CN110263863A - Fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2 - Google Patents
Fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2 Download PDFInfo
- Publication number
- CN110263863A CN110263863A CN201910547744.8A CN201910547744A CN110263863A CN 110263863 A CN110263863 A CN 110263863A CN 201910547744 A CN201910547744 A CN 201910547744A CN 110263863 A CN110263863 A CN 110263863A
- Authority
- CN
- China
- Prior art keywords
- network
- bilinearity
- mushroom
- feature
- feature extraction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
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
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of the fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2, key step are as follows: (1) establish based on transfer learning and bilinear fine granularity mushroom phenotype identification model;(2) transfer learning and training are carried out based on identification model;(3) it is pre-processed after image being inputted identification model;(4) feature extraction is carried out to pretreated image data.The present invention combines the feature that two symmetrical InceptionResNetV2 feature extraction networks extract, and obtains more fine-grained feature, keeps recognition effect more preferable;And use the transfer learning training method based on model, the good feature extraction network parameter weight of the pre-training on ImageNet data set is moved on mushroom fine granularity phenotypic data collection, better constringency performance can be reached within the shorter training time, keep recognition result more preferable.
Description
Technical field
The invention belongs to computer, artificial intelligence and field of image processings, and in particular to it is a kind of based on transfer learning with
The fine granularity mushroom phenotype recognition methods of bilinearity InceptionResNetV2.
Background technique
Fine granularity image recognition (Fine-grained Image Recognition) at present be applied to vehicle cab recognition,
The fields such as birds identification, but because the categorical measure of mushroom is more, and different subclass similarities are high, identification difficulty is big, currently without
It specially can be used for the phenotypically recognized product of mushroom.
Although currently there are some fine granularity image recognition technologys, cannot preferable particulate be carried out to mushroom
Spend phenotype identification.Specifically mainly there are following some problems to need to solve:
(1) Model Weight for how using the transfer learning method based on model that pre-training on ImageNet data set is good
It moves in the fine granularity phenotype identification model of mushroom, reduces required data volume and training time, obtain preferably primality
Energy and constringency performance.
(2) how to converge operation using bilinearity to be combined the characteristics of image that two feature extraction networks extract,
It obtains more fine-grained feature and is used for image recognition.
(3) how the stronger InceptionResNetV2 feature extraction network progress image of ability in feature extraction is used
Feature extraction obtains better feature progress bilinearity and converges operation.
Summary of the invention
It is an object of the invention to overcome above-mentioned the shortcomings of the prior art, provide a kind of based on transfer learning and two-wire
The fine granularity mushroom phenotype recognition methods of property InceptionResNetV2, this method can be according to fine granularity mushroom phenotypic data
Collection training pattern simultaneously identifies different types of fine granularity mushroom phenotype image.
To achieve the above object, the technical solution adopted by the present invention is that: one kind be based on transfer learning and bilinearity
The fine granularity mushroom phenotype recognition methods of InceptionResNetV2 comprising following steps:
Step 1 is established based on transfer learning and bilinear fine granularity mushroom phenotype identification model;
Step 2 carries out transfer learning and training based on identification model;
Step 3 pre-processes after image is inputted identification model;
Step 4 carries out feature extraction to pretreated image data;Using symmetrical structure
InceptionResNetV2 feature extraction network extracts the feature vector in image, then to the feature vector extracted and its
Spontaneous transposition carries out bilinearity and converges operation obtaining the bilinearity eigenmatrix of each position of picture, and by bilinearity feature square
Battle array is converted into bilinearity feature vector, is followed by softmax layers finally by full articulamentum and carries out more points to bilinearity feature vector
Class obtains probability of all categories.
Further, the pretreatment in the step 3 includes centralization, normalization, scaling, random cropping and Random Level
Overturning.
Further, after the image data input Network Recognition model of arbitrary size, entire data set is subtracted first
Average value simultaneously carries out centralization and normalized divided by the standard deviation of entire data set, and image is scaled to short side and is later
448 pixels, and the square-shaped image region of 448*448 is cut out using the mode of random cropping from image, it is last random
Flip horizontal is carried out to image.
Further, using the InceptionResNetV2 net in Inception series of network model in the step 4
Network carries out feature extraction, and joined residual block in InceptionResNetV2 feature extraction network.
Further, first 7 layers of the InceptionResNetV2 network are by three-layer coil lamination, one layer of maximum pond
Layer, two layers of convolutional layer, one layer of maximum pond layer composition, are repeated 10 times the residual error Inception module having there are three branch later,
Further through a better simply Inception module, using 20 tools, there are two the residual error Inception modules of branch, again
By the Inception module of 4 branches, finally by 10 tools, there are two the residual error Inception modules of branch, then
Output result is obtained by a convolutional layer.
Further, bilinear model B is made of four-tuple, as shown in formula (1),
B=(fA,fB,P,C) (4)
Wherein fAAnd fBIt is characteristic function, P is the pond function of model, and C is the classification function of mushroom;
Feature of the feature of output on each position using matrix apposition combination from, as shown in formula (2),
bilinear(L,I,fA,fB)=fA(L,I)TfB(L,I) (5)
Wherein L indicates that position and scale, I indicate picture;If the dimension difference for extracting feature of two characteristic functions
For (K, M) and (K, N), then after bilinear bilinearity converges operation, dimension becomes (M, N), if using summation Chi Hualai
The feature of comprehensive each position, then as shown in formula (3),
Wherein Φ (I) indicates that global picture feature indicates;
Bilinearity feature vector x=Φ (I) is finally passed through into symbol square root transformationAnd increase L2
RegularizationClassifier is inputted again obtains classification results to the end.
Further, training process is divided into two steps:
(1) what InceptionResNetV2 feature extraction network fixed first loaded obtains on ImageNet data set
Pre-training parameter, only allow the parameter of the last full articulamentum random initializtion of training;
(2) after network convergence, then the parameter of solid InceptionResNetV2 feature extraction network is solved, use is lesser
Learning rate is finely adjusted.
Further, total training process is as follows:
(1) building is based on transfer learning and bilinear fine granularity mushroom phenotype identification model, wherein including
InceptionResNetV2 is as feature extraction network;
(2) ImageNet pre-training model initialization InceptionResNetV2 feature extraction network is used, is used
Glorot normal initialization device initializes full connection layer parameter;
(3) parameter of InceptionResNetV2 feature extraction network is fixed, the training process after being allowed to can not pass through
Backpropagation updates the parameter value of this part;
(4) training sample after obtaining image preprocessing in input channel, batch size are 8, and image size is
448*448;
(5) the batch training sample for getting (4) inputs network model, converges operation by feature extraction and bilinearity
And full articulamentum, probability of all categories is calculated finally by softmax;
(6) use classes cross entropy loss function calculates the penalty values of network model;
(7) by calculating gradient value, using SGD optimizer, it is 1.0 that initial learning rate, which is arranged, and learning rate decays to 1e-
8, momentum Momentum is set as 0.9, and error back propagation is returned whole network, updates the parameter of full articulamentum;
(8) judge whether to reach given number of iterations 100 or meet the 10 iteration variations of verifying penalty values to be no more than 0.001
Morning stop method condition, if, it is believed that network has been restrained, then enters step (9), if otherwise reentering step (4);
(9) change the learning rate of SGD optimizer to 0.001;
(10) fixation to InceptionResNetV2 feature extraction network pre-training parameter is released, network is led to
Cross the parameter value that backpropagation updates this part;
(11) training sample after obtaining image preprocessing in input channel, batch size are 8, and image size is
448*448;
(12) the batch training sample for getting (11) inputs network model, converges behaviour by feature extraction and bilinearity
Work and full articulamentum, probability of all categories is calculated finally by softmax;
(13) use classes cross entropy loss function calculates the penalty values of network model;
(14) by calculating gradient value, using SGD optimizer, it is 0.001 that initial learning rate, which is arranged, and learning rate decays to
1e-8, momentum Momentum are set as 0.9, and error back propagation is returned whole network, updates each layer of network of parameter;
(15) judge whether to reach given number of iterations 70 or meet the 10 iteration variations of verifying penalty values to be no more than 0.001
Morning stop method condition, if, it is believed that network has been restrained, then enters step (16), if otherwise reentering step (11);
(16) accuracy rate, accurate rate, recall rate, the F1 value of network model are calculated by test set.
The invention has the benefit that (1) is converged using bilinearity, by two symmetrical InceptionResNetV2 features
It extracts the feature that network extracts to combine, obtains more fine-grained feature, keep recognition effect more preferable.(2) it has used and has been based on
The transfer learning training method of model migrates the good feature extraction network parameter weight of the pre-training on ImageNet data set
Onto mushroom fine granularity phenotypic data collection, it can reach better constringency performance within the shorter training time, make recognition result
More preferably.
The present invention respectively with use the standard of symmetrical VGG16 and symmetrical VGG19 model on mushroom fine granularity phenotypic data collection
Four true rate, accurate rate, recall rate, F1 value results are compared, as shown in table 1.
1 result of table
As can be seen from the table, being learned based on migration using symmetrical InceptionResNetV2 network proposed by the present invention
Practise it is best with bilinear fine granularity mushroom phenotype identification model effect, reached 0.90 accuracy rate, 0.91 accurate rate,
0.90 recall rate, 0.90 F1 value, and other methods about 2%~6% are higher by indices.
Detailed description of the invention
Fig. 1 is the fine granularity mushroom phenotype identification model frame of bilinearity InceptionResNetV2.
Fig. 2 is pretreatment process figure.
Fig. 3 is Inception module network structure.
Fig. 4 is InceptionResNetV2 overall network structure.
Fig. 5 is transfer learning schematic diagram.
Fig. 6 is trained flow chart.
Specific embodiment
The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.
One, network model
The present invention chooses InceptionResNetV2 network as the feature extraction network in Bilinear CNN network,
Wish the effect that overall network model is promoted by the deeper stronger ability in feature extraction of network bring.
It results in based on transfer learning and bilinear fine granularity mushroom phenotype identification model, overall network structure
As shown in Figure 1, first passing around the pre- place of centralization, normalization, random cropping, Random Level overturning after image input network model
Reason process obtains the spy extracted from image via the InceptionResNetV2 feature extraction network of symmetrical structure later
Vector is levied, then are converged by operation and obtains each position of picture for the feature vector extracted and its spontaneous transposition progress bilinearity
Bilinearity eigenmatrix, and convert bilinearity feature vector for bilinearity eigenmatrix, be followed by finally by full articulamentum
Softmax layers carry out more classification to bilinearity feature vector and obtain probability of all categories.The mushroom classification being related in the present invention
Title has: Amanita vaginata var.vaginata, Xerocomus subtomentosus, Conocybe
Albipes, Cortinarius rubellus, Helvella crispa, Cuphophyllus flavipes, Hygrocybe
Reidii, Inocybe erubescens, Lyophyllum fumosum, Russula pectinatoides, Tricholoma
Fulvum, Tricholoma sciodes, Lycoperdon utriforme, Rhodocollybia butyracea
f.asema。
1, image input and pretreatment
The average value of entire data set can be subtracted first after the image data input network model of arbitrary size and divided by whole
The standard deviation of a data set carries out centralization and normalized, and its purpose is to allow near data zooming to 0 value and do not change
The distribution of parameter evidence reduces difference of the different samples when calculating gradient, accelerates the convergence of network.
It is 448 pixels that image, which can be scaled to short side, later, and is cut out from image using the mode of random cropping
The square-shaped image region of 448*448, last image can be by random flip horizontals.Random cropping and Random Level overturning etc. are pre-
Processing means are provided to increase the diversity of data set, and network model is made to have better Generalization Capability.Due to mushroom data
The characteristics of collection, the direction of growth of mushroom be from lower to upper, therefore only with flip horizontal rather than flip vertical.
Image color is indicated using tri- channels RGB, therefore passing through pretreated image data size is 448*448*
3, the subsequent data can be admitted to feature extraction network and be handled, and total preprocessing process is as shown in Figure 2.
2, feature extraction network
Feature extraction Web vector graphic based on transfer learning Yu bilinear fine granularity mushroom phenotype identification model
InceptionResNetV2 network is constituted.As shown in figure 3, being remerged by using multiple convolution kernels in 4 branch parallel processing
Structure (Bottleneck Layer), improve the width of network, increase network to the ability to accept of different size scale,
Thus the problems such as deep neural network parameter is too many, computation complexity is too big, gradient disperse is solved.Use fractionation convolution operation
Reasonable dimension decomposition is carried out, such dimension operation splitting can save in the case where not losing largely keeping minutia
About permitted multiparameter, reduces calculating consumption to accelerate the convergence rate of network, while further having deepened the depth of network
Degree, improves the non-linear of network.
The overall structure of InceptionResNetV2 feature extraction network as shown in figure 4, by reference to Microsoft residual error net
Network joined the design of residual block, enables parameter to skip some layers by the shortcut in some networks and is propagated, in this way
Design can solve deeper network structure Gradient disappearance the problems such as, thus make ultra deep network training become can
Can, under the network structure of deeper, better training effect can be obtained.First 7 layers of InceptionResNetV2 network are
It is made of three-layer coil lamination, one layer of maximum pond layer, two layers of convolutional layer, one layer of maximum pond layer, 10 tools is repeated later
There are three the residual error Inception modules of branch to have further through a better simply Inception module using 20 times
The residual error Inception module of Liang Ge branch finally has by 10 times further through the Inception module of 4 branches
The residual error Inception module of Liang Ge branch, then output result has been obtained by a convolutional layer.
The parameter of InceptionResNetV2 feature extraction networks major layer is as shown in table 1, wherein only listing first 7 layers
Convolution sum maximum pond layer and the merging layer of each residual error Inception module, convolutional layer, residual error layer and the last one convolution
Layer, then Batch Normalization layers and ReLU layers after each convolutional layer.Image from being into input layer dimension size
448*448*3 starts, and increases the depth of image by continuous convolution, and maximum pond layer halves the dimension of image, residual error
The dimension size that Inception module maintains image is constant, and every by a residual error Inception module, the length of image
Width reduces depth and increases, and dimension size when finally exporting is 12*12*1536, and Headquarters of the General Staff quantity is 54336736.
The parameter of 1 InceptionResNetV2 feature extraction networks major layer of table
3, bilinearity is converged and classifies
Bilinearity refers to that, for function f (x, y), as the one of parameter such as x of fixation, function f (x, y) is to another
A parameter y is linear.In the present invention, bilinear model B is made of four-tuple, as shown in formula (1),
B=(fA,fB,P,C) (7)
Wherein fAAnd fBIt is characteristic function, P is the pond function of model, and C is the classification function of mushroom.
Characteristic function f is that the effect of the feature extraction network in the present invention is that the picture of input and position are mapped to c × D
The feature of size, D refer to depth.The feature exported in the present invention by the feature on each position using matrix apposition combine and
Come, as shown in formula (2),
bilinear(L,I,fA,fB)=fA(L,I)TfB(L,I) (8)
Wherein L indicates that position and scale, I indicate picture.If the dimension difference for extracting feature of two characteristic functions
For (K, M) and (K, N), then after bilinear bilinearity converges operation, dimension becomes (M, N).If using summation Chi Hualai
The feature of comprehensive each position, then as shown in formula (3),
Wherein Φ (I) indicates that global picture feature indicates.
Bilinearity feature vector x=Φ (I) is finally passed through into symbol square root transformationAnd increase L2
RegularizationClassifier is inputted again obtains classification results to the end.
In the present invention, the characteristic length and width extracted by InceptionResNetV2 feature extraction network is equal
It is 12, depth 1536.Converging operation firstly the need of by three-dimensional feature vector reshape to feature vector progress bilinearity is
Two-dimensional feature vector obtains the feature vector of 144*1536.Next feature vector progress transposition is obtained into 1536*144 dimension
The feature vector for spending size carries out matrix apposition using the feature vector after former feature vector and transposition, i.e. bilinearity converges behaviour
Make, obtains the bilinearity feature vector that dimension size is 1536*1536.Bilinearity feature vector, which is shown laid flat in size, is
2359296 one-dimensional bilinearity feature vector, in addition symbol square root transformation and L2 regularization layer, then use full articulamentum,
More classification are carried out by softmax, the parameter amount of full articulamentum is 33030158.
Two, transfer learning
In the present invention, the transfer learning based on model has been used, the ImageNet with about 14,190,000 pictures is used
Data set includes many classifications in ImageNet data set, wherein there are plant, mushrooms etc. and mushroom of the present invention as source domain
The similar classification of goal task, by the way that the good Model Weight of pre-training on ImageNet data set is moved to mushroom of the invention
On data set, as shown in figure 5, not only reducing required data volume, moreover it is possible to obtain higher initial performance, faster training speed
Degree, better constringency performance.
Pre-training model is obtained from Keras pre-training model library, then load into based on transfer learning with it is bilinear
In fine granularity mushroom phenotype identification model, trained process is divided into two steps:
(1) what InceptionResNetV2 feature extraction network fixed first loaded obtains on ImageNet data set
Pre-training parameter, only allow the parameter of the last full articulamentum random initializtion of training.
(2) after network convergence, then the parameter of solid InceptionResNetV2 feature extraction network is solved, use is lesser
Learning rate is finely adjusted.
The reason of first step fixation InceptionResNetV2 network pre-training parameter, is that the full articulamentum of addition is
Random initializtion, biggish penalty values can be generated at the beginning and then generate biggish gradient, and it is good to be easily destroyed pre-training
Parameter, so to reuse lesser learning rate after the convergence of full articulamentum finely tunes entire model.
The pre-training model optimizer of transfer learning of the invention uses stochastic gradient descent (Stochastic
Gradient Descent, SGD) algorithm as optimizer,
Total training process is as shown in Figure 6, the specific steps are as follows:
(1) building is based on transfer learning and bilinear fine granularity mushroom phenotype identification model, wherein including
InceptionResNetV2 is as feature extraction network;
(2) ImageNet pre-training model initialization InceptionResNetV2 feature extraction network is used, is used
Glorot normal initialization device initializes full connection layer parameter;
(3) parameter of InceptionResNetV2 feature extraction network is fixed, the training process after being allowed to can not pass through
Backpropagation updates the parameter value of this part;
(4) training sample after obtaining image preprocessing in input channel, batch size are 8, and image size is
448*448;
(5) the batch training sample for getting (4) inputs network model, converges operation by feature extraction and bilinearity
And full articulamentum, probability of all categories is calculated finally by softmax;
(6) use classes cross entropy loss function calculates the penalty values of network model;
(7) by calculating gradient value, using SGD optimizer, it is 1.0 that initial learning rate, which is arranged, and learning rate decays to 1e-
8, momentum Momentum is set as 0.9, and error back propagation is returned whole network, updates the parameter of full articulamentum;
(8) judge whether to reach given number of iterations 100 or meet the 10 iteration variations of verifying penalty values to be no more than 0.001
Morning stop method condition, if, it is believed that network has been restrained, then enters step (9), if otherwise reentering step (4);
(9) change the learning rate of SGD optimizer to 0.001;
(10) fixation to InceptionResNetV2 feature extraction network pre-training parameter is released, network is led to
Cross the parameter value that backpropagation updates this part;
(11) training sample after obtaining image preprocessing in input channel, batch size are 8, and image size is
448*448;
(12) the batch training sample for getting (11) inputs network model, converges behaviour by feature extraction and bilinearity
Work and full articulamentum, probability of all categories is calculated finally by softmax;
(13) use classes cross entropy loss function calculates the penalty values of network model;
(14) by calculating gradient value, using SGD optimizer, it is 0.001 that initial learning rate, which is arranged, and learning rate decays to
1e-8, momentum Momentum are set as 0.9, and error back propagation is returned whole network, updates each layer of network of parameter;
(15) judge whether to reach given number of iterations 70 or meet the 10 iteration variations of verifying penalty values to be no more than 0.001
Morning stop method condition, if, it is believed that network has been restrained, then enters step (16), if otherwise reentering step (11);
(16) accuracy rate, accurate rate, recall rate, the F1 value of network model are calculated by test set.
The basic principles, main features and advantages of the invention have been shown and described above.Those skilled in the art
It should be appreciated that the protection scope that the above embodiments do not limit the invention in any form, all to be obtained using modes such as equivalent replacements
The technical solution obtained, falls in protection scope of the present invention.
Part that the present invention does not relate to is the same as those in the prior art or can be realized by using the prior art.
Claims (8)
1. a kind of fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2, special
Sign be the following steps are included:
Step 1 is established based on transfer learning and bilinear fine granularity mushroom phenotype identification model;
Step 2 carries out transfer learning and training based on identification model;
Step 3 pre-processes after image is inputted identification model;
Step 4 carries out feature extraction to pretreated image data;It is special using the InceptionResNetV2 of symmetrical structure
Sign extracts the feature vector in network extraction image, then carries out bilinearity to the feature vector extracted and its spontaneous transposition
Converge operation and obtain the bilinearity eigenmatrix of each position of picture, and by bilinearity eigenmatrix be converted into bilinearity feature to
Amount is followed by softmax layers finally by full articulamentum and obtains probability of all categories to the more classification of bilinearity feature vector progress.
2. the fine granularity mushroom table according to claim 1 based on transfer learning Yu bilinearity InceptionResNetV2
Type recognition methods, which is characterized in that the pretreatment in the step 3 include centralization, normalization, scaling, random cropping and with
Machine flip horizontal.
3. the fine granularity mushroom table according to claim 2 based on transfer learning Yu bilinearity InceptionResNetV2
Type recognition methods, which is characterized in that after the image data input Network Recognition model of arbitrary size, subtract entire data set first
Average value and carry out centralization and normalized divided by the standard deviation of entire data set, image is scaled to short side and is later
448 pixels, and the square-shaped image region of 448*448 is cut out using the mode of random cropping from image, it is last random
Flip horizontal is carried out to image.
4. the fine granularity mushroom table according to claim 1 based on transfer learning Yu bilinearity InceptionResNetV2
Type recognition methods, which is characterized in that using in Inception series of network model in the step 4
InceptionResNetV2 network carries out feature extraction, and joined in InceptionResNetV2 feature extraction network residual
Poor block.
5. the fine granularity mushroom table according to claim 4 based on transfer learning Yu bilinearity InceptionResNetV2
Type recognition methods, which is characterized in that first 7 layers of the InceptionResNetV2 network are by three-layer coil lamination, one layer of maximum pond
Change layer, two layers of convolutional layer, one layer of maximum pond layer composition, being repeated 10 times tool later, there are three the residual error Inception moulds of branch
Block, further through a better simply Inception module, the residual error Inception module having using 20 times there are two branch,
Further through the Inception module of 4 branches, the residual error Inception module finally having by 10 times there are two branch,
Output result is obtained by a convolutional layer again.
6. the fine granularity mushroom table according to claim 1 based on transfer learning Yu bilinearity InceptionResNetV2
Type recognition methods, which is characterized in that bilinear model B is made of four-tuple, as shown in formula (1),
B=(fA,fB,P,C) (1)
Wherein fAAnd fBIt is characteristic function, P is the pond function of model, and C is the classification function of mushroom;
Feature of the feature of output on each position using matrix apposition combination from, as shown in formula (2),
bilinear(L,I,fA,fB)=fA(L,I)TfB(L,I) (2)
Wherein L indicates that position and scale, I indicate picture;If the dimension for extracting feature of two characteristic functions be respectively (K,
M) and (K, N), then after bilinear bilinearity converges operation, dimension becomes (M, N), if comprehensive each using summation Chi Hualai
The feature of a position, then as shown in formula (3),
Wherein Φ (I) indicates that global picture feature indicates;
Bilinearity feature vector x=Φ (I) is finally passed through into symbol square root transformationAnd increase L2 canonical
ChangeClassifier is inputted again obtains classification results to the end.
7. the fine granularity mushroom table according to claim 1 based on transfer learning Yu bilinearity InceptionResNetV2
Type recognition methods, which is characterized in that training process is divided into two steps:
(1) what InceptionResNetV2 feature extraction network fixed first loaded obtains pre- on ImageNet data set
Training parameter, the parameter for the full articulamentum random initializtion for only allowing training last;
(2) after network convergence, then the parameter of solid InceptionResNetV2 feature extraction network is solved, uses lesser study
Rate is finely adjusted.
8. the fine granularity mushroom according to claim 1 or claim 7 based on transfer learning Yu bilinearity InceptionResNetV2
Phenotype recognition methods, which is characterized in that total training process is as follows:
(1) building is based on transfer learning and bilinear fine granularity mushroom phenotype identification model, wherein including
InceptionResNetV2 is as feature extraction network;
(2) ImageNet pre-training model initialization InceptionResNetV2 feature extraction network is used, just using Glorot
Normal initializer initializes full connection layer parameter;
(3) parameter of InceptionResNetV2 feature extraction network is fixed, the training process after being allowed to can not be by reversed
Propagate the parameter value for updating this part;
(4) training sample after obtaining image preprocessing in input channel, batch size are 8, and image size is 448*
448;
(5) the batch training sample got (4) inputs network model, by feature extraction and bilinearity converge operation and
Full articulamentum calculates probability of all categories finally by softmax;
(6) use classes cross entropy loss function calculates the penalty values of network model;
(7) by calculating gradient value, using SGD optimizer, it is 1.0 that initial learning rate, which is arranged, and learning rate decays to 1e-8, is moved
Amount Momentum is set as 0.9, and error back propagation is returned whole network, updates the parameter of full articulamentum;
(8) judge whether to reach given number of iterations 100 or meet the morning that the 10 iteration variations of verifying penalty values are no more than 0.001
Stop method condition, if, it is believed that network has been restrained, then enters step (9), if otherwise reentering step (4);
(9) change the learning rate of SGD optimizer to 0.001;
(10) fixation to InceptionResNetV2 feature extraction network pre-training parameter is released, passes through network anti-
To the parameter value for propagating this part of update;
(11) training sample after obtaining image preprocessing in input channel, batch size are 8, and image size is 448*
448;
(12) the batch training sample got (11) inputs network model, by feature extraction and bilinearity converge operation with
And full articulamentum, probability of all categories is calculated finally by softmax;
(13) use classes cross entropy loss function calculates the penalty values of network model;
(14) by calculating gradient value, using SGD optimizer, it is 0.001 that initial learning rate, which is arranged, and learning rate decays to 1e-
8, momentum Momentum is set as 0.9, and error back propagation is returned whole network, updates each layer of network of parameter;
(15) judge whether to reach given number of iterations 70 or meet the morning that the 10 iteration variations of verifying penalty values are no more than 0.001
Stop method condition, if, it is believed that network has been restrained, then enters step (16), if otherwise reentering step (11);
(16) accuracy rate, accurate rate, recall rate, the F1 value of network model are calculated by test set.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910547744.8A CN110263863B (en) | 2019-06-24 | 2019-06-24 | Fine-grained fungus phenotype identification method based on transfer learning and bilinear InceptionResNet V2 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910547744.8A CN110263863B (en) | 2019-06-24 | 2019-06-24 | Fine-grained fungus phenotype identification method based on transfer learning and bilinear InceptionResNet V2 |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110263863A true CN110263863A (en) | 2019-09-20 |
CN110263863B CN110263863B (en) | 2021-09-10 |
Family
ID=67920752
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910547744.8A Active CN110263863B (en) | 2019-06-24 | 2019-06-24 | Fine-grained fungus phenotype identification method based on transfer learning and bilinear InceptionResNet V2 |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110263863B (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110781921A (en) * | 2019-09-25 | 2020-02-11 | 浙江农林大学 | Depth residual error network and transfer learning-based muscarinic image identification method and device |
CN111401122A (en) * | 2019-12-27 | 2020-07-10 | 航天信息股份有限公司 | Knowledge classification-based complex target asymptotic identification method and device |
CN111462068A (en) * | 2020-03-30 | 2020-07-28 | 华南理工大学 | Bolt and nut detection method based on transfer learning |
CN111613287A (en) * | 2020-03-31 | 2020-09-01 | 武汉金域医学检验所有限公司 | Report coding model generation method, system and equipment based on Glow network |
CN111666984A (en) * | 2020-05-20 | 2020-09-15 | 海南电网有限责任公司电力科学研究院 | Intelligent overvoltage identification method based on transfer learning |
CN111709326A (en) * | 2020-05-29 | 2020-09-25 | 安徽艾睿思智能科技有限公司 | Deep learning-based cascaded granular crop quality analysis method and system |
CN111860601A (en) * | 2020-06-22 | 2020-10-30 | 北京林业大学 | Method and device for predicting large fungus species |
CN111914802A (en) * | 2020-08-17 | 2020-11-10 | 中国电波传播研究所(中国电子科技集团公司第二十二研究所) | Ionosphere return scattering propagation pattern identification method based on transfer learning |
CN112052904A (en) * | 2020-09-09 | 2020-12-08 | 陕西理工大学 | Method for identifying plant diseases and insect pests based on transfer learning and convolutional neural network |
CN112364899A (en) * | 2020-10-27 | 2021-02-12 | 西安科技大学 | Abrasive grain ferrographic image intelligent identification method based on virtual image and transfer learning |
CN112633370A (en) * | 2020-12-22 | 2021-04-09 | 中国医学科学院北京协和医院 | Detection method, device, equipment and medium for filamentous fungus morphology |
WO2021115123A1 (en) * | 2019-12-12 | 2021-06-17 | 苏州科技大学 | Method for footprint image retrieval |
CN113128593A (en) * | 2021-04-20 | 2021-07-16 | 南京林业大学 | Plant fine-grained identification method based on bilinear convolutional neural network |
CN113642518A (en) * | 2021-08-31 | 2021-11-12 | 山东省计算中心(国家超级计算济南中心) | Cell membrane coloring integrity judging method for her2 pathological image based on transfer learning |
CN114722928A (en) * | 2022-03-29 | 2022-07-08 | 河海大学 | Blue-green algae image identification method based on deep learning |
CN111027626B (en) * | 2019-12-11 | 2023-04-07 | 西安电子科技大学 | Flow field identification method based on deformable convolution network |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108509976A (en) * | 2018-02-12 | 2018-09-07 | 北京佳格天地科技有限公司 | The identification device and method of animal |
US20180260655A1 (en) * | 2014-11-07 | 2018-09-13 | Adobe Systems Incorporated | Local feature representation for image recognition |
CN109086836A (en) * | 2018-09-03 | 2018-12-25 | 淮阴工学院 | A kind of automatic screening device of cancer of the esophagus pathological image and its discriminating method based on convolutional neural networks |
CN109389245A (en) * | 2018-09-06 | 2019-02-26 | 浙江鸿程计算机系统有限公司 | A kind of multifactor fusion school district school age population prediction technique based on deep neural network |
CN109508655A (en) * | 2018-10-28 | 2019-03-22 | 北京化工大学 | The SAR target identification method of incomplete training set based on twin network |
CN109508650A (en) * | 2018-10-23 | 2019-03-22 | 浙江农林大学 | A kind of wood recognition method based on transfer learning |
CN109635712A (en) * | 2018-12-07 | 2019-04-16 | 杭州电子科技大学 | Spontaneous micro- expression type method of discrimination based on homogeneous network |
-
2019
- 2019-06-24 CN CN201910547744.8A patent/CN110263863B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180260655A1 (en) * | 2014-11-07 | 2018-09-13 | Adobe Systems Incorporated | Local feature representation for image recognition |
CN108509976A (en) * | 2018-02-12 | 2018-09-07 | 北京佳格天地科技有限公司 | The identification device and method of animal |
CN109086836A (en) * | 2018-09-03 | 2018-12-25 | 淮阴工学院 | A kind of automatic screening device of cancer of the esophagus pathological image and its discriminating method based on convolutional neural networks |
CN109389245A (en) * | 2018-09-06 | 2019-02-26 | 浙江鸿程计算机系统有限公司 | A kind of multifactor fusion school district school age population prediction technique based on deep neural network |
CN109508650A (en) * | 2018-10-23 | 2019-03-22 | 浙江农林大学 | A kind of wood recognition method based on transfer learning |
CN109508655A (en) * | 2018-10-28 | 2019-03-22 | 北京化工大学 | The SAR target identification method of incomplete training set based on twin network |
CN109635712A (en) * | 2018-12-07 | 2019-04-16 | 杭州电子科技大学 | Spontaneous micro- expression type method of discrimination based on homogeneous network |
Non-Patent Citations (1)
Title |
---|
陈鹤森: ""基于深度学习的细粒度图像识别研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (26)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110781921A (en) * | 2019-09-25 | 2020-02-11 | 浙江农林大学 | Depth residual error network and transfer learning-based muscarinic image identification method and device |
CN111027626B (en) * | 2019-12-11 | 2023-04-07 | 西安电子科技大学 | Flow field identification method based on deformable convolution network |
WO2021115123A1 (en) * | 2019-12-12 | 2021-06-17 | 苏州科技大学 | Method for footprint image retrieval |
US11809485B2 (en) | 2019-12-12 | 2023-11-07 | Suzhou University of Science and Technology | Method for retrieving footprint images |
CN111401122B (en) * | 2019-12-27 | 2023-09-26 | 航天信息股份有限公司 | Knowledge classification-based complex target asymptotic identification method and device |
CN111401122A (en) * | 2019-12-27 | 2020-07-10 | 航天信息股份有限公司 | Knowledge classification-based complex target asymptotic identification method and device |
CN111462068A (en) * | 2020-03-30 | 2020-07-28 | 华南理工大学 | Bolt and nut detection method based on transfer learning |
CN111462068B (en) * | 2020-03-30 | 2023-03-21 | 华南理工大学 | Bolt and nut detection method based on transfer learning |
CN111613287A (en) * | 2020-03-31 | 2020-09-01 | 武汉金域医学检验所有限公司 | Report coding model generation method, system and equipment based on Glow network |
CN111613287B (en) * | 2020-03-31 | 2023-08-04 | 武汉金域医学检验所有限公司 | Report coding model generation method, system and equipment based on Glow network |
CN111666984B (en) * | 2020-05-20 | 2023-08-25 | 海南电网有限责任公司电力科学研究院 | Overvoltage intelligent identification method based on transfer learning |
CN111666984A (en) * | 2020-05-20 | 2020-09-15 | 海南电网有限责任公司电力科学研究院 | Intelligent overvoltage identification method based on transfer learning |
CN111709326A (en) * | 2020-05-29 | 2020-09-25 | 安徽艾睿思智能科技有限公司 | Deep learning-based cascaded granular crop quality analysis method and system |
CN111709326B (en) * | 2020-05-29 | 2023-05-12 | 安徽省科亿信息科技有限公司 | Cascade type particle crop quality analysis method and system based on deep learning |
CN111860601A (en) * | 2020-06-22 | 2020-10-30 | 北京林业大学 | Method and device for predicting large fungus species |
CN111860601B (en) * | 2020-06-22 | 2023-10-17 | 北京林业大学 | Method and device for predicting type of large fungi |
CN111914802B (en) * | 2020-08-17 | 2023-02-07 | 中国电波传播研究所(中国电子科技集团公司第二十二研究所) | Ionosphere return scattering propagation pattern identification method based on transfer learning |
CN111914802A (en) * | 2020-08-17 | 2020-11-10 | 中国电波传播研究所(中国电子科技集团公司第二十二研究所) | Ionosphere return scattering propagation pattern identification method based on transfer learning |
CN112052904A (en) * | 2020-09-09 | 2020-12-08 | 陕西理工大学 | Method for identifying plant diseases and insect pests based on transfer learning and convolutional neural network |
CN112364899A (en) * | 2020-10-27 | 2021-02-12 | 西安科技大学 | Abrasive grain ferrographic image intelligent identification method based on virtual image and transfer learning |
CN112633370A (en) * | 2020-12-22 | 2021-04-09 | 中国医学科学院北京协和医院 | Detection method, device, equipment and medium for filamentous fungus morphology |
CN113128593A (en) * | 2021-04-20 | 2021-07-16 | 南京林业大学 | Plant fine-grained identification method based on bilinear convolutional neural network |
CN113642518A (en) * | 2021-08-31 | 2021-11-12 | 山东省计算中心(国家超级计算济南中心) | Cell membrane coloring integrity judging method for her2 pathological image based on transfer learning |
CN113642518B (en) * | 2021-08-31 | 2023-08-22 | 山东省计算中心(国家超级计算济南中心) | Transfer learning-based her2 pathological image cell membrane coloring integrity judging method |
CN114722928A (en) * | 2022-03-29 | 2022-07-08 | 河海大学 | Blue-green algae image identification method based on deep learning |
CN114722928B (en) * | 2022-03-29 | 2024-04-16 | 河海大学 | Blue algae image recognition method based on deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN110263863B (en) | 2021-09-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110263863A (en) | Fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2 | |
CN103927531B (en) | It is a kind of based on local binary and the face identification method of particle group optimizing BP neural network | |
CN104462494B (en) | A kind of remote sensing image retrieval method and system based on unsupervised feature learning | |
CN109934761A (en) | Jpeg image steganalysis method based on convolutional neural networks | |
CN110084221A (en) | A kind of serializing face critical point detection method of the tape relay supervision based on deep learning | |
CN111696101A (en) | Light-weight solanaceae disease identification method based on SE-Inception | |
CN106778918A (en) | A kind of deep learning image identification system and implementation method for being applied to mobile phone terminal | |
CN111339818B (en) | Face multi-attribute recognition system | |
CN110363253A (en) | A kind of Surfaces of Hot Rolled Strip defect classification method based on convolutional neural networks | |
Ding et al. | Research on daily objects detection based on deep neural network | |
Hara et al. | Towards good practice for action recognition with spatiotemporal 3d convolutions | |
CN106339753A (en) | Method for effectively enhancing robustness of convolutional neural network | |
CN109783887A (en) | A kind of intelligent recognition and search method towards Three-dimension process feature | |
CN107679572A (en) | A kind of image discriminating method, storage device and mobile terminal | |
CN110490298A (en) | Lightweight depth convolutional neural networks model based on expansion convolution | |
CN108205703A (en) | Multi-input multi-output matrix average value pooling vectorization implementation method | |
CN109165733A (en) | Multi-input multi-output matrix maximum pooling vectorization implementation method | |
CN112906747A (en) | Knowledge distillation-based image classification method | |
CN110674326A (en) | Neural network structure retrieval method based on polynomial distribution learning | |
CN108182316A (en) | A kind of Electromagnetic Simulation method and its electromagnetism brain based on artificial intelligence | |
CN111282281B (en) | Image processing method and device, electronic equipment and computer readable storage medium | |
CN110046568A (en) | A kind of video actions recognition methods based on Time Perception structure | |
Zhang et al. | Summary of convolutional neural network compression technology | |
Ma et al. | A survey of sparse-learning methods for deep neural networks | |
CN110188621A (en) | A kind of three-dimensional face expression recognition methods based on SSF-IL-CNN |
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 |