CN110414578A - A kind of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion - Google Patents
A kind of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion Download PDFInfo
- Publication number
- CN110414578A CN110414578A CN201910647933.2A CN201910647933A CN110414578A CN 110414578 A CN110414578 A CN 110414578A CN 201910647933 A CN201910647933 A CN 201910647933A CN 110414578 A CN110414578 A CN 110414578A
- Authority
- CN
- China
- Prior art keywords
- training
- model
- colour gamut
- picture
- dynamic
- 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
- 238000012549 training Methods 0.000 title claims abstract description 48
- 238000000034 method Methods 0.000 title claims abstract description 27
- 238000006243 chemical reaction Methods 0.000 title claims abstract description 25
- 238000013526 transfer learning Methods 0.000 title claims abstract description 20
- 238000012545 processing Methods 0.000 claims abstract description 29
- 230000002708 enhancing effect Effects 0.000 claims abstract description 19
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 15
- 238000012360 testing method Methods 0.000 claims abstract description 12
- 238000002474 experimental method Methods 0.000 claims abstract description 10
- 238000005516 engineering process Methods 0.000 claims abstract description 7
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 4
- 230000004927 fusion Effects 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 15
- 230000009466 transformation Effects 0.000 claims description 12
- BUGBHKTXTAQXES-UHFFFAOYSA-N Selenium Chemical compound [Se] BUGBHKTXTAQXES-UHFFFAOYSA-N 0.000 claims description 7
- 229910052711 selenium Inorganic materials 0.000 claims description 7
- 239000011669 selenium Substances 0.000 claims description 7
- 241000894007 species Species 0.000 claims description 7
- 230000000007 visual effect Effects 0.000 claims description 6
- 241001465754 Metazoa Species 0.000 claims description 5
- 239000000284 extract Substances 0.000 claims description 5
- 239000006002 Pepper Substances 0.000 claims description 3
- 230000006835 compression Effects 0.000 claims description 3
- 238000007906 compression Methods 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 abstract description 8
- 238000011160 research Methods 0.000 abstract description 3
- 230000008569 process Effects 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000013508 migration Methods 0.000 description 3
- 230000005012 migration Effects 0.000 description 3
- 238000012800 visualization Methods 0.000 description 3
- 238000013480 data collection Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000000513 principal component analysis Methods 0.000 description 2
- 102100040401 DNA topoisomerase 3-alpha Human genes 0.000 description 1
- 101000611068 Homo sapiens DNA topoisomerase 3-alpha Proteins 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000012550 audit Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000013434 data augmentation Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000003475 lamination Methods 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 238000009877 rendering Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/53—Querying
- G06F16/535—Filtering based on additional data, e.g. user or group profiles
-
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- 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
- G06F18/23—Clustering techniques
Abstract
The transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion that the invention discloses a kind of, include the following steps: 1), building local data sets: training dataset needed for building the building experiment of web crawlers collection network picture manually;2), colour gamut enhancing and Pixel-level processing;3), data cleansing;4), model selects;5), model training: convolutional neural networks structure is finely adjusted by the Fusion Features that the feature extracted describes operator and local data sets, it is trained using the enhanced data set of colour gamut, dynamic batch processing training is added simultaneously, obtains the new model that can be used;6), model deployment test.The invention proposes the colour gamut enhancing algorithms based on PCA to improve 4% or so than the model accuracy rate without using the technology of current mainstream with the multiple batches of training of dynamic, provides a good direction for the vast scholar using deep learning research.
Description
Technical field
The present invention relates to image classification fields, are related specifically to a kind of color based on PCA for deep learning data set
Domain enhances algorithm and is directed to transfer learning dynamic batch training algorithms.
Background technique
The VGGnet model of the invention of Oxonian Visual Geometry team in 2014 is in ISLVRC positioning match
It is won the championship title with 25.3% error rate;Christian Szegedy of the same year Google et al. proposes the win of GoogleNet model
Obtained the champion of ISLVRC in 2014.We can directly use for reference the outstanding model of forefathers and be applied to solve new ask in this way
Topic builds model, training pattern or the better effect of acquirement faster.It in other words, is from one or more the purpose of transfer learning
Knowledge, experience are extracted in a originating task, are then applied in the middle of a target domain.Same time instructs for deep learning model
Practice the problem of data set lacks, related science man proposes traditional data enhanced scheme, utilizes translation, rotation, scaling, noise people
For the various distortions of simulation picture in shooting process, good achievement is also achieved.
However, levels of audit quality is uneven due to associated depth Learning Studies personnel training limited sample size, and lack phase
The knowledge of professional domain is closed, this results in often cannot get the effect in ideal using migration models.And it is some high-quality at present
Data set resource grasped by big companies always, this requires we construct deep learning data set to fully consider it is positive and negative
Sample under sample and different distortion status is to increase the learning ability of model.Traditional data enhancing is based on computer graphic
The expansion of the pixel scale of shape, picture be based on RGB mode, computer be difficult the light and shade, the tone that intuitively capture color with
And profile, in this way to model identification accuracy rate promotion it is extremely limited, maintain essentially in about 2%.
Summary of the invention
It is a kind of based on the multiple batches of training of dynamic and colour gamut it is an object of the invention to aiming at the shortcomings in the prior art, provide
The transfer learning method of conversion, to solve the above problems.
Technical problem solved by the invention can be realized using following technical scheme:
A kind of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion, includes the following steps:
1) local data sets, are constructed: training data needed for building the building experiment of web crawlers collection network picture manually
Collection;
2), colour gamut enhancing and Pixel-level processing: HSV is converted by RGB for picture by python code and is sent into PCA calculation
Method is clustered, and is analyzed according to cluster result principal component, and the channel little to picture classification result is rejected;At Pixel-level
Reason includes carrying out random cropping, overturning and Fuzzy Processing to picture;
3), data cleansing: by python build automatic assembly line tool pipeline it is first not up to standard to resolution ratio into
And then row processing combines human assistance to be filtered main body unconspicuous picture in kind;
4), model selects: by transfer learning, using MobileNet_1.0_224, Mobilenet_0.75_ respectively
192, MobileNet_0.50_160, Mobilenet_0.25_128, InvecptionV3 model extracts model in ImageNet
The bottleneck obtained on data set describes operator;
5) Fusion Features of operator and local data sets, model training: are described to convolutional Neural by the feature extracted
Network structure is finely adjusted, and is trained using the enhanced data set of colour gamut, while dynamic batch processing training is added, and is obtained new
The model that can be used;
6), model deployment test: complete APP is built using Java language and carries out real-time species detection, and records correlation
Experiment parameter analyzes experimental result.
Further, the web crawlers is built using Selenium automatic technology, including statement browser object,
The URL for needing to crawl, the lower one page node and batch preservation picture of searching picture.
Further, the colour gamut enhancing includes traditional data enhancing and colour gamut conversion, and traditional data enhancing includes random
Rotation, local deformation, mirror image switch and salt-pepper noise;Colour gamut conversion uses PCA clustering algorithm, first converts HSV lattice for picture
Formula, using a transformation matrix A, picture after conversion is the image of three wave bands, respectively H band, S-band and V-band,
Original image matrix X=[X1, X2, X3] one group of new principal component matrix Y=[Y of transformation formation1, Y2, Y3], formula are as follows: [Y1, Y2,
Y3]=A [X1, X2, X3], wherein X1, X2, X3Vector respectively represents wave band 1,2,3, CfFor the covariance matrix of X:
Wherein λ is the feature vector of C, then A=λT;PCA transformation will carry out orthogonal transformation to all wave bands of HSV image, mention
Principal component wave band is taken, correlation is minimum between each principal component after making conversion, to information be concentrated on former in Y matrix
A principal component reduces information redundancy, optimizes data;Preceding 2 principal components contain the profile information of animal, filter out to identification figure
Do not have helpful jammr band;Again the picture memory compression after exporting is to 66%.
Further, the multiple batches of training method of the model training is as follows:
With the increase of exercise wheel number during model training, system is by adjust automatically batch processing size, by originally defaulting
Batch processing size batch_size=64, increase to batch_size=75, be finally stabilized to batch_size=100;In
In deep learning, batch processing size often decides the ability of model learning, by constantly adjusting batch processing size, so that model
It can adapt to the different impression visuals field;The information of picture different dimensions is mutually merged, dynamic batchparameters is voluntarily added
random_batchsize;Parameter variation range is that 65~100, random_batchsize indicates how much model training is criticized for after
Handle the variation range of data.
Compared with prior art, beneficial effects of the present invention are as follows:
It is main the present invention relates to a kind of multiple batches of training method of dynamic and based on the colour gamut converting algorithm of deep learning data set
It is used for image recognition and pattern match.This method extracts InceptionV3 first, and MobileNet model is in ImageNet
Bottleneck on data set describes operator, is then finely adjusted manually to model, freezes bottleneck layer, and the pond layer of oneself is added, volume
Lamination and full articulamentum.This method integrated use traditional data enhancing technology and colour gamut enhancing handle experimental data set.
Using the transfer learning multiple batches of trained InceptionV3 of dynamic, MobileNet model, and to the parameter of two models, accurate
Rate, response speed have done comprehensive comparison.The colour gamut based on PCA has been used to enhance the multiple batches of training of algorithm and dynamic than master at present
The model accuracy rate raising 4% or so without using the technology of stream, wherein Inception3 to the TOP3 accuracy rate of test set more
It is to be increased to 95% or more, this provides a good direction for the vast scholar using deep learning research.Experiment is most
Trained model is deployed in Android device afterwards, realizes the real-time grading of rare animal, it was demonstrated that the present invention has very high
Accuracy rate and stable runnability.
The present invention has greatly liberated dependence of the deep learning to data set, and data set is allowed preferably to help model training,
The effect of 1+1 > 2 is played, this method can make full use of experimental data set, the effect of transfer learning is played to the greatest extent, from
And it brings great convenience to related scientific research personnel and the people for engaging in image recognition work.
Detailed description of the invention
Fig. 1 is Selenium workflow.
Fig. 2 is part rare animal data set used by this experiment.
Fig. 3 is the realization process that PCA-HSV colour gamut enhances algorithm.
Fig. 4 is the legend carried out after the conversion of PCA-HSV colour gamut, and left figure is converted by PCA, and right figure is original image.
Fig. 5 is the specific structural details of the MobileNet after model fine tuning.
Fig. 6 is Mobilenet transfer learning normatron analogous diagram.
Fig. 7 is visual part convolutional neural networks layer in migration models.
Fig. 8 is this patent Mobilenet+ colour gamut enhancing+dynamic batch training accuracy rate visualization result.
Fig. 9 is this patent InpectionV3+ colour gamut enhancing+dynamic batch training accuracy rate visualization result.
Figure 10 is APP selection and interface of taking pictures.
Figure 11 is species identification result interface.
Figure 12 is a kind of schematic diagram of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion.
Specific embodiment
To be easy to understand the technical means, the creative features, the aims and the efficiencies achieved by the present invention, below with reference to
Specific embodiment, the present invention is further explained.
Referring to Fig. 1~Figure 12, a kind of transfer learning based on the multiple batches of training of dynamic and colour gamut conversion of the present invention
Method includes the following steps:
1) local data sets, are constructed: training data needed for building the building experiment of web crawlers collection network picture manually
Collection;
2) HSV face, colour gamut enhancing and Pixel-level processing: is converted by rgb color mode for picture by python code
Color model is simultaneously sent into Principal Component Analysis PCA algorithm and is clustered, and analyzes according to cluster result principal component, rejects to figure
The little channel of piece classification results;Pixel-level processing is mainly the processing etc. such as to carry out random cropping, overturning to picture and obscure;
3), data cleansing: by python build automatic assembly line tool pipeline it is first not up to standard to resolution ratio into
And then row processing combines human assistance to be filtered main body unconspicuous picture in kind;
4), model selects: by transfer learning, using MobileNet_1.0_224, Mobilenet_0.75_ respectively
192, MobileNet_0.50_160, Mobilenet_0.25_128, InvecptionV3 model extracts model in ImageNet
The bottleneck obtained on data set describes operator;
5) Fusion Features of operator and local data sets, model training: are described to convolutional Neural by the feature extracted
Network structure is finely adjusted, and is trained using enhanced data set, while dynamic batch processing training is added, obtain it is new can
With the model used;
6), model deployment test: complete APP is built using Java language and carries out real-time species detection, and records correlation
Experiment parameter analyzes experimental result.
It is as follows that web crawlers in step 1) builds details:
This method uses browser automated test frame Selenium automatic technology.Selenium is an automation
Testing tool can drive browser to execute the operations such as specific movement, such as click, drop-down using it.For some utilizations
For the page of JavaScript dynamic rendering, this grasp mode is highly effective.Generic browser is to Selenium, it is necessary to
Corresponding plug-in unit is installed.Specific process for using is shown in Fig. 1, including statement browser object, the URL for needing to crawl, searches picture
Lower one page node and batch save picture.Selenium browser automated test frame crawls the process of network picture, first
It first states browser object, constructs URL using python language, click next page using the movement of browser plug-in analog subscriber,
The picture crawled batch is preserved.Relevant experimental data collection is shown in Fig. 2, tests the data set brief introduction voluntarily constructed: data
Collection covers 74 species, each 200 picture of species altogether.It wherein include original image and color gamut conversion, the operation such as Random-Rotation.
It is mainly as follows that colour gamut in step 2) converts the method for including to be related to:
The colour gamut enhancing includes data enhancing and colour gamut conversion.Traditional data enhancing: become including Random-Rotation, part
Shape, mirror image switch and salt-pepper noise;Colour gamut conversion uses PCA clustering algorithm, first converts HSV format for picture, utilizes one
Transformation matrix A, by taking the multispectral image of three wave bands of picture as an example, generally H band, S-band and V-band, specific algorithm
Process is shown in Fig. 3.HSV transformation is that RGB image is converted into HSV image model, is turned first to the image of RGB using python tool
It is melted into HSV mode, obtained picture is subjected to principal component analysis, unrelated channel is rejected, then picture is synthesized, is reduced into
HSV.Original image matrix X=[X1, X2, X3] one group of new principal component matrix Y=[Y of transformation formation1, Y2, Y3], formula are as follows:
[Y1, Y2, Y3]=A [X1, X2, X3], wherein X1, X2, X3Vector respectively represents wave band 1,2,3, CfFor the covariance matrix of X:
Wherein λ is the feature vector of C, then A=λT;PCA transformation will carry out orthogonal transformation to all wave bands of HSV image, mention
Principal component wave band is taken, correlation is minimum between each principal component after making conversion, so that information is concentrated mainly in Y matrix
Preceding several principal components reduce information redundancy, optimize data.Preceding 2 principal components contain the profile information of most animals, filter
Except to identification figure do not have helpful jammr band.And the picture memory compression after exporting again is to 66%.Specific PCA processing
Evaluation table is as follows, and Fig. 4 is related legend.
1 PCA of table handles evaluation table
The multiple batches of training method that model training stage is related in step 5) is as follows:
With the increase of exercise wheel number during model training, system is by adjust automatically batch processing size, by originally defaulting
Batch processing size batch_size=64, increase to batch_size=75, be finally stabilized to batch_size=100.In
In deep learning, batch processing size often decides the ability of model learning, by constantly adjusting batch processing size, so that model
It can adapt to different receptive fields.The information of picture different dimensions is mutually merged, referring in particular to official's script of Google
retrain.py.Dynamic batchparameters random_batchsize is voluntarily added for algorithm;Parameter variation range is 65
~100, random_batchsize indicate model training how much for rear batching data variation range.
The fine tuning of correlation model is described as follows:
Delete the last average pond layer of Inception_V3 and Mobilenet, full articulamentum and Softmax layers, for
Local data sets are added one layer of maximum pond layer, step-length 7, and one layer of convolutional network layer (3 × 3 are added after maximum pond layer
× 1024) it, then reconnects one layer of average pond layer and makees down-sampled, step-length 7.It is eventually adding one layer of full articulamentum and obtains 512
× 70 two-dimentional tensor, experiment use Softmax as classification valuation functions.The detail of the model is shown in Fig. 5.
Experimental result is visualized and its is analyzed as follows in step 6):
Improved Mobilenet model emulation result is shown in Fig. 6.Using Tensorboard to the mobileNet after fine tuning
Model is visualized, and input represents the input of picture array, and MobilenetV1 represents the MobileNet mould of migration
Type, is added the full articulamentum of oneself later, and classification results are assessed using Cross_entropy.
Fig. 7 is shown in transfer learning Inception_V3 frozen bottleneck layer part convolutional neural networks visualization.
Wherein analysis of experimental results is as follows:
Model accuracy tests five kinds of models more altogether.1.0_224,0.75_199 under MobileNet model,
0.50_168,0.25_128 and Inception_V3, wherein 224,199,168,128 refer to the resolution ratio of input picture.
Default indicates the number that last full articulamentum is only adjusted to model, and data augmentation is indicated to model application
Colour gamut conversion, random_batchsize expression to model apply dynamic batch system.
The visual presentation of MobileNet model accuracy rate result is shown in Fig. 8, Inception_V3 model accuracy rate result
Visual presentation is shown in Fig. 9.
Eventually by the continuous debugging of model, model is deployed to Android mobile phone platform and is tested.
Picture selects and takes pictures to obtain interface and see Figure 10.
Species identification result shows that Figure 11 is seen at interface.
The different model accuracys rate of table 2 compare
Experimental situation setting is described as follows:
Use (SuSE) Linux OS, laptop running memory 4G, hard-disk capacity 20G, by TensorFlow depth
Learning framework is trained, and is verified and is tested.Training dataset is divided into 80% training set using 10 folding cross validations, and 10% tests
Card collection and 10% test set.Training pattern attempted a variety of model summers different parameters combination, to explore MobileNet and
The optimum performance of InceptionV3, while it being directed to experimental result, the Space-time Complexity of each model is analyzed.
Unified Model is trained using back-propagation algorithm.All layers are by identical learning rate 0.001, with random
Batch processing size carries out algorithm tuning.
Experimental result is visualized using TensorBoard, the accuracy of main presentation training, loss late.Wherein mould
Type accuracy computation method is calculated using Softmax formula.
The basic principles, main features and advantages of the present invention have been shown and described above.The technology of the industry
Personnel are it should be appreciated that the present invention is not limited to the above embodiments, and the above embodiments and description only describe this
The principle of invention, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these changes
Change and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and its
Equivalent thereof.
Claims (4)
1. a kind of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion, characterized by the following steps:
1) local data sets, are constructed: training dataset needed for building the building experiment of web crawlers collection network picture manually;
2), colour gamut enhancing and Pixel-level processing: HSV is converted by RGB for picture by python code and be sent into PCA algorithm into
Row cluster, analyzes principal component according to cluster result, rejects the channel little to picture classification result;Pixel-level processing packet
It includes and random cropping, overturning and Fuzzy Processing is carried out to picture;
3), data cleansing: build that automatic assembly line tool pipeline is not first up to standard to resolution ratio to be located by python
And then reason combines human assistance to be filtered main body unconspicuous picture in kind;
4), model select: by transfer learning, respectively using MobileNet_1.0_224, Mobilenet_0.75_192,
MobileNet_0.50_160, Mobilenet_0.25_128, InvecptionV3 model extract model in ImageNet data
The bottleneck obtained on collection describes operator;
5) Fusion Features of operator and local data sets, model training: are described to convolutional neural networks by the feature extracted
Structure is finely adjusted, and is trained using the enhanced data set of colour gamut, while dynamic batch processing training is added, obtain it is new can
With the model used;
6), model deployment test: complete APP is built using Java language and carries out real-time species detection, and records related experiment
Parameter analyzes experimental result.
2. the transfer learning method according to claim 1 based on the multiple batches of training of dynamic and colour gamut conversion, feature exist
In building using Selenium automatic technology for: the web crawlers, including statement browser object, the URL for needing to crawl,
The lower one page node and batch for searching picture save picture.
3. the transfer learning method according to claim 1 based on the multiple batches of training of dynamic and colour gamut conversion, feature exist
In: colour gamut enhancing include traditional data enhancing and colour gamut conversion, traditional data enhancing include Random-Rotation, local deformation,
Mirror image switch and salt-pepper noise;Colour gamut conversion uses PCA clustering algorithm, first converts HSV format for picture, is converted using one
Matrix A, the picture after conversion is the image of three wave bands, respectively H band, S-band and V-band, original image matrix X=
[X1, X2, X3] one group of new principal component matrix Y=[Y of transformation formation1, Y2, Y3], formula are as follows: [Y1, Y2, Y3]=A [X1, X2, X3],
Wherein X1, X2, X3Vector respectively represents wave band 1,2,3, CfFor the covariance matrix of X:
Wherein λ is the feature vector of C, then A=λT;PCA transformation will to all wave bands of HSV image carry out orthogonal transformation, extract it is main at
Subrane, make conversion after each principal component between correlation it is minimum, thus before information is concentrated in Y matrix it is several it is main at
Point, information redundancy is reduced, data are optimized;Preceding 2 principal components contain the profile information of animal, filter out and do not help to identification figure
The jammr band helped;Again the picture memory compression after exporting is to 66%.
4. the transfer learning method according to claim 1 based on the multiple batches of training of dynamic and colour gamut conversion, feature exist
In: the multiple batches of training method of the model training is as follows:
With the increase of exercise wheel number, system adjust automatically batch processing size, by batch originally defaulted during model training
Size batch_size=64 is managed, increases to batch_size=75, is finally stabilized to batch_size=100;In depth
In habit, batch processing size often decides the ability of model learning, by constantly adjusting batch processing size, model is fitted
Answer the different impression visuals field;The information of picture different dimensions is mutually merged, dynamic batchparameters random_ is voluntarily added
batchsize;Parameter variation range is that 65~100, random_batchsize indicates model training how much for rear batch processing number
According to variation range.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910647933.2A CN110414578A (en) | 2019-07-16 | 2019-07-16 | A kind of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910647933.2A CN110414578A (en) | 2019-07-16 | 2019-07-16 | A kind of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110414578A true CN110414578A (en) | 2019-11-05 |
Family
ID=68361883
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910647933.2A Pending CN110414578A (en) | 2019-07-16 | 2019-07-16 | A kind of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110414578A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110991349A (en) * | 2019-12-05 | 2020-04-10 | 中国科学院重庆绿色智能技术研究院 | Lightweight vehicle attribute identification method based on metric learning |
CN111382787A (en) * | 2020-03-06 | 2020-07-07 | 芯薇(上海)智能科技有限公司 | Target detection method based on deep learning |
CN111507474A (en) * | 2020-06-18 | 2020-08-07 | 四川大学 | Neural network distributed training method for dynamically adjusting Batch-size |
CN117055853A (en) * | 2022-11-27 | 2023-11-14 | 华东师范大学 | PCA analysis software based on ToF-SIMS mass spectrum data and use method |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103034874A (en) * | 2011-09-29 | 2013-04-10 | 上海中医药大学 | Face gloss analytical method based on inspection diagnosis of traditional Chinese medical science |
CN104537639A (en) * | 2014-11-25 | 2015-04-22 | 西安应用光学研究所 | Translation evaluation based optimized spectrum image data fusion method |
CN108764455A (en) * | 2018-05-17 | 2018-11-06 | 南京中兴软件有限责任公司 | Parameter adjustment method, device and storage medium |
CN109522855A (en) * | 2018-11-23 | 2019-03-26 | 广州广电银通金融电子科技有限公司 | In conjunction with low resolution pedestrian detection method, system and the storage medium of ResNet and SENet |
-
2019
- 2019-07-16 CN CN201910647933.2A patent/CN110414578A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103034874A (en) * | 2011-09-29 | 2013-04-10 | 上海中医药大学 | Face gloss analytical method based on inspection diagnosis of traditional Chinese medical science |
CN104537639A (en) * | 2014-11-25 | 2015-04-22 | 西安应用光学研究所 | Translation evaluation based optimized spectrum image data fusion method |
CN108764455A (en) * | 2018-05-17 | 2018-11-06 | 南京中兴软件有限责任公司 | Parameter adjustment method, device and storage medium |
CN109522855A (en) * | 2018-11-23 | 2019-03-26 | 广州广电银通金融电子科技有限公司 | In conjunction with low resolution pedestrian detection method, system and the storage medium of ResNet and SENet |
Non-Patent Citations (2)
Title |
---|
YISHENG SONG等: "Species recognition technology based on migration learning and data augmentation", 《2018 5TH INTERNATIONAL CONFERENCE ON SYSTEMS AND INFORMATICS (ICSAI)》 * |
宋益盛等: "基于迁移学习和数据增强技术的物种识别", 《现代计算机》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110991349A (en) * | 2019-12-05 | 2020-04-10 | 中国科学院重庆绿色智能技术研究院 | Lightweight vehicle attribute identification method based on metric learning |
CN111382787A (en) * | 2020-03-06 | 2020-07-07 | 芯薇(上海)智能科技有限公司 | Target detection method based on deep learning |
CN111507474A (en) * | 2020-06-18 | 2020-08-07 | 四川大学 | Neural network distributed training method for dynamically adjusting Batch-size |
CN111507474B (en) * | 2020-06-18 | 2022-07-01 | 四川大学 | Neural network distributed training method for dynamically adjusting Batch-size |
CN117055853A (en) * | 2022-11-27 | 2023-11-14 | 华东师范大学 | PCA analysis software based on ToF-SIMS mass spectrum data and use method |
CN117055853B (en) * | 2022-11-27 | 2024-02-13 | 华东师范大学 | PCA analysis method based on TOF-SIMS mass spectrum data |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110414578A (en) | A kind of transfer learning method based on the multiple batches of training of dynamic and colour gamut conversion | |
WO2022160771A1 (en) | Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model | |
CN107273800B (en) | Attention mechanism-based motion recognition method for convolutional recurrent neural network | |
CN110287849A (en) | A kind of lightweight depth network image object detection method suitable for raspberry pie | |
CN111340738B (en) | Image rain removing method based on multi-scale progressive fusion | |
CN107463954B (en) | A kind of template matching recognition methods obscuring different spectrogram picture | |
CN111160533A (en) | Neural network acceleration method based on cross-resolution knowledge distillation | |
CN104268590B (en) | The blind image quality evaluating method returned based on complementary combination feature and multiphase | |
CN111723239B (en) | Video annotation method based on multiple modes | |
CN106295613A (en) | A kind of unmanned plane target localization method and system | |
CN110309723A (en) | A kind of driving behavior recognition methods based on characteristics of human body's disaggregated classification | |
CN112307995A (en) | Semi-supervised pedestrian re-identification method based on feature decoupling learning | |
CN112489164B (en) | Image coloring method based on improved depth separable convolutional neural network | |
CN112749663B (en) | Agricultural fruit maturity detection system based on Internet of things and CCNN model | |
CN111476133B (en) | Unmanned driving-oriented foreground and background codec network target extraction method | |
CN109829924A (en) | A kind of image quality evaluating method based on body feature analysis | |
CN108898092A (en) | Multi-spectrum remote sensing image road network extracting method based on full convolutional neural networks | |
Pan et al. | FDPPGAN: remote sensing image fusion based on deep perceptual patchGAN | |
Tan et al. | A Lightweight Underwater Object Detection Model: FL-YOLOV3-TINY | |
CN110033077A (en) | Neural network training method and device | |
CN108986103A (en) | A kind of image partition method merged based on super-pixel and more hypergraphs | |
CN112488963A (en) | Method for enhancing crop disease data | |
Zhang et al. | Remote sensing image retrieval based on DenseNet model and CBAM | |
CN114049567A (en) | Self-adaptive soft label generation method and application in hyperspectral image classification | |
Jiao et al. | A Novel Data Augmentation Method to Enhance the Training Dataset for Road Extraction from Historical Maps |
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: 20191105 |
|
RJ01 | Rejection of invention patent application after publication |