CN110399917B - Image classification method based on hyper-parameter optimization CNN - Google Patents

Image classification method based on hyper-parameter optimization CNN Download PDF

Info

Publication number
CN110399917B
CN110399917B CN201910671268.0A CN201910671268A CN110399917B CN 110399917 B CN110399917 B CN 110399917B CN 201910671268 A CN201910671268 A CN 201910671268A CN 110399917 B CN110399917 B CN 110399917B
Authority
CN
China
Prior art keywords
cnn
hyper
particle
iteration
parameters
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910671268.0A
Other languages
Chinese (zh)
Other versions
CN110399917A (en
Inventor
付俊
王思淼
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Northeastern University China
Original Assignee
Northeastern University China
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Northeastern University China filed Critical Northeastern University China
Priority to CN201910671268.0A priority Critical patent/CN110399917B/en
Publication of CN110399917A publication Critical patent/CN110399917A/en
Application granted granted Critical
Publication of CN110399917B publication Critical patent/CN110399917B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Medical Informatics (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides an image classification method based on hyper-parameter optimization (CNN), and belongs to the technical field of image recognition. The method selects and selects the structural parameters of the convolutional layer C1 and the pooling layer P1 as the hyper-parameters of the invention according to the structural characteristics of the CNN architecture, and limits the value range of the hyper-parameters to be (X) l ,X u ). Then, the hyper-parameters of the CNN are optimized by adopting two periodic variation PSO algorithms of global variation and local variation, so that the condition that the optimization of the hyper-parameters of the traditional PSO algorithm stays at local optimum is avoided, and the image classification performance which is more competitive than that of the traditional PSO algorithm is obtained. The efficiency and the cost of deep learning CNN hyper-parameter optimization are obviously improved, the image classification potential of a CNN framework is furthest exerted, hardware resources and calculation cost when the CNN carries out image classification are saved, and the method has a certain application value in engineering practice.

Description

Image classification method based on hyper-parameter optimization CNN
Technical Field
The invention relates to the technical field of image recognition, in particular to an image classification method based on hyper-parameter optimization (CNN).
Background
The image classification technology has been developed more and more mature, and CNN architectures suitable for different scene classifications are infinite, but a complex CNN structure usually consumes hardware resources and calculation cost. Before the CNN is used for image classification training, some parameters in the CNN need to be set in advance, the parameters are called as hyper-parameters, and the classification performance of the CNN image can be improved to the greatest extent on the premise of not changing the structure of the CNN by selecting a group of optimal hyper-parameters. Therefore, it is particularly important in engineering practice to select an appropriate hyper-parameter to fully release the image classification performance of the CNN architecture.
There have been some efforts in the study of hyper-parametric optimization methods for image classification, and earlier studies focused on the use of machine-learned hyper-parametric optimization methods for CNN. The hyper-parameter optimization method is mainly divided into model-free optimization and model-based optimization, the most advanced method of the hyper-parameter optimization method comprises simple grid and random search, and the most advanced method comprises a heuristic optimization algorithm based on population and Bayesian optimization (GP) based on Gaussian process. Heuristic optimization algorithms are of particular interest for CNN hyper-parametric optimization, where particle swarm optimization has proven to be very effective in solving multiple tasks in many areas due to its simplicity and versatility, and it has great potential for large-scale parallelization. The search efficiency of the super-parameter optimization based on the particle swarm optimization is far superior to that of other super-parameter optimization algorithms such as grid search and random search, the search time of the super-parameter optimization is shortened, and the problems of low optimization efficiency, time consumption and the like of the traditional super-parameter optimization are solved. However, the particle swarm optimization has the problem of being prone to fall into local optimization, which causes the super-parameter optimization to stay in the local optimization rather than a group of globally optimal super-parameters, so that a group of super-parameters which enable the CNN performance to be optimal cannot be searched to a certain extent, and the CNN image classification cannot achieve the optimal result.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides an image classification method based on hyper-parameter optimization (CNN).
The technical scheme adopted by the invention is as follows:
an image classification method based on hyper-parameter optimization (CNN) comprises the following steps:
step 1: preprocessing an image data set needing to be classified, and proportionally dividing the image data set into a training set T, a testing set C and a verification set V, wherein the verification set V and the training set T are extracted from image data and satisfy | T | =10 LiV |.
And 2, step: and (3) building a CNN (convolutional neural network) framework which comprises a convolutional layer C1, a pooling layer P1, a convolutional layer C2 and a pooling layer P2 and is activated and terminated by Softmax. Selecting the structural parameters of the convolutional layer C1 and the pooling layer P1 as the hyper-parameters of the invention according to the structural characteristics of the CNN architecture, and determining the value range of the hyper-parameters as (X) l ,X u )。
And step 3: optimizing hyper-parameters of the CNN framework by adopting a periodic variation PSO algorithm on the verification set to obtain a group of hyper-parameter values X i (g)；
Step 3-1: initial particle velocity, fitness function value, individual optimum position P g Global optimum position P i Setting the iteration number g as 0, the iteration precision delta as more than or equal to 0, the particle search space dimension as D and the number of particles as N;
the current position vector of each particle in the population is X i ＝(x i,1 ,x i,2 ,...,x i,D ) I =1,2,.., N, with the current velocity vector being V i ＝(v i,1 ,v i,2 ,...,v i,D ) I =1,2,.., N, with the individual optimum position vector P g ＝(p g，1 ,p g,2 ,...,p g,D ) G =1,2,.., N, with the global optimum position vector P i ＝(p i1 ,p i2 ,...,p iD )，i＝1,2,...,N；
Step 3-2: at the g-th iteration, each particle updates its velocity and position:
V i (g+1)＝ωV i (g)+c 1 r 1 (P i (g)-X i (g))+c 2 r 2 (P g (g)-X i (g)) (1)
X i (g+1)＝X i (g)+V i (g) (2)
where ω is the inertia factor, c 1 And c 2 Are learning factors, all are constants, r 1 And r 2 Is a random number from the range (0,1);
step 3-3: selecting a global mutation operator to change the positions of all particles in the whole population, or selecting a local mutation operator to change the positions of elite particles in the population:
the formula of the global mutation operator and the local mutation operator for the position change is as follows:
Figure GDA0004083066020000021
Figure GDA0004083066020000022
wherein A is 1 ，A 2 Is a self-defined rangeFactor, constant, r 3 ，r 4 0,1),
Figure GDA0004083066020000023
for elite particles, q is the number of new particles generated by the local mutation operator, f 1 Is the global variation frequency, f 2 Is the local variation frequency;
step 3-4: checking whether the speed and the position of the particles are out of range, and if so, replacing the corresponding particle value by the exceeding boundary value, wherein the specific judgment method comprises the following steps:
if V i (g)≤V l Then V is i (g)＝V 1 (ii) a If V i (g)≥V u Then V is i (g)＝V 1 (ii) a If X i (g)≤X 1 Then X i (g)＝X 1 (ii) a If X i (g)≥X u Then X i (g)＝X 1 ；
Wherein (V) 1 ,V u ) Is the velocity range of the particle, (X) 1 ,X u ) Is the position range of the particle;
step 3-5: the optimal position obtained by executing the steps 3-1 to 3-4 is the required over-parameter value X i (g)。
And 4, step 4: will exceed parameter X i (g) Inputting the optimized CNN into the CNN, and training the optimized CNN by using the training set obtained in the step 1;
and 5: inputting the test set obtained in the step 1 into the trained CNN to obtain a classification result of the test set C;
step 6: judging whether the iteration reaches a termination condition;
step 6-1: calculating the fitness function value of each particle period variation PSO:
Figure GDA0004083066020000031
wherein, CNN (X) i (g) Is the accuracy of the classification result obtained in step 5 of claim 1, X i (g) The super-molecule obtained in step 3 of claim 1A parameter value;
step 6-2: respectively updating individual historical optimal positions P by comparing the fitness function values of the particles of the iteration obtained in the step 6-1 i (g) And the group optimal position P g (g) Obtaining the optimal particle X of the iteration min (g)：
Figure GDA0004083066020000032
Figure GDA0004083066020000033
Step 6-3: judging the fitness value of the optimal particle to increase by less than a threshold value represented by epsilon, and judging the position update of the optimal particle in the population to be less than
Figure GDA0004083066020000034
The minimum step length is expressed, and whether the iteration number g reaches the maximum iteration number g or not is judged max If one of the above termination conditions is satisfied, the iteration is terminated.
And 7: if the termination condition is not reached, executing the step 3 to the step 6 to continue iteration;
and step 8: if the end condition is reached, the final optimal hyper-parameter is obtained and recorded as X min (g)；
And step 9: the final optimal hyperparameter X min (g) And substituting the image data into the CNN, and classifying the image of the whole data set to obtain a classification result.
Adopt the produced beneficial effect of above-mentioned technical scheme to lie in: the image classification method based on the hyperparametric optimization CNN effectively solves the problem that the particle swarm algorithm is easy to fall into local optimization, and further improves the convergence speed and the convergence precision of the particle swarm algorithm; the searching performance of the hyper-parameter optimization method is better, so that the problem of unsatisfactory CNN classification precision caused by improper hyper-parameter selection is avoided to a certain extent, the image classification precision is improved, and the performance of CNN image processing is exerted to the greatest extent.
Drawings
FIG. 1 is a flowchart of an image classification method based on hyper-parametric optimization CNN according to the present invention;
fig. 2 is a segment of a handwritten digit recognition MNIST dataset for image classification in a first embodiment of the invention;
FIG. 3 is a diagram showing the change of the classification performance of the MNIST data set with the iteration number before and after the improvement of the hyper-parameter optimization method in the first embodiment of the present invention;
FIG. 4 is a fragment of an object recognition cifar-10 dataset for image classification according to a second embodiment of the present invention;
fig. 5 is a graph showing the classification performance of the cifar-10 data set as a function of iteration number before and after the improvement of the hyper-parameter optimization method according to the second embodiment of the present invention.
Detailed Description
The following detailed description of embodiments of the present invention is provided in connection with the accompanying drawings and examples. The following examples are intended to illustrate the invention but are not intended to limit the scope of the invention.
Example 1
As shown in fig. 1, the method of the present embodiment is as follows.
Step 1: the reference data set MNIST data set is chosen as the data set to be classified, the data set being shown in fig. 2 as a segment, the data set having 70000 grayscale images, each image being 28 × 28 pixels and containing 10 classes, each class having 7000 images. 60000 images in the data set are randomly selected as a training set, the remaining 10000 images are selected as a testing set, and 6000 images are randomly selected from the training set as a verification set.
Step 2: and building a CNN architecture which comprises a convolutional layer C1, a pooling layer P1, a convolutional layer C2 and a pooling layer P2 and is activated and terminated by Softmax. Selecting the structural parameters of the convolutional layer C1 and the pooling layer P1 as the hyper-parameters of the invention according to the structural characteristics of the CNN architecture, and determining the value range of the hyper-parameters as (X) l ,X u ) As shown in table 1.
TABLE 1 parameters of convolutional and max-pooling layers and their allowable ranges
Figure GDA0004083066020000041
And step 3: optimizing hyper-parameters of the CNN framework by adopting a periodic variation PSO algorithm on the verification set to obtain a group of hyper-parameter values X i (g)；
Step 3-1: initial particle velocity, fitness function value, individual optimum position P g Global optimum position P i Setting the iteration number g as 0, the iteration precision delta is more than or equal to 0, the particle search space dimension D =4, and the number of particles N =10;
the current position vector of each particle in the population is X i ＝(x i,1 ,x i,2 ,...,x i,4 ) I =1,2, 10, with the current velocity vector being V i ＝(v i,1 ,v i,2 ,...,v i,4 ) I =1,2, 10, the individual optimal position vector is P g ＝(p g，1 ,p g,2 ,...,p g,4 ) G =1,2, 10, the global optimum position vector is P i ＝(p i1 ,p i2 ,...+p i4 )，i＝1,2,...,10；
Step 3-2: at the g-th iteration, each particle updates its velocity and position:
V i (g+1)＝ωV i (g)+c 1 r 1 (P i (g)-X i (g))+c 2 r 2 (P g (g)-X i (g)) (1)
X i (g+1)＝X i (g)+V i (g) (2)
wherein ω = c 1 ＝c 2 ＝0.5，r 1 And r 2 Is a random number from the range (0,1);
step 3-3: selecting a global mutation operator to change the positions of all particles in the whole population, or selecting a local mutation operator to change the positions of elite particles in the population:
the formula of the global mutation operator and the local mutation operator for the position change is as follows:
Figure GDA0004083066020000051
Figure GDA0004083066020000052
wherein A is 1 ＝A 2 ＝1，r 3 、r 4 Is a random number in the range of (0,1),
Figure GDA0004083066020000053
for elite particles, the number of new particles generated by the local mutation operator q =3, global mutation frequency f 1 =10, local variation frequency f 2 ＝2；
Step 3-4: checking whether the speed and position of the particles are out of range, and if so, replacing the corresponding particle value with the boundary value beyond the speed and position, wherein the specific judgment method comprises the following steps:
if V i (g)≤V l Then V is i (g)＝V 1 (ii) a If V i (g)≥V u Then V is i (g)＝V 1 (ii) a If X i (g)≤X 1 Then X i (g)＝X 1 (ii) a If X i (g)≥X u Then X i (g)＝X 1 ；
Wherein (V) 1 ,V u ) The particle velocity range is (-2,2), (X) 1 ,X u ) The position range values for the particles are as described in table 1;
step 3-5: the optimal position obtained by executing the steps 3-1 to 3-4 is the required over-parameter value X i (g)。
And 4, step 4: will exceed parameter X i (g) Inputting the optimized CNN into the CNN, and training the optimized CNN by using the training set obtained in the step 1;
and 5: inputting the test set obtained in the step 1 into the trained CNN to obtain a classification result of the image;
step 6: judging whether the iteration reaches a termination condition;
step 6-1: calculating a fitness function value of each particle period variation PSO, and drawing a fitness function curve, as shown in FIG. 3:
Figure GDA0004083066020000061
wherein, CNN (X) i (g) Is the accuracy, X, of the classification result obtained in step 5 of claim 1 i (g) The value of the hyperparametric value obtained in step 3 of claim 1;
step 6-2: respectively updating individual historical optimal positions P by comparing the particle fitness function values of the iteration obtained in the step 6-1 i (g) And the group optimal position P g (g) Obtaining the optimal particle X of the iteration min (g)：
Figure GDA0004083066020000062
Figure GDA0004083066020000063
Step 6-3: judging whether the increase of the optimal particle fitness value in five continuous generations is smaller than a threshold value epsilon =0.0001, and judging whether the update of the optimal particle position in the population in 5 continuous generations is smaller than the minimum step length
Figure GDA0004083066020000064
And judging whether the iteration time g reaches the maximum iteration time 20, and if one of the termination conditions is met, terminating the iteration.
And 7: if the termination condition is not reached, executing the step 3 to the step 6;
and 8: if the end condition is reached, the final optimal hyper-parameter is obtained and recorded as X min (g)；
And step 9: the final optimal hyperparameter X min (g) And substituting the image data into the CNN, and classifying the image of the whole data set to obtain a classification result.
Example 2
As shown in fig. 1, the method of the present embodiment is as follows.
Step 1: the reference dataset object identification cifar-10 dataset was chosen as the dataset to be classified, the segment of the dataset is shown in fig. 4, the dataset has 60000 32 × 32 pixel color images, which contain 10 classes, each of which has 6000 images. 50000 images in the data set are randomly selected as a training set, the remaining 10000 images are used as a test set, and 5000 images in the training set are randomly selected as a verification set.
Step 2 to step 9 are the same as those in example 1, and the classification performance of the cifar-10 data sets before and after the improvement of the obtained hyper-parameter optimization method is changed with the iteration number as shown in fig. 5.
The image classification accuracy comparison of the handwritten digit recognition MNIST data set and the object recognition cifar-10 data set before and after the improvement of the hyper-parameter optimization method is shown in the table 2:
TABLE 2 accuracy of CNN image classification for different datasets before and after hyper-parameter optimization
Figure GDA0004083066020000071
The result shows that the method of the invention improves the image classification accuracy to a certain extent on the premise of not changing the classification CNN framework, exerts the image classification potential of the CNN framework to the maximum extent, saves the hardware resources and the calculation cost when the CNN carries out image classification, and has a certain application value in engineering practice.

Claims (3)

1. An image classification method based on hyper-parameter optimization CNN is characterized by comprising the following steps:
step 1: preprocessing an image data set to be classified, and dividing the image data set into a training set T, a test set C and a verification set V in proportion;
step 2: building a CNN architecture, and selecting a hyper-parameter and a value range thereof according to the structural characteristics of the CNN architecture;
and step 3: optimizing CNN architecture by adopting periodic variation PSO algorithm on verification setHyper-parametric to obtain a set of hyper-parameter values X i (g)；
And 4, step 4: will exceed parameter X i (g) Inputting the optimized CNN into the CNN, and training the optimized CNN by using the training set obtained in the step 1;
and 5: inputting the test set C obtained in the step 1 into the trained CNN to obtain a classification result of the test set C;
step 6: judging whether the iteration reaches a termination condition;
and 7: if the termination condition is not reached, executing the step 3 to the step 6 to continue iteration;
and 8: if the end condition is reached, the final optimal hyper-parameter is obtained and is marked as X min (g)；
And step 9: the final optimal hyperparameter X is obtained min (g) Substituting into CNN, classifying the images of the whole data set to obtain a classification result;
the process of optimizing the hyper-parameters of the CNN architecture by adopting the periodic variation PSO algorithm in the verification set in the step 3 is as follows:
step 3-1: initializing particle speed, fitness function value and individual optimal position P g Global optimum position P i Setting the iteration number g as 0, the iteration precision delta as more than or equal to 0, the particle search space dimension as D and the number of particles as N;
the current position vector of each particle in the population is X i ＝(x i,1 ,x i,2 ,...,x i,D ) I =1,2,.., N, with the current velocity vector being V i ＝(v i,1 ,v i,2 ,...,v i,D ) I =1,2,.., N, with the individual optimum position vector P g ＝(p g，1 ,p g,2 ,...,p g,D ) G =1,2,.., N, with the global optimum position vector P i ＝(p i1 ,p i2 ,...,p iD )，i＝1,2,...,N；
Step 3-2: at the g-th iteration, each particle updates its velocity and position:
V i (g+1)＝ωV i (g)+c 1 r 1 (P i (g)-X i (g))+c 2 r 2 (P g (g)-X i (g)) (1)
X i (g+1)＝X i (g)+V i (g) (2)
where ω is the inertia factor, c 1 And c 2 Are learning factors, all are constants, r 1 And r 2 Is a random number from the range (0,1);
step 3-3: selecting a global mutation operator to change the positions of all particles in the whole population, or selecting a local mutation operator to change the positions of elite particles in the population:
the formula of the global mutation operator and the local mutation operator for the position change is as follows:
Figure FDA0004083066010000021
Figure FDA0004083066010000022
wherein A is 1 ，A 2 Is a self-defined amplitude factor, which is a constant, r 3 ，r 4 Is a random number in the range of (0,1),
Figure FDA0004083066010000023
for elite particles, q is the number of new particles generated by the local mutation operator, f 1 Is the global variation frequency, f 2 Is the local variation frequency;
step 3-4: checking whether the speed and position of the particles are out of range, and if so, replacing the corresponding particle value with the boundary value beyond the speed and position, wherein the specific judgment method comprises the following steps:
if V i (g)≤V l Then V is i (g)＝V 1 (ii) a If V i (g)≥V u Then V is i (g)＝V 1 (ii) a If X i (g)≤X 1 Then X i (g)＝X 1 (ii) a If X i (g)≥X u Then X i (g)＝X 1 ；
Wherein (V) 1 ,V u ) Is the velocity range of the particle, (X) 1 ,X u ) Is the position range of the particle;
step 3-5: the optimal position obtained by executing the steps 3-1 to 3-4 is the required over-parameter value X i (g)；
The process of judging whether the iteration reaches the termination condition in the step 6 is as follows:
step 6-1: calculating the fitness function value of each particle period variation PSO:
Figure FDA0004083066010000024
wherein, CNN (X) i (g) Is the accuracy, X, of the classification result obtained in said step 5 i (g) Obtaining the super parameter value obtained in the step 3;
step 6-2: respectively updating individual historical optimal positions P by comparing the particle fitness function values of the iteration obtained in the step 6-1 i (g) And the group optimal position P g (g) Obtaining the optimal particle X of the iteration min (g)：
Figure FDA0004083066010000025
Figure FDA0004083066010000026
And 6-3: judging the fitness value of the optimal particle to increase by less than a threshold value represented by epsilon, and judging the position update of the optimal particle in the population to be less than
Figure FDA0004083066010000031
The minimum step length is expressed, and whether the iteration number g reaches the maximum iteration number g or not is judged max If one of the above termination conditions is satisfied, the iteration is terminated.
2. The method for classifying images based on hyperparametric optimized CNN of claim 1, wherein the verification set V and the training set T in step 1 are extracted from image data, and satisfy | T | =10 calvities.
3. The image classification method based on hyper-parametric optimization (CNN) according to claim 1, wherein the CNN framework constructed in the step 2 comprises a convolutional layer C1, a pooling layer P1, a convolutional layer C2 and a pooling layer P2, and is terminated by Softmax activation, the structural parameters of the convolutional layer C1 and the pooling layer P1 are selected as the hyper-parameters according to the structural characteristics of the CNN framework, and the value range of the hyper-parameters is determined as (X) l ,X u )。
CN201910671268.0A 2019-07-24 2019-07-24 Image classification method based on hyper-parameter optimization CNN Active CN110399917B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910671268.0A CN110399917B (en) 2019-07-24 2019-07-24 Image classification method based on hyper-parameter optimization CNN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910671268.0A CN110399917B (en) 2019-07-24 2019-07-24 Image classification method based on hyper-parameter optimization CNN

Publications (2)

Publication Number Publication Date
CN110399917A CN110399917A (en) 2019-11-01
CN110399917B true CN110399917B (en) 2023-04-18

Family

ID=68324921

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910671268.0A Active CN110399917B (en) 2019-07-24 2019-07-24 Image classification method based on hyper-parameter optimization CNN

Country Status (1)

Country Link
CN (1) CN110399917B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110942090B (en) * 2019-11-11 2024-03-29 北京迈格威科技有限公司 Model training method, image processing device, electronic equipment and storage medium
CN111160459A (en) * 2019-12-30 2020-05-15 上海依图网络科技有限公司 Device and method for optimizing hyper-parameters
CN112197876A (en) * 2020-09-27 2021-01-08 中国科学院光电技术研究所 Single far-field type depth learning wavefront restoration method based on four-quadrant discrete phase modulation

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
CN109085469A (en) * 2018-07-31 2018-12-25 中国电力科学研究院有限公司 A kind of method and system of the signal type of the signal of cable local discharge for identification
CN109919202A (en) * 2019-02-18 2019-06-21 新华三技术有限公司合肥分公司 Disaggregated model training method and device

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11003994B2 (en) * 2017-12-13 2021-05-11 Cognizant Technology Solutions U.S. Corporation Evolutionary architectures for evolution of deep neural networks

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
CN109085469A (en) * 2018-07-31 2018-12-25 中国电力科学研究院有限公司 A kind of method and system of the signal type of the signal of cable local discharge for identification
CN109919202A (en) * 2019-02-18 2019-06-21 新华三技术有限公司合肥分公司 Disaggregated model training method and device

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
改进的基于粒子群优化的支持向量机特征选择和参数联合优化算法;张进等;《计算机应用》;20160510(第05期);全文 *

Also Published As

Publication number Publication date
CN110399917A (en) 2019-11-01

Similar Documents

Publication Publication Date Title
CN110399917B (en) Image classification method based on hyper-parameter optimization CNN
CN107392919B (en) Adaptive genetic algorithm-based gray threshold acquisition method and image segmentation method
CN108875933B (en) Over-limit learning machine classification method and system for unsupervised sparse parameter learning
CN110322445B (en) Semantic segmentation method based on maximum prediction and inter-label correlation loss function
CN111882040A (en) Convolutional neural network compression method based on channel number search
CN116503676B (en) Picture classification method and system based on knowledge distillation small sample increment learning
CN109934826A (en) A kind of characteristics of image dividing method based on figure convolutional network
CN109034062B (en) Weak supervision abnormal behavior detection method based on time sequence consistency
CN108985342A (en) A kind of uneven classification method based on depth enhancing study
Barman et al. Transfer learning for small dataset
CN110991621A (en) Method for searching convolutional neural network based on channel number
CN111931801B (en) Dynamic route network learning method based on path diversity and consistency
CN105512675A (en) Memory multi-point crossover gravitational search-based feature selection method
CN114882278A (en) Tire pattern classification method and device based on attention mechanism and transfer learning
CN112348153A (en) Group intelligent optimization acceleration method based on simplex local search
CN113191445A (en) Large-scale image retrieval method based on self-supervision countermeasure Hash algorithm
Wang et al. A novel imperialist competitive algorithm for multithreshold image segmentation
Zhang et al. The WuC-Adam algorithm based on joint improvement of Warmup and cosine annealing algorithms
CN107274357B (en) Gray level image enhancement processing system with optimal parameters
CN114581470B (en) Image edge detection method based on plant community behaviors
CN113592174B (en) Knowledge-driven two-dimensional polygonal cloth piece automatic discharging method
Zhou et al. Effective vision transformer training: A data-centric perspective
Lai et al. A LPSO-SGD algorithm for the Optimization of Convolutional Neural Network
CN113095489A (en) Inductive multi-level graph network representation learning method
Leconte et al. AskewSGD: an annealed interval-constrained optimisation method to train quantized neural networks

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