variation, and N is the grouping in the initial population to be calculated
Number, as a result needs to round up;Individual in grouping is initialized, and guarantees that each individual is individual for minimum, then right
Then each group of carry out label, label range carry out the mutation operation of group number number from 1 to N to individual in each group.This reality
Applying population at individual number PZ in example is 100, and the variation threshold Num of setting is 20.Can be divided into 5 groups first, each group have 20 most
The chromosome of small individual;
S3. by all individuals of N group initial population N-1 times mutation operation, initial individuals group is obtained, the value of N is positive
Integer;Both first group was without making a variation, and all individuals all carry out 1 variation, all individuals in third group in second group
2 variations are all carried out, and so on.
S4. the judge of environment fitness will be carried out to all individuals according to environment fitness function, fitness function formula is such as
Under:
Fit (Chromosome)=" Val " _ accuracy+ α log (C)
Wherein Fit is environment fitness function, and Chromosome is the individual of selection, and " Val " _ accuracy is individual
Performance in test data set, C are the complexities of model;When the response functional value of individual is constant, this final image is obtained
The model structure of identification mission;Model complexity is higher, and the value for calculating log (C) is bigger.And there is one to need what is set to answer
This penalty term is controlled between 0 to 1 this range, can control model complexity to entire Fitness by miscellaneous degree parameter alpha
The influence of function.When the variation of the response functional value of individual, step S5 is executed;
S5. parent chromosome is obtained from the individual in initial individuals group using parent chromosome selection algorithm, then with
Machine chooses mutation operation from mutation operation set and is once made a variation, and executes step S4;
Further, taking chromosome solution space includes the network layer in deep learning, and network layer is specially convolutional Neural net
The convolutional layer and pond layer abstracted in network, the full articulamentum abstracted in full Connection Neural Network, active coating, batch normalizing
Change layer and Softmax layers;The attribute information of the corresponding level of chromosome solution space includes the convolution kernel size of convolutional layer, convolution nucleus number
The information such as amount and convolution step-length.Convolution kernel magnitude range is the integer between 3 to 7, and convolution kernel quantitative range is between 10 to 100
Integer, convolution step-length range be 1 to 3 between integer.While level determines, also other data of this layer are carried out
The attribute information of both corresponding levels of chromosome solution space is recorded, such as when one convolutional layer of coding, while also to record convolution kernel
The information such as size, convolution nuclear volume and convolution step-length, these information can completely restore the structure of this in network layer.
Further, mutation operation set includes learning rate, insertion convolutional layer, deletes convolutional layer and change convolution step-length.
Further, the detailed process in step S5 are as follows:
S5-1. first according to tournament algorithm, i.e., Size chromosome is randomly choosed from all chromosome of this generation
As candidates;Then chromosome selection is carried out using roulette wheel selection, i.e., by all individuals according to environment fitness
The floating-point number interval to 0 to 1 is normalized in value;Finally according to each individual adaptation degree proportion, from this space with
Machine chooses a floating number, and the location of this floating number is corresponding individual by the individual as final choice;Finally repeat
This process M times can be obtained M parent chromosome;The formula of roulette algorithm is as follows:
Wherein ciFor the individual for participating in selection, Size is the individual sum for participating in championship, and fit is environment fitness letter
Number;
S5-2. M parent chromosome is chosen from mutation operation set to mutation operation at random and carries out P variation, is obtained
New child chromosome, N is equal to the product of M and P, therefore number of individuals M*P is identical as number of individuals N in parent population in new population.M's
Value is less than N.Tournament algorithm and roulette algorithm solve the diversity evolved designed for guaranteeing randomness.
The present invention can solve neural network model when handling problem of image recognition, and structure is difficult to select, hyper parameter
It is difficult to the problems such as determining.And use didactic model buildings method, it is possible to reduce the manpower object during model buildings
Power consumption, improves the efficiency of model buildings.Finally in algorithmic procedure, optimizes initial population and generate algorithm, environment adaptation
The many places algorithm details such as function, parent selection algorithm are spent, final classification model can be improved builds effect.Shown in sum up, this
Invention has been used neural network and evolution algorithm, can have been consumed with less manpower and material resources when in problem of image recognition,
Build the higher final mask of effect.
Embodiment 2
According to the method described above, it is tested.
(1) apply the present invention to the image recognition data set of MNIST handwritten numeral, the sample of MNIST data set is shown in figure
4, the data set number hand-written from 250 different peoples, amount to 70000 pictures, will use wherein 60000 as network
Training dataset, remaining 10000 are used as test data set.
(2) chromosome solution space will be carried out and mutation operation set determines.The selection of solution space are as follows: and full articulamentum: 0, volume
Lamination: 1, ReLU:2, batch normalization layer: 3, Sotmax layers: 4 };In selection Fig. 2 of mutation operation
{ ALTERLEARNINGRATE, IDENTITY, ALTER STRIDE, INSERT CONVOLUTION, REMOVE
CONVOLUTION, ALTER NUMBER OF CHANNELS, FILTER SIZE, INSERT FULLY CONNECTED }.
(3) training parameter will be configured, as shown in table 1.
Parameter name | Parameter description | Parameter value |
Number of epochs | Training set trains pass | 3 |
Batch size | Criticize size | 64 |
Weight initialisation | Weights initialisation method | Xavier |
Optimiser | Optimizer selection | Adam |
The setting of 1 training parameter of table
(4) network model training, training process such as Fig. 1 are carried out according to ready-portioned data set and configured training parameter
Shown, to obtain final classification model structure, model structure is as shown in figure 5, model can be incited somebody to action according to the pictorial information of input
Picture is assigned in respective class, pictorial information as shown in Figure 4, will be grouped into the class of 7 this number;
(5) network weight that training obtains in (4) is used to test sample in test set, it is finally hand-written to MNIST
The discrimination of number identification data set has reached 99.33%
Finally, it is stated that the above examples are only used to illustrate the technical scheme of the present invention and are not limiting, although referring to compared with
Good embodiment describes the invention in detail, those skilled in the art should understand that, it can be to skill of the invention
Art scheme is modified or replaced equivalently, and without departing from the objective and range of technical solution of the present invention, should all be covered at this
In the scope of the claims of invention.
Claims (4)
1. a kind of image-recognizing method based on depth Evolutionary Neural Network, which comprises the following steps:
S1. according to problem of image recognition data, choose chromosome solution space and mutation operation parameter sets, mutation operation set
It is selected as the correspondence mutation operation of chromosome solution space selection, obtains the attribute information of the corresponding level of corresponding chromosome solution space;
S2. it is based on initialization of population algorithm, obtains the grouping number N in initial population, the calculation formula of initialization of population algorithm
It is as follows:
Wherein, PZ is the size of initial population, and Num is the threshold value of variation, and N is the grouping number in the initial population to be calculated,
As a result it needs to round up;
S3. by all individuals of N group initial population N-1 times mutation operation, initial individuals group is obtained, the value of N is positive integer;
S4. the judge of environment fitness will be carried out to all individuals according to environment fitness function, fitness function formula is as follows:
Fit (Chromosome)=" Val " _ accuracy+ α log (C)
Wherein Fit is environment fitness function, and Chromosome is the individual of selection, and " Val " _ accuracy is that individual is being surveyed
The performance on data set is tried, C is the complexity of model;When the response functional value of individual is constant, this final image recognition is obtained
The model structure of task;When the variation of the response functional value of individual, step S5 is executed;
S5. parent chromosome is obtained from the individual in initial individuals group using parent chromosome selection algorithm, then at random from
Mutation operation is chosen in mutation operation set and carries out a mutation operation, executes step S4.
2. a kind of image-recognizing method based on depth Evolutionary Neural Network according to claim 1, which is characterized in that take
Chromosome solution space is full articulamentum, convolutional layer, ReLU and batch normalization layer.
3. a kind of image-recognizing method based on depth Evolutionary Neural Network according to claim 1, which is characterized in that become
ETTHER-OR operation set includes learning rate, insertion convolutional layer, deletes convolutional layer and change convolution step-length.
4. a kind of image-recognizing method based on depth Evolutionary Neural Network according to claim 1, which is characterized in that step
Detailed process in rapid S5 are as follows:
S5-1. first according to tournament algorithm, i.e., Size chromosome conduct is randomly choosed from all chromosome of this generation
Candidates;Then using roulette wheel selection carry out chromosome selection, i.e., by it is all individual according to environment fitness value into
Row normalizes to 0 to 1 floating-point number interval;Finally according to each individual adaptation degree proportion, selected at random from this space
A floating number is taken, the location of this floating number is corresponding individual by the individual as final choice;Finally repeat this mistake
Journey M times can be obtained M parent chromosome;The formula of roulette algorithm is as follows:
Wherein ciFor the individual for participating in selection, Size is the individual sum for participating in championship, and fit is environment fitness function;
S5-2. M parent chromosome is chosen from mutation operation set to mutation operation at random and carries out P variation, is obtained new
Child chromosome, N is equal to the product of M and P, therefore number of individuals M*P is identical as number of individuals N in parent population in new population.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910410375.8A CN110135498A (en) | 2019-05-17 | 2019-05-17 | A kind of image-recognizing method based on depth Evolutionary Neural Network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910410375.8A CN110135498A (en) | 2019-05-17 | 2019-05-17 | A kind of image-recognizing method based on depth Evolutionary Neural Network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110135498A true CN110135498A (en) | 2019-08-16 |
Family
ID=67574784
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910410375.8A Pending CN110135498A (en) | 2019-05-17 | 2019-05-17 | A kind of image-recognizing method based on depth Evolutionary Neural Network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110135498A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111292260A (en) * | 2020-01-17 | 2020-06-16 | 四川翼飞视科技有限公司 | Construction method of evolutionary neural network and hyperspectral image denoising method based on evolutionary neural network |
CN111353313A (en) * | 2020-02-25 | 2020-06-30 | 四川翼飞视科技有限公司 | Emotion analysis model construction method based on evolutionary neural network architecture search |
CN111814946A (en) * | 2020-03-17 | 2020-10-23 | 同济大学 | Image description automatic generation method based on multi-body evolution |
CN112836794A (en) * | 2021-01-26 | 2021-05-25 | 深圳大学 | Method, device and equipment for determining image neural architecture and storage medium |
CN113081025A (en) * | 2021-03-10 | 2021-07-09 | 华中科技大学 | New crown pneumonia intelligent diagnosis system and device based on lung sounds |
CN114220127A (en) * | 2021-12-29 | 2022-03-22 | 安徽大学 | Image identification method based on gradient-guided evolutionary algorithm |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101212681A (en) * | 2007-12-25 | 2008-07-02 | 海信集团有限公司 | Quick motion search method |
US20110142316A1 (en) * | 2009-10-29 | 2011-06-16 | Ge Wang | Tomography-Based and MRI-Based Imaging Systems |
CN105279555A (en) * | 2015-10-28 | 2016-01-27 | 清华大学 | Self-adaptive learning neural network implementation method based on evolutionary algorithm |
CN105629722A (en) * | 2016-02-23 | 2016-06-01 | 河海大学常州校区 | Adaptive-control multi-pool sewage dynamic distribution method and system based on fuzzy rule |
CN108334949A (en) * | 2018-02-11 | 2018-07-27 | 浙江工业大学 | A kind of tachytelic evolution method of optimization depth convolutional neural networks structure |
CN108846349A (en) * | 2018-06-08 | 2018-11-20 | 电子科技大学 | A kind of face identification method based on dynamic Spiking neural network |
CN109165720A (en) * | 2018-09-05 | 2019-01-08 | 深圳灵图慧视科技有限公司 | Neural network model compression method, device and computer equipment |
-
2019
- 2019-05-17 CN CN201910410375.8A patent/CN110135498A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101212681A (en) * | 2007-12-25 | 2008-07-02 | 海信集团有限公司 | Quick motion search method |
US20110142316A1 (en) * | 2009-10-29 | 2011-06-16 | Ge Wang | Tomography-Based and MRI-Based Imaging Systems |
CN105279555A (en) * | 2015-10-28 | 2016-01-27 | 清华大学 | Self-adaptive learning neural network implementation method based on evolutionary algorithm |
CN105629722A (en) * | 2016-02-23 | 2016-06-01 | 河海大学常州校区 | Adaptive-control multi-pool sewage dynamic distribution method and system based on fuzzy rule |
CN108334949A (en) * | 2018-02-11 | 2018-07-27 | 浙江工业大学 | A kind of tachytelic evolution method of optimization depth convolutional neural networks structure |
CN108846349A (en) * | 2018-06-08 | 2018-11-20 | 电子科技大学 | A kind of face identification method based on dynamic Spiking neural network |
CN109165720A (en) * | 2018-09-05 | 2019-01-08 | 深圳灵图慧视科技有限公司 | Neural network model compression method, device and computer equipment |
Non-Patent Citations (1)
Title |
---|
涂序彦 等著: "《广义人工智能》", 31 August 2012, 国防工业出版社 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111292260A (en) * | 2020-01-17 | 2020-06-16 | 四川翼飞视科技有限公司 | Construction method of evolutionary neural network and hyperspectral image denoising method based on evolutionary neural network |
CN111353313A (en) * | 2020-02-25 | 2020-06-30 | 四川翼飞视科技有限公司 | Emotion analysis model construction method based on evolutionary neural network architecture search |
CN111814946A (en) * | 2020-03-17 | 2020-10-23 | 同济大学 | Image description automatic generation method based on multi-body evolution |
CN111814946B (en) * | 2020-03-17 | 2022-11-15 | 同济大学 | Multi-body evolution-based automatic image description generation method |
CN112836794A (en) * | 2021-01-26 | 2021-05-25 | 深圳大学 | Method, device and equipment for determining image neural architecture and storage medium |
CN112836794B (en) * | 2021-01-26 | 2023-09-29 | 深圳大学 | Method, device, equipment and storage medium for determining image neural architecture |
CN113081025A (en) * | 2021-03-10 | 2021-07-09 | 华中科技大学 | New crown pneumonia intelligent diagnosis system and device based on lung sounds |
CN113081025B (en) * | 2021-03-10 | 2022-04-29 | 华中科技大学 | New crown pneumonia intelligent diagnosis system and device based on lung sounds |
CN114220127A (en) * | 2021-12-29 | 2022-03-22 | 安徽大学 | Image identification method based on gradient-guided evolutionary algorithm |
CN114220127B (en) * | 2021-12-29 | 2024-05-10 | 安徽大学 | Image recognition method based on gradient guided evolutionary algorithm |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110135498A (en) | A kind of image-recognizing method based on depth Evolutionary Neural Network | |
Salmeron et al. | Learning fuzzy cognitive maps with modified asexual reproduction optimisation algorithm | |
CN104077595B (en) | Deep learning network image recognition methods based on Bayesian regularization | |
Kruschke | Models of categorization | |
Fiszelew et al. | Finding optimal neural network architecture using genetic algorithms | |
CN110059716B (en) | Construction of CNN-LSTM-SVM network model and MOOC conjugate prediction method | |
CN109325516B (en) | Image classification-oriented ensemble learning method and device | |
CN109389171A (en) | Medical image classification method based on more granularity convolution noise reduction autocoder technologies | |
CN115169227B (en) | Design concept generation network construction method and concept scheme automatic generation method | |
Dell’Anna et al. | Evolving fuzzy logic systems for creative personalized socially assistive robots | |
Smith et al. | Diagrammatic student models: Modeling student drawing performance with deep learning | |
Hasenjäger et al. | Active learning in neural networks | |
Duan | Automatic generation and evolution of personalized curriculum based on genetic algorithm | |
Chanamarn et al. | Enhancing efficient study plan for student with machine learning techniques | |
CN110188978A (en) | A kind of university student's profession recommended method based on deep learning | |
CN114154839A (en) | Course recommendation method based on online education platform data | |
Sultana et al. | Knowledge discovery from recommender systems using deep learning | |
CN112860856A (en) | Intelligent problem solving method and system for arithmetic application problem | |
Dutt et al. | Hand written character recognition using artificial neural network | |
CN114943278B (en) | Continuous online group incentive method and device based on reinforcement learning and storage medium | |
CN110853707A (en) | Gene regulation and control network reconstruction method based on deep learning | |
Prosi et al. | The dynamical regime and its importance for evolvability, task performance and generalization | |
Chang et al. | Unsupervised query-based learning of neural networks using selective-attention and self-regulation | |
Khan et al. | Pruning convolutional filters using batch Bridgeout | |
CN109635942A (en) | A kind of imitative brain excitement state and aepression working condition nerve network circuit structure and method |
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: 20190816 |
|
RJ01 | Rejection of invention patent application after publication |