CN108197666A - A kind of processing method, device and the storage medium of image classification model - Google Patents
A kind of processing method, device and the storage medium of image classification model Download PDFInfo
- Publication number
- CN108197666A CN108197666A CN201810087876.2A CN201810087876A CN108197666A CN 108197666 A CN108197666 A CN 108197666A CN 201810087876 A CN201810087876 A CN 201810087876A CN 108197666 A CN108197666 A CN 108197666A
- Authority
- CN
- China
- Prior art keywords
- model
- image
- training
- training network
- parameter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of processing method of image classification model, including：Initialisation image disaggregated model；Determine sequentially to choose from image classification model the first of filter model treats training network and determines that including the filter model of sequence selection and the second of sorter model treats training network；Markd image pattern is not had according to acquisition, treats that training network carries out the training of unsupervised mode to first, the parameter of training network median filter model is treated with update first；Treat that training network extracts characteristics of image from image pattern based on second；Based on the characteristics of image extracted and the label of respective image sample, treat that the sorter model in training network carries out the training for having monitor mode to second；Based on the error of the sorter model output result after training, the parameter of training network median filter model is treated in update second.The present invention further simultaneously discloses a kind of processing unit and storage medium of image classification model.
Description
Technical field
The present invention relates to the image processing techniques in computer realm more particularly to a kind of processing sides of image classification model
Method, device and storage medium.
Background technology
With the fast development of electronic technology and internet particularly mobile Internet, electronic equipment particularly intelligent mobile
The function of terminal is stronger and stronger, and user can install various application programs according to self-demand on intelligent mobile terminal, with
Complete various affairs.For example, realize that image identifies by the application program being mounted on intelligent mobile terminal.
At present, in order to which image is identified, to pick out the classification belonging to image, usually by depth in the relevant technologies
It practises algorithm such as depth belief network (DBN, Deep Belief Network) to apply in image identification, that is, utilizes in neuron
Between the weight that generates so that neural network can generate training data with maximum probability.However, in the training process of DBN
In, when back-propagation algorithm is used constantly to adjust the parameters in DBN model, due in reverse propagated error
In the process, error can constantly tiring out with limited Boltzmann machine (RBM, Restricted Boltzmann Machine) number of plies
It accumulates and the decline of exponential form occurs, that is, the phenomenon that gradient disperse occur, reduced so as to cause image recognition accuracy so that mould
The effect of type training study is unable to reach expected results.
Therefore, how quickly for the accuracy rate of raising image identification, the relevant technologies there is no effective solution.
Invention content
In view of this, an embodiment of the present invention is intended to provide a kind of processing method of image classification model, device and storages to be situated between
Matter, to solve the problems, such as that the relevant technologies are difficult to the effectively quick accuracy rate for improving image identification.
In order to achieve the above objectives, the technical solution of the embodiment of the present invention is realized in：
In a first aspect, the embodiment of the present invention provides a kind of processing method of image classification model, the method includes：
Initialisation image disaggregated model；
Determine sequentially to choose from described image disaggregated model the first of filter model treats training network and determines
The filter model chosen including the sequence and the second of sorter model treats training network；
Markd image pattern is not had according to acquisition, treats that training network carries out the instruction of unsupervised mode to described first
Practice, to update the described first parameter for treating training network median filter model；
Treat that training network extracts characteristics of image from described image sample based on described second；
Based on the characteristics of image extracted and the label of respective image sample, the classification in training network is treated to described second
Device model carries out the training for having monitor mode；
Based on the error of the sorter model output result after training, training network median filter mould is treated in update described second
The parameter of type.
Second aspect, the embodiment of the present invention provide a kind of processing unit of image classification model, and described device includes：Initially
Change module, determining module, the first training module, extraction module, the second training module and update module；Wherein,
The initialization module, for initialisation image disaggregated model；
The determining module, for determining that sequentially chosen from described image disaggregated model the first of filter model treats
Training network and determine that including the filter model of sequence selection and the second of sorter model treats training network；
First training module for not having markd image pattern according to acquisition, is waited to train to described first
Network carries out the training of unsupervised mode, to update the described first parameter for treating training network median filter model；
The extraction module treats that training network extracts characteristics of image from described image sample for being based on described second；
Second training module, for based on the characteristics of image extracted and the label of respective image sample, to described
Second treats that the sorter model in training network carries out the training for having monitor mode；
The update module, for the error based on the sorter model output result after training, update described second is treated
The parameter of training network median filter model.
The third aspect, the embodiment of the present invention provide a kind of storage medium, are stored thereon with executable program, described executable
The step of processing method of image classification model provided in an embodiment of the present invention is realized when program is executed by processor.
Fourth aspect, the embodiment of the present invention also provide a kind of processing unit of image classification model, including memory, processing
On a memory and the executable program that can be run by the processor, the processor operation is described to be can perform for device and storage
The step of processing method of image classification model provided in an embodiment of the present invention is performed during program.
The processing method of image classification model, device and the storage medium that the embodiment of the present invention is provided, by from image
Determine that first treats that training network and second treats training network in disaggregated model, according to not having markd image pattern, to first
It treats that training network carries out unsupervised mode training, the parameter of training network median filter model is treated with update first；Based on extraction
Characteristics of image and respective image sample label, treat that training network has carried out monitor mode training to second, to update second
Treat the parameter of training network median filter model.In this way, unsupervised learning mode and supervised learning mode, which are alternated, to be made
With constantly being adjusted to the parameters in image classification model so that the convergence rate of image classification model faster, is more easy to
Reach plateau, reduce the time of model training study；Meanwhile the embodiment of the present invention uses the image classification mould of above-mentioned structure
Type carries out image identification to image data to be identified, can avoid the occurrence of asking because of the gradient disperse that back-propagation algorithm is brought
Topic greatly improves the accuracy rate of image identification, promotes user experience.
Description of the drawings
Fig. 1, which is that one of the processing method of image classification model provided in an embodiment of the present invention is optional, realizes flow signal
Figure；
Fig. 2 is that another optional realization flow of the processing method of image classification model provided in an embodiment of the present invention is shown
It is intended to；
Fig. 3 is convolution provided in an embodiment of the present invention restricted Boltzmann machine (CRBM, Convolution
Restricted Boltzmann Machine) an optional structure diagram；
Fig. 4 A are a provided in an embodiment of the present invention first optional structure diagram for treating training network；
Fig. 4 B are a provided in an embodiment of the present invention second optional structure diagram for treating training network；
Fig. 5 A are provided in an embodiment of the present invention first another optional structure diagram for treating training network；
Fig. 5 B are provided in an embodiment of the present invention second another optional structure diagram for treating training network；
Fig. 6 is an optional functional structure signal of the processing unit of image classification model provided in an embodiment of the present invention
Figure；
Fig. 7 is that another optional functional structure of the processing unit of image classification model provided in an embodiment of the present invention is shown
It is intended to；
Fig. 8 is an optional hardware configuration signal of the processing unit of image classification model provided in an embodiment of the present invention
Figure.
Specific embodiment
The characteristics of in order to more fully hereinafter understand the embodiment of the present invention and technology contents, below in conjunction with the accompanying drawings to this hair
The realization of bright embodiment is described in detail, appended attached drawing purposes of discussion only for reference, is not used for limiting the present invention.
Fig. 1, which is that one of the processing method of image classification model provided in an embodiment of the present invention is optional, realizes flow signal
Figure, the processing method of image classification model can be applied in server or terminal device；As shown in Figure 1, in the embodiment of the present invention
Image classification model processing method realization flow, may comprise steps of：
Step 101：Initialisation image disaggregated model.
Step 102：Determine sequentially chosen from image classification model the first of filter model treat training network and
Determine that including the filter model of sequence selection and the second of sorter model treats training network.
Step 103：Markd image pattern is not had according to acquisition, treats that training network carries out unsupervised mode to first
Training, the parameter that training network median filter model is treated with update first.
Step 104：Treat that training network extracts characteristics of image from image pattern based on second.
Step 105：Based on the characteristics of image extracted and the label of respective image sample, treated in training network to second
Sorter model carries out the training for having monitor mode.
Step 106：Based on the error of the sorter model output result after training, update second is treated to filter in training network
The parameter of device model.
In the present embodiment, image classification model is a kind of mixed model, which has a certain number of filters
The model structure that wave device model and sorter model combine.Of filter model can be set according to actual conditions
Number, i.e., be set as one or more by the number of filter model.It, can will be multiple when the number of filter model is multiple
Filter model is linked in sequence according to the preset order of connection.
During being trained using the method for above-mentioned steps 101 to step 106 to image classification model, it is combined with
The mode that supervised learning and unsupervised learning carry out in turn carrys out training image disaggregated model network namely to from image classification model
In determine first treat that training network carries out the training of unsupervised mode, second determined from image classification model is waited to train
Network carries out the training for having monitor mode, treats that training network and second treats training network median filter model to reach update first
Parameter purpose.Since the parameters in image classification model are trained to splendid state so that the image classification
Model has good image classification ability, therefore, this image classification model is applied in the image data base of acquisition, can be with
Image data generic is efficiently identified out, improves the accuracy rate of image identification.
It should be noted that in view of the duration and training effect of model training, it can not possibly be unlimited in practical application
Repetition training image classification model, therefore, usually setting first treats that training network and second treats the training network training number of plies
Respective threshold sets first to treat that the training number of plies threshold value of training network and second treats the training number of plies threshold value of training network, and
First treats that the training number of plies of training network and second treats that the training number of plies of training network is identical.That is, it waits to instruct to first
After the completion of practicing the wheel training that network carries out unsupervised mode, then treat that training network carries out the wheel for having monitor mode to second
Training after the completion of each round training, all detects whether the completed trained number of plies reaches pre-set trained number of plies threshold value,
When detecting the trained number of plies not up to training number of plies threshold value, then the training number of plies is added 1, and continue initialisation image disaggregated model
Model training is carried out, until when detecting that the trained number of plies reaches trained number of plies threshold value, then can terminate model training process.
In practical applications, training number of plies threshold value can be adjusted accordingly, but training number of plies threshold value according to actual conditions
It can not be too big.Usually training number of plies threshold value can be set as 5 layers or 6 layers, at this point, the parameters base in image classification model
This is trained to splendid state.
The specific implementation process of the processing method of image classification model of the embodiment of the present invention is done below further in detail
Explanation.
Fig. 2 is that another optional realization flow of the processing method of image classification model provided in an embodiment of the present invention is shown
It is intended to, the processing method of image classification model can be applied in server or terminal device；As shown in Fig. 2, the embodiment of the present invention
In image classification model processing method realization flow, may comprise steps of：
Step 201：Initialisation image disaggregated model.
In the present embodiment, image classification model can include filter model and sorter model；Wherein, according to reality
Border situation can set the number of filter model, that is, the number for setting filter model is one or more.That is,
Image classification model in the present embodiment is a kind of mixed model, which has a certain number of filter models,
And the model structure that sorter model combines.
It should be noted that when setting filter model number for it is multiple when, can by multiple filter models according to
The preset order of connection is linked in sequence.
Here, initialisation image disaggregated model, it can be understood as the wave filter mould that initialisation image disaggregated model includes
The parameter of sorter model that the parameter and initialisation image disaggregated model of type include.Wherein, the parameter of filter model
Can be weights or bias, the parameter of sorter model can be for characterizing probability value of each classification results appearance etc., originally
Inventive embodiments are not specifically limited herein.
The server of the processing method application of the image classification model of the present embodiment can be remote server or high in the clouds clothes
Business device, the terminal device of the processing method application of the image classification model can be intelligent electronic device, as a kind of preferable
Embodiment, intelligent electronic device can be smart mobile phone or tablet computer.
Step 202：Determine sequentially chosen from image classification model the first of filter model treat training network and
Determine that including the filter model of sequence selection and the second of sorter model treats training network.
In alternative embodiment of the present invention, following manner realization may be used in this step 202：It is suitable from image classification model
Sequence chooses target filter model to be trained, and selected target filter model is first in image classification model
During filter model,
The input layer and hidden layer for choosing target filter model treat training network as first；
Input layer, hidden layer and the sorter model for choosing target filter model treat training network as second.
In another alternative embodiment of the present invention, following manner realization may be used in this step 202：
Target filter model to be trained, and selected target filter mould are sequentially chosen from image classification model
When type is not first filter model in image classification model,
Choose the previous filtering of input layer, hidden layer and target filter the model connection of target filter model
The pond layer of device model treats training network as first；
Choose the input layer, hidden layer, the previous wave filter being connect with target filter model of target filter model
The pond layer and sorter model of model treat training network as second.
Below using filter model to be formed convolution depth belief network (CDBN, Convolution Deep Belief
Network for constituent element CRBM), by sorter model for for Softmax graders, training network is treated to first
Structure and second treats that the structure of training network illustrates respectively.
It is first right below before the structure and second that training network is treated to first treat that the structure of training network illustrates
The composition structure of CRBM is simply introduced.
Fig. 3 is an optional structure diagram of CRBM provided in an embodiment of the present invention, as shown in figure 3, a CRBM
Structure mainly include three parts, from the point of view of the direction according to the stream compression that data output is input to by data, this three parts point
It Wei not input layer V, hidden layer H and pond layer P.And the structure of the RBM in the relevant technologies only includes input layer V and hidden layer H, and
Other layers are not present behind hidden layer H.However, as can be seen from Fig. 3, the difference lies in CRBM's is implicit by CRBM and RBM
One layer, i.e. pond layer P are newly added to behind layer H again.In addition, CRBM and RBM are also differ in that, CRBM is in input layer V
Be identical for the weights on whole positions of data-oriented on hidden layer H.In practical application, although not providing input figure
As having same size or input picture to must satisfy two-dimensional structure, but for the ease of calculating, it is assumed that enable input picture
Size is Nv×Nv, therefore, size Nv×NvBinary cell matrix constitute in entire CRBM network structures
Input layer, wherein, NvRepresent the width and height of the image that are inputted in input layer.The general of " group " is defined on hidden layer
It reads, is N by sizeH×NHBinary cell matrix be defined as a group, K NH×NHGroup, which is combined together, to be built into
Hidden layer, therefore, NH 2K binary cell constitutes the hidden layer in entire CRBM network structures, N hereHRepresent implicit
The width and height of image inputted in layer, K represent the number of the group defined in hidden layer.The wave filter used in hidden layer
Number NWFor NW=NV-NH+ 1, filter size size is NW×NW, wherein, a wave filter is connected with a group.At K
In group, the weights of wave filter are equal in entire implicit layer unit.For bias, on hidden layer, a group
A corresponding bias bk, and for input layer, the bias c of entire input layer is shared.
Here, equation below (1) may be used to represent in the energy function of CRBM：
Wherein, E (v, h) represents the energy value of input layer and implicit interlayer, WkRepresent hidden layer with inputting the K in layer unit
The weights of a convolutional filtering, k represent the quantity of weights, and v represents the weights of input layer, bkRepresent what is shared in hidden layer node
The bias of hidden layer, hk i,jRepresent the unit of kth layer in hidden layer.
Using the core algorithm of gibbs sampler, when known to the state of input layer, state profit that can be based on input layer
With the conditional probability shown in equation below (2), probability when implicit layer state is 1 is obtained：
p(hk ij=1 | v)=σ ((Wk*v)ij+bk) (2)
Wherein, p (hk ij=1 | v) represent the conditional probability when implying layer state under inputting layer state known case and being 1,
WkRepresent that hidden layer represents the quantity of weights with inputting the weights of K convolutional filtering in layer unit, k, v represents the power of input layer
Value, bkRepresent the bias for the hidden layer shared in hidden layer node.
Using reconstructing mode, when known to the state of hidden layer, can the state based on hidden layer utilize equation below (3)
Probability when input layer state is 1 is obtained in shown conditional probability：
Wherein, p (vij=1 | h) represent conditional probability when input layer state is 1 under implicit layer state known case, Wk
Represent that hidden layer represents the quantity of weights with inputting the weights of K convolutional filtering in layer unit, k, c is represented in input layer
In share input layer bias.
It can obtain through the above, three parameters are mainly included in CRBM models：1) hidden layer and the K in input layer unit
The weights W of a convolutional filteringk, weights quantity is k=1 ... ..K, wherein, a wave filter can cover Nw×NwA image slices
Element；2) the bias c for the input layer shared in input layer；3) bias for the hidden layer shared in hidden layer node
bk。
Since natural image possesses specific constant attribute, the characteristic attribute of some position and remaining position in natural image
The characteristic attribute put is identical；That is, if the characteristic attribute of some part or position to image carries out
Statistics, then, this feature attribute is equally applicable to the other positions of this image, therefore, it is possible to pass through duplicate feature category
Any position of inquiry learning whole image.Assuming that there are one the natural image that size is 32 × 32, appoint from this image
The parts of images that meaning interception size is 7 × 7 to the sample data of this block 7 × 7 obtain after feature learning as sample data
The feature taken may be used for carrying out feature learning to whole image.The feature obtained from the data block that this size is 7 × 7
Attribute carries out convolution operation, so as to from all positions of whole image with natural image that original size is 32 × 32
Obtain the activation value of different characteristic.
Assuming that there are one the image that size is r × c, and the convolution for thering is k size to be a × b (a≤r, b≤c)
Core, carries out convolution operation with these convolution kernels and this image respectively, then obtains a feature squares of k × (r-a+1) × (c-b+1)
Battle array, wherein, the value that the sum of product addition of respective element obtains between image and convolution kernel is that convolution operation obtains later
Eigenmatrix each element value.If training grader uses the spy being characterized in by being obtained after convolution operation
Sign, then calculation amount undoubtedly can be very big, and it is possible to which there is a situation where over-fittings, this is actually infeasible.
To solve the above problems, introducing the concept in pond in the structure of CRBM, pond is by parts all in image
Feature gather together, using this characteristic in pond, the feature obtained after convolution is split, is divided into several rulers
The very little fraction for m × n, and maximum value or the average value of these fractions split is obtained to get to by pond
Change the feature after operation.The feature that feature after pondization is used to operate is used as training grader.If by image
In pond region be set as a continuous region in image, and in the constant or identical implicit layer unit extraction of pondization
Feature, then pond unit is translation invariant, even if image has done small translation, the pond feature of acquisition is still
The same, it will not change.
This operation of introducing pondization of the embodiment of the present invention, not only can be to avoid because greatly dropping during all features used
The dimension of low data, greatly reduces data operation quantity, the phenomenon that can also avoiding the occurrence of model over-fitting, reaches the good of model
Classification performance.
It is described based on the above-mentioned structure to CRBM, the structure and second for treating training network to first below in conjunction with the accompanying drawings are waited to instruct
The structure for practicing network is described in detail.
In practical applications, the image classification model in the present embodiment includes one or more CRBM, when image classification mould
When type includes multiple CRBM, each CRBM is trained respectively, and the sequence of training is：Have first to first CRBM
Supervision and the training of unsupervised mode, the parameters in first CRBM are improved and export corresponding characteristics of image,
Then, using the corresponding characteristics of image of output as the input of second CRBM, and so on, for each CRBM, all utilize
There are supervision and unsupervised mode to be trained in turn, so that each CRBM in image classification model is trained to splendid
State, parameter achieve the effect that best.
For example, CRBM is sequentially chosen from image classification model to be trained, when the sequence from image classification model
When the CRBM of selection is first CRBM, as shown in Figure 4 A, provided in an embodiment of the present invention first treats the structure group of training network
As the input layer and hidden layer of selected CRBM；As shown in Figure 4 B, provided in an embodiment of the present invention second training network is treated
Structure composition be selected CRBM input layer, hidden layer and Softmax sorter models.When from image classification mould
When the CRBM sequentially chosen in type is not first CRBM (such as second, it is a and subsequent to be of course applied for third
CRBM), as shown in Figure 5A, provided in an embodiment of the present invention first treat training network structure composition be selected CRBM it is defeated
Enter the pond layer of layer, hidden layer and the previous CRBM being connect with selected CRBM；As shown in Figure 5 B, the present invention is implemented
The second of example offer treats input layer, hidden layer and the selected CRBM that the structure composition of training network is selected CRBM
The pond layer of previous CRBM of connection and Softmax sorter models.
Step 203：Markd image pattern is not had according to acquisition, treats that training network carries out unsupervised mode to first
Training, the parameter that training network median filter model is treated with update first.
The training of unsupervised mode in the present embodiment is alternatively referred to as unsupervised learning (Unsupervised Learning)
A kind of network training mode that the training of mode, typically target class label etc. use, the information which is utilized
Be not the image pattern that known, specific namely training method is directed to it is not have markd image pattern.Here,
Unsupervised learning mode often has following two methods：A kind of is the direct method based on PDF estimation, and another kind is
Indirect clustering method is a kind of method that similitude using between image sample data is measured.As one kind compared with
Good embodiment, unsupervised learning mode is using indirect clustering method.
In the present embodiment, this step 203 specifically includes：First the filter model that training network includes is treated in initialization first
Parameter；Then, according to distance between the feature for not having markd image pattern, the feature of image pattern is clustered,
To obtain the newer parameter of filter model for treating that training network includes to first.
Here, various existing or new distance calculating method can be used, calculate the spy for not having markd image pattern
The distance between sign, is not specifically limited here；Wherein, the distance between feature of image pattern can be with Euclidean distance come table
Show, can also be represented with COS distance, the embodiment of the present invention does not limit herein.
It is because of any two image sample it should be noted why calculating the distance between feature of image pattern
The distance between this feature, can represent the similarity between the feature of two image patterns, i.e., the spy of two image patterns
Similitude between sign.Here cluster, it can be understood as the similarity between the feature of image pattern is measured, by phase
Gather like immediate image pattern feature is spent for one kind.
Step 204：Treat that training network extracts characteristics of image from image pattern based on second, it is special based on the image extracted
The label of sign and respective image sample, treats that the sorter model in training network carries out the training for having monitor mode to second.
The training for having monitor mode in the present embodiment is alternatively referred to as supervised learning (Supervised Learning) side
The training of formula, be it is a kind of the parameters in sorter model are constantly changed using labeled good data so that
Network reaches the optimal mode of learning of effect, supervised learning mode is usually applied to neural network and support vector machines
Training process.Specifically, supervised learning is to be trained study, and these learning samples using labeled good sample
Classification be categorized good sample.
Here, unsupervised learning and supervised learning are used in mixed way to the semi-supervised learning mode that constitutes, it is semi-supervised
The characteristics of mode of learning is can to reach trained by less marked data and many data not being marked
Network and the purpose that image is identified.
In the present embodiment, in this step 204 based on the characteristics of image extracted and the mark of respective image sample
Note, treats that the sorter model in training network have for the training of monitor mode to second, specifically may be used such as lower section
Formula is realized：Structure is with the label of the parameter of sorter model, the characteristics of image extracted and respective image sample
Cost function；Determine the updated value of the parameter of sorter model when cost function meets the condition of convergence.
Sorter model in the present embodiment can be Softmax sorter models, wherein, Softmax graders can be with
Represented using Softmax regression algorithms, belong to the probability each classified in multiple classification for calculating image pattern, below it is right
The expression formula of Softmax regression algorithms is briefly described.Softmax recurrence is a kind of logistic regression of citation form, is needle
To a kind of logistic regression of multicategory classification problem.Softmax recurrence can solve the problems, such as it is that different classes of data is avoided to be in
The state to repel each other.Assuming that there are one training set { (x(1),y(1)),...,(x(m),y(m)), wherein, y(i)∈{1,2,
... k }, x represents original input picture sample, y represent output as a result, using one comprising k dimensional vectors p (y=i | x) to all
The probability that classification results occur is indicated, the expression formula y=h of settingθ(x) as shown in formula (4)：
In above-mentioned expression formula (4), the parameters of sorter model use θ respectively1,θ2,...,θkIt represents, all
Probability addition result be 1.By above-mentioned formula (4) it is found that outputing the feature of k dimensions from the pond layer of CRBM, the spy that k is tieed up
Sign is input to Softmax graders (Feature Conversion for tieing up k is the corresponding probability of k classification), and the corresponding probability of k classification
Gaussian distributed.
Parameter based on sorter model, with reference to the characteristics of image and the regularization term of the label of respective image sample extracted
Afterwards, the cost function expression formula obtained is：
Local derviation is asked to the 1st parameter of j-th of classification, obtained partial derivative is：
After the processing minimized to J (θ), it is possible to reach Softmax classification return target, i.e., by pair
Parameter in cost function seeks the mode of local derviation, determines the update of the parameter of sorter model when cost function meets the condition of convergence
Value.
Step 205：Based on the error of the sorter model output result after training, update second is treated to filter in training network
The parameter of device model.
In the present embodiment, following manner may be used to realize in this step 205：It is defeated according to the sorter model after training
Go out the error of result, treat that the parameter of training network median filter model is adjusted to first, to adjust to meeting the condition of convergence,
The undated parameter of the filter model in training network is treated with acquisition second.
Here, training net is treated since training data is input to first during deep learning, when training neural network
The input layer of network median filter model such as CRBM by hidden layer, finally reaches pond layer and exports as a result, at this point, image sample
Error is likely to form between this desired value and output result, then calculates the error between estimated value and actual value, therefore, adopts
With back-propagation algorithm, by error from pond layer to hidden layer backpropagation, until input layer is traveled to, in the mistake of backpropagation
Cheng Zhong, according to the error of the sorter model output result after training, training network median filter model is treated in constantly adjustment first
Parameter, until meet the condition of convergence, filter model in training network is treated to obtain second the and constantly iteration above process
Undated parameter.Wherein, the initial value size of the parameters of filter model is random.
Step 206：The ginseng that image sample data input is treated into training network median filter model via updated first
Number and updated second treats the image classification model that the parameter of training network median filter model determines.
In the present embodiment, image classification model is a kind of mixed model, which has a certain number of filtering
The model structure that device model and sorter model combine.The number of filter model can be set according to actual conditions,
The number of filter model is set as one or more.It, can be by multiple filters when the number of filter model is multiple
Wave device model is linked in sequence according to the preset order of connection.
The difference between the RBM in filter model such as CRBM, with the relevant technologies in image classification model essentially consists in
For CRBM in a sub-picture, weights this parameter in the hidden layer and input layer in this image whole region is shared；
In addition, pond layer is further included in CRBM.The network structure of CRBM as shown in Figure 3, wherein pond layer can represent with P,
In CRBM models, there is one layer to be known as detection layers, i.e. hidden layer, which is the hidden layer by being obtained after convolution operation
And pond layer has the element number of identical group of number, is set as K group, in the layer of pond, each group in K group includes NP
×NPA unit, wherein k ∈ { 1 ..., K }, choose a smaller integer value, such as 2 or 3, using C to this integer value into
Row represents that pond layer P compress the expression of hidden layer using C, and hidden layer H is divided into block of the size for C × C, from every
One P layers of binary cell p is extracted in a blockk, therefore, NP=NH/C。
The signal that input layer V is transmitted bottom-up can be received by hidden layer, and the associated expression of V is as follows：
In above-mentioned expression formula, useIt is indicated come the element number to kth layer in hidden layer, uses WkCome to
K convolution kernel is indicated, and uses bkIt is indicated come the bias to kth layer in hidden layer.
After image data is input to input layer and is received by hidden layer, and then, each fritter is sampled respectively,
As the multinomial of input, comprising implicit layer unit in each fritter, utilizeImplicit layer unit is indicated,
Increased energy is used after opening unitIt represents, then the conditional probability of hidden layer and pond layer can be expressed as below：
If it is known that hidden layer, then input layer can be asked for using expression formula (10).And above-mentioned expression formula (8) and
(9) then in the case where giving the state of input layer, the conditional probability of hidden layer and pond layer how is asked for.Due to pond
Do not include free parameter in layer, therefore during training CRBM, hidden layer is broadcast to input using back-propagation algorithm
Layer.After several CRBM are combined together, the operation of maximum probability pondization is introduced, that is, will in the tip portion of CRBM structures
Image is divided into after several mutually independent parts, and probability value maximum in this part is selected to use.
It, can not possibly unconfined repetition training image in practical application in view of the duration and training effect of model training
Disaggregated model, therefore, usually setting first treat that training network and second treats the respective threshold of the training network training number of plies, that is, set
First treats that the training number of plies threshold value of training network and second treats the training number of plies threshold value of training network, and first treats training network
The training number of plies is identical with the training number of plies that second treats training network.That is, treating that it is unsupervised that training network carries out to first
After the completion of one wheel training of mode, then treat that training network carries out the wheel training for having monitor mode to second, in each training in rotation
After the completion of white silk, all detect whether the completed trained number of plies reaches pre-set trained number of plies threshold value, when detecting trained layer
When number not up to trains number of plies threshold value, then the training number of plies is added 1, and continue initialisation image disaggregated model and carry out model training,
Until when detecting that the trained number of plies reaches trained number of plies threshold value, then it can terminate model training process.
In practical applications, training number of plies threshold value can be adjusted accordingly, but training number of plies threshold value according to actual conditions
It can not be too big.Usually training number of plies threshold value can be set as 5 layers or 6 layers, at this point, the parameters base in image classification model
This is trained to splendid state.
Step 207：Obtain each probability value of image classification model output；Wherein, each probability value represents image pattern respectively
The probability size of each data generic in data；According to each probability value, the classification for meeting Probability Condition is chosen, according to selected
The classification identification image sample data taken.
It here, can be by image sample data input picture disaggregated model to be identified, to image sample data to be identified
In each data vector represent converted, using the result after transformation as image sample data generic probability progress it is defeated
Go out, to obtain each probability value of the difference generic of each data in image sample data to be identified.Specifically, based on figure
As the excitation function of nodes different in disaggregated model, the vector of the image sample data of input is represented to convert, will be converted
Result as classification vector represent and its corresponding probability.The classification for meeting Probability Condition in the present embodiment can be for probability most
The high classification as image sample data to be identified.That is, it is selected from each probability value that image classification model exports
The corresponding classification of probability peak is taken, as the image category finally identified.
It should be noted that image classification model mentioned here, is to treat to filter in training network by updated first
The parameter of device model and updated second treats the image classification model that the parameter of training network median filter model determines, and
The implementation procedure that parameter in the image classification model has been subjected to above-mentioned steps 201 to step 206 carried out it is continuous modification and it is complete
It is kind, reach best training effect.
Using the technical solution of the embodiment of the present invention, by by both unsupervised learning mode and supervised learning mode knot
It closes, and alternates use, the parameters in image classification model are constantly adjusted so that the receipts of image classification model
It holds back speed faster, is more easy to reach plateau, reduces the time of model training study；Meanwhile the embodiment of the present invention is using above-mentioned
The image classification model of structure carries out image identification to image data to be identified, can avoid the occurrence of because of back-propagation algorithm band
The problem of gradient disperse come, greatly improves the accuracy rate of image identification.
In order to realize the processing method of above-mentioned image classification model, the embodiment of the present invention additionally provides a kind of image classification mould
The processing unit of type, the processing unit of the image classification model are applied in server or terminal device, and Fig. 6 is implemented for the present invention
One optional illustrative view of functional configuration of the processing unit for the image classification model that example provides；As shown in fig. 6, the image classification
The processing unit of model includes initialization module 61, determining module 62, the first training module 63, extraction module 64, second and trains
Module 65 and update module 66.Each program module is described in detail below.
Initialization module 61, for initialisation image disaggregated model；
Determining module 62, for determining that sequentially chosen from image classification model the first of filter model treats training net
Network and determine the filter model for including sequence selection and sorter model second treat training network；
First training module 63, for not having markd image pattern according to acquisition, treated to first training network into
The parameter of training network median filter model is treated in the training of the unsupervised mode of row with update first；
Extraction module 64 treats that training network extracts characteristics of image from image pattern for being based on second；
Second training module 65, for based on the characteristics of image extracted and the label of respective image sample, being treated to second
Sorter model in training network carries out the training for having monitor mode；
Update module 66, for the error based on the sorter model output result after training, training net is treated in update second
The parameter of network median filter model.
In alternative embodiment of the present invention, for the determining filtering sequentially chosen from image classification model of determining module 62
The first of device model treats training network and determines that including the filter model of sequence selection and the second of sorter model waits to instruct
Practice for network, following manner may be used to realize：
When sequentially choosing target filter model to be trained, and selected target filter from image classification model
When model is first filter model in image classification model, the input layer and hidden layer of target filter model can be chosen
Training network is treated as first, and input layer, hidden layer and the sorter model for choosing target filter model are treated as second
Training network.
In another alternative embodiment of the present invention, that is sequentially chosen from image classification model is determined for determining module 62
The first of filter model treats training network and determines to include the second of the filter model and sorter model sequentially chosen
It treats for training network, following manner may be used to realize：
When sequentially choosing target filter model to be trained, and selected target filter from image classification model
When model is not first filter model in image classification model, the input layer of target filter model can be chosen, implied
Layer and the pond layer of the previous filter model of target filter model connection treat training network as first, choose mesh
Mark the input layer of filter model, hidden layer, the previous filter model being connect with target filter model pond layer, with
And sorter model treats training network as second.
In the present embodiment, markd image pattern is not had according to acquisition for the first training module 63, to first
It treats that training network carries out the training of unsupervised mode, is treated for the parameter of training network median filter model with update first, it can
To realize in the following way：
First, the parameter for the filter model that training network includes is treated in initialization first；Then, it is markd according to not having
Distance between the feature of image pattern clusters the feature of image pattern, treats what training network included to first to obtain
The newer parameter of filter model.
In the present embodiment, for the second training module 65 based on the characteristics of image extracted and the mark of respective image sample
Note, treats that the sorter model in training network have for the training of monitor mode to second, and following manner may be used
It realizes：
First, build using the label of the parameter of sorter model, the characteristics of image extracted and respective image sample as
The cost function of son；Then, it is determined that when cost function meets the condition of convergence parameter of sorter model updated value.
For error of the update module 66 based on the sorter model output result after training, training network is treated in update second
For the parameter of median filter model, following manner may be used to realize：
According to the error of the sorter model output result after training, the ginseng for treating training network median filter model to first
Number is adjusted, and to adjust to the condition of convergence is met, the undated parameter of the filter model in training network is treated in acquisition second.
Fig. 7 is that another optional functional structure of the processing unit of image classification model provided in an embodiment of the present invention is shown
It is intended to；As shown in fig. 7, the processing unit of the image classification model further includes：Input module 67, for being updated in update module 66
After second treats the parameter of training network median filter model, image sample data is inputted and waits to train via updated first
The parameter of network median filter model and updated second treats the image point that the parameter of training network median filter model determines
Class model；
Acquisition module 68, for obtaining each probability value of image classification model output；Wherein, each probability value represents to scheme respectively
The probability size of each data generic in decent notebook data；
Module 69 is chosen, for according to each probability value, choosing the classification for meeting Probability Condition；
Identification module 610, for identifying image sample data according to selected classification.
It should be noted that：Above-described embodiment provide image classification model processing unit to image classification model into
It, can as needed will be above-mentioned only with the division progress of above-mentioned each program module for example, in practical application during row processing
Processing distribution is completed by different program modules, i.e., is divided into the internal structure of the processing unit of image classification model different
Program module, to complete all or part of processing described above.In addition, the image classification model that above-described embodiment provides
Processing unit and the processing method embodiment of image classification model belong to same design, and specific implementation process refers to method implementation
Example, which is not described herein again.
In practical applications, above procedure mould initialization module 61 in the block, determining module 62, the first training module 63,
Extraction module 64, the second training module 65, update module 66, input module 67, choose module 69 and identification module 610 can be by
On server or terminal device central processing unit (CPU, Central Processing Unit), microprocessor (MPU,
Micro Processor Unit), digital signal processor (DSP, Digital Signal Processor) or scene can compile
The realizations such as journey gate array (FPGA, Field Programmable Gate Array)；Above procedure mould acquisition module 68 in the block
It (can be included by communications module in practical applications：Base communication external member, operating system, communication module, standard interface and association
View etc.) and dual-mode antenna realization.
In order to realize the processing method of above-mentioned image classification model, the embodiment of the present invention additionally provides a kind of image classification mould
The hardware configuration of the processing unit of type.The processing of the image classification model of the embodiment of the present invention is realized in description with reference to the drawings
Device, the processing unit of the image classification model can be implemented in a variety of manners, such as server such as Cloud Server, terminal are set
Standby such as desktop computer, laptop, smart mobile phone computer equipment.Below to the image classification mould of the embodiment of the present invention
The hardware configuration of the processing unit of type is described further, it will be understood that Fig. 8 illustrate only the processing dress of image classification model
The example arrangement rather than entire infrastructure put can implement part-structure or entire infrastructure shown in Fig. 8 as needed.
Referring to Fig. 8, Fig. 8 is an optional hardware of the processing unit of image classification model provided in an embodiment of the present invention
Structure diagram can be applied in practical application in the various servers or terminal device of aforementioned operation application program, Fig. 8 institutes
The processing unit 800 of the image classification model shown includes：At least one processor 801, memory 802, user interface 803 and extremely
A few network interface 804.Various components in the processing unit 800 of the image classification model are coupled by bus system 805
Together.It is appreciated that bus system 805 is used to implement the connection communication between these components.Bus system 805, which is removed, includes number
Except bus, power bus, controlling bus and status signal bus in addition are further included.But for the sake of clear explanation, in fig. 8
Various buses are all designated as bus system 805.
Wherein, user interface 803 can include display, keyboard, mouse, trace ball, click wheel, button, button, sense of touch
Plate or touch screen etc..
It is appreciated that memory 802 can be volatile memory or nonvolatile memory, may also comprise volatibility and
Both nonvolatile memories.
Memory 802 in the embodiment of the present invention is used to store various types of data to support the place of image classification model
Manage the operation of device 800.The example of these data includes：Appoint for what is operated in the processing unit 800 of image classification model
What computer program, such as executable program 8021 and operating system 8022, realizes the image classification model of the embodiment of the present invention
The program of processing method may be embodied in executable program 8021.
The processing method for the image classification model that the embodiment of the present invention discloses can be applied in processor 801, Huo Zheyou
Processor 801 is realized.Processor 801 may be a kind of IC chip, have the processing capacity of signal.In the process of realization
In, each step of the processing method of above-mentioned image classification model can pass through the integrated logic circuit of the hardware in processor 801
Or the instruction of software form is completed.Above-mentioned processor 801 can be general processor, DSP or other programmable logic
Device, discrete gate or transistor logic, discrete hardware components etc..The present invention can be realized or be performed to processor 801
Processing method, step and the logic diagram of each image classification model provided in embodiment.General processor can be microprocessor
Device or any conventional processor etc..The step of the processing method of image classification model provided with reference to the embodiment of the present invention
Suddenly, hardware decoding processor can be embodied directly in and perform completion or with the hardware in decoding processor and software module group
Conjunction performs completion.Software module can be located in storage medium, which is located at memory 802, and the reading of processor 801 is deposited
Information in reservoir 802, the step of completing the processing method of image classification model provided in an embodiment of the present invention with reference to its hardware.
In embodiments of the present invention, the processing unit 800 of image classification model includes memory 802, processor 801 and deposits
The executable program 8021 that can be run on memory 802 and by processor 801 is stored up, processor 801 runs executable program
It is realized when 8021：Initialisation image disaggregated model；Determine sequentially chosen from image classification model the first of filter model
It treats training network and determines that including the filter model of sequence selection and the second of sorter model treats training network；According to
What is obtained does not have markd image pattern, treats that training network carries out the training of unsupervised mode to first, is treated with update first
The parameter of training network median filter model；Treat that training network extracts characteristics of image from image pattern based on second；Based on institute
The characteristics of image of extraction and the label of respective image sample treat that the sorter model in training network has carried out supervision side to second
The training of formula；Based on the error of the sorter model output result after training, training network median filter model is treated in update second
Parameter.
As a kind of embodiment, processor 801 is realized when running executable program 8021：It is suitable from image classification model
Sequence chooses target filter model to be trained, and selected target filter model is first in image classification model
During filter model, the input layer and hidden layer of choosing target filter model treat training network as first；Choose target filter
Input layer, hidden layer and the sorter model of wave device model treat training network as second.
As a kind of embodiment, processor 801 is realized when running executable program 8021：It is suitable from image classification model
Sequence chooses target filter model to be trained, and selected target filter model be not in image classification model first
During a filter model, the previous of the input layer of target filter model, hidden layer and the connection of target filter model is chosen
The pond layer of a filter model treats training network as first；Choose input layer, hidden layer and the mesh of target filter model
The pond layer and sorter model of the previous filter model of mark filter model connection treat training network as second.
As a kind of embodiment, processor 801 is realized when running executable program 8021：Training net is treated in initialization first
The parameter for the filter model that network includes；According to distance between the feature for not having markd image pattern, to image pattern
Feature is clustered, to obtain the newer parameter of filter model for treating that training network includes to first.
As a kind of embodiment, processor 801 is realized when running executable program 8021：Structure is with sorter model
The cost function that the label of parameter, the characteristics of image extracted and respective image sample is；It determines that cost function meets to receive
The updated value of the parameter of sorter model when holding back condition.
As a kind of embodiment, processor 801 is realized when running executable program 8021：According to the grader after training
Model exports the error of result, treats that the parameter of training network median filter model is adjusted to first, is received with adjusting to satisfaction
Condition is held back, the undated parameter of the filter model in training network is treated with acquisition second.
As a kind of embodiment, processor 801 is realized when running executable program 8021：Training net is treated in update second
After the parameter of network median filter model, image sample data is inputted and treats training network median filter via updated first
The parameter of model and updated second treats the image classification model that the parameter of training network median filter model determines；It obtains
Each probability value of image classification model output；Wherein, each probability value represents the affiliated class of each data in image sample data respectively
Other probability size；According to each probability value, the classification for meeting Probability Condition is chosen, image pattern is identified according to selected classification
Data.
In the exemplary embodiment, the embodiment of the present invention additionally provides a kind of storage medium, the storage medium can be CD,
The storage mediums such as flash memory or disk are chosen as non-moment storage medium.
Executable program 8021 is stored on storage medium provided in an embodiment of the present invention, executable program 8021 is handled
Device 801 is realized when performing：Initialisation image disaggregated model；Determine the filter model sequentially chosen from image classification model
First treats training network and determines that including the filter model of sequence selection and the second of sorter model treats training network；
Markd image pattern is not had according to acquisition, treats that training network carries out the training of unsupervised mode to first, to update the
One treats the parameter of training network median filter model；Treat that training network extracts characteristics of image from image pattern based on second；Base
In the characteristics of image and the label of respective image sample that are extracted, treat that the sorter model in training network has carried out prison to second
Superintend and direct the training of mode；Based on the error of the sorter model output result after training, training network median filter is treated in update second
The parameter of model.
As a kind of embodiment, executable program 8021 is realized when being performed by processor 801：From image classification model
Sequence chooses target filter model to be trained, and selected target filter model is first in image classification model
During a filter model, the input layer and hidden layer of choosing target filter model treat training network as first；Choose target
Input layer, hidden layer and the sorter model of filter model treat training network as second.
As a kind of embodiment, executable program 8021 is realized when being performed by processor 801：From image classification model
Sequence chooses target filter model to be trained, and selected target filter model be not in image classification model the
During one filter model, before choosing input layer, hidden layer and target filter the model connection of target filter model
The pond layer of one filter model treats training network as first；Choose the input layer of target filter model, hidden layer, with
The pond layer and sorter model of the previous filter model of target filter model connection treat training net as second
Network.
As a kind of embodiment, executable program 8021 is realized when being performed by processor 801：Initialization first is waited to train
The parameter for the filter model that network includes；According to distance between the feature for not having markd image pattern, to image pattern
Feature clustered, to obtain the newer parameter of filter model for treating that training network includes to first.
As a kind of embodiment, executable program 8021 is realized when being performed by processor 801：Structure is with sorter model
Parameter, the characteristics of image that is extracted and respective image sample the label cost function that is；Determine that cost function meets
The updated value of the parameter of sorter model during the condition of convergence.
As a kind of embodiment, executable program 8021 is realized when being performed by processor 801：According to the classification after training
Device model exports the error of result, treats that the parameter of training network median filter model is adjusted to first, to adjust to satisfaction
The condition of convergence treats the undated parameter of the filter model in training network with acquisition second.
As a kind of embodiment, executable program 8021 is realized when being performed by processor 801：It waits to train in update second
After the parameter of network median filter model, image sample data is inputted and treats to filter in training network via updated first
The parameter of device model and updated second treats the image classification model that the parameter of training network median filter model determines；It obtains
Each probability value that image classification model is taken to export；Wherein, each probability value is represented respectively belonging to each data in image sample data
The probability size of classification；According to each probability value, the classification for meeting Probability Condition is chosen, image sample is identified according to selected classification
Notebook data.
Using the technical solution of the embodiment of the present invention, training network and the are treated by determining first from image classification model
Two treat training network, according to not having markd image pattern, treat that training network carries out unsupervised mode training to first, with more
New first treats the parameter of training network median filter model；The label of characteristics of image and respective image sample based on extraction, it is right
Second treats that training network has carried out monitor mode training, and the parameter of training network median filter model is treated with update second.In this way,
Unsupervised learning mode and supervised learning mode are alternated into use, the parameters in image classification model are carried out not
Disconnected adjustment so that the convergence rate of image classification model faster, is more easy to reach plateau, reduce model training study when
Between；Meanwhile the embodiment of the present invention carries out image identification using the image classification model of above-mentioned structure to image data to be identified,
The problem of gradient disperse brought by back-propagation algorithm can be avoided the occurrence of, greatly improves the accuracy rate of image identification, is promoted
User experience.
It should be understood by those skilled in the art that, the embodiment of the present invention can be provided as method, system or executable program
Product.Therefore, the shape of the embodiment in terms of hardware embodiment, software implementation or combination software and hardware can be used in the present invention
Formula.Moreover, the present invention can be used can use storage in one or more computers for wherein including computer usable program code
The form of executable program product that medium is implemented on (including but not limited to magnetic disk storage and optical memory etc.).
The present invention be with reference to according to the method for the embodiment of the present invention, the flow of equipment (system) and executable program product
Figure and/or block diagram describe.It should be understood that it can be realized by executable program instructions every first-class in flowchart and/or the block diagram
The combination of flow and/or box in journey and/or box and flowchart and/or the block diagram.These executable programs can be provided
The processor of all-purpose computer, special purpose computer, Embedded Processor or reference programmable data processing device is instructed to produce
A raw machine so that the instruction performed by computer or with reference to the processor of programmable data processing device is generated for real
The device of function specified in present one flow of flow chart or one box of multiple flows and/or block diagram or multiple boxes.
These executable program instructions, which may also be stored in, can guide computer or with reference to programmable data processing device with spy
Determine in the computer-readable memory that mode works so that the instruction generation being stored in the computer-readable memory includes referring to
Enable the manufacture of device, the command device realize in one flow of flow chart or multiple flows and/or one box of block diagram or
The function of being specified in multiple boxes.
These executable program instructions can also be loaded into computer or with reference in programmable data processing device so that count
Calculation machine or with reference to performing series of operation steps on programmable device to generate computer implemented processing, so as in computer or
It is used to implement with reference to the instruction offer performed on programmable device in one flow of flow chart or multiple flows and/or block diagram one
The step of function of being specified in a box or multiple boxes.
The foregoing is only a preferred embodiment of the present invention, is not intended to limit the scope of the present invention, it is all
All any modification, equivalent and improvement made within the spirit and principles in the present invention etc. should be included in the protection of the present invention
Within the scope of.
Claims (14)
1. a kind of processing method of image classification model, which is characterized in that the method includes：
Initialisation image disaggregated model；
Determine sequentially to choose from described image disaggregated model the first of filter model treats training network and determines to include
The filter model and the second of sorter model that the sequence is chosen treats training network；
Markd image pattern is not had according to acquisition, treats that training network carries out the training of unsupervised mode to described first,
To update the described first parameter for treating training network median filter model；
Treat that training network extracts characteristics of image from described image sample based on described second；
Based on the characteristics of image extracted and the label of respective image sample, the grader mould in training network is treated to described second
Type carries out the training for having monitor mode；
Based on the error of the sorter model output result after training, training network median filter model is treated in update described second
Parameter.
2. the processing method of image classification model according to claim 1, which is characterized in that described to determine from described image
The first of the filter model sequentially chosen in disaggregated model treats training network and determines to include the filtering that the sequence is chosen
The second of device model and sorter model treats training network, including：
Target filter model to be trained, and selected target filter mould are sequentially chosen from described image disaggregated model
When type is first filter model in described image disaggregated model,
The input layer and hidden layer for choosing the target filter model treat training network as described first；
The input layer, hidden layer and the sorter model for choosing the target filter model are waited to instruct as described second
Practice network.
3. the processing method of image classification model according to claim 1, which is characterized in that described to determine from described image
The first of the filter model sequentially chosen in disaggregated model treats training network and determines to include the filtering that the sequence is chosen
The second of device model and sorter model treats training network, including：
Target filter model to be trained, and selected target filter mould are sequentially chosen from described image disaggregated model
When type is not first filter model in described image disaggregated model,
Choose the previous of input layer, hidden layer and the target filter model connection of the target filter model
The pond layer of filter model treats training network as described first；
Choose the input layer, hidden layer, the previous filter being connect with the target filter model of the target filter model
The pond layer and the sorter model of wave device model treat training network as described second.
4. the processing method of image classification model according to claim 1, which is characterized in that not the having according to acquisition
Markd image pattern treats that training network carries out the training of unsupervised mode to described first, waits to instruct to update described first
Practice the parameter of network median filter model, including：
Initialize the described first parameter for treating the filter model that training network includes；
According to distance between the feature for not having markd image pattern, the feature of image pattern is clustered, to obtain
It obtains and the newer parameter of filter model that training network includes is treated to described first.
5. the processing method of image classification model according to claim 1, which is characterized in that described based on the figure extracted
As feature and the label of respective image sample, treat that the sorter model in training network carries out having monitor mode to described second
Training, including：
Build the generation for being with the label of the parameter of the sorter model, the characteristics of image extracted and respective image sample
Valency function；
Determine that the cost function meets the updated value of the parameter of the sorter model during condition of convergence；
Training network median filter mould is treated in the error of the sorter model output result based on after training, update described second
The parameter of type, including：
According to the error of the sorter model output result after training, the ginseng for treating training network median filter model to described first
Number is adjusted, to adjust to meeting the condition of convergence, to obtain the described second update for treating filter model in training network
Parameter.
6. the processing method of image classification model according to claim 1, which is characterized in that in the update described second
After the parameter for treating training network median filter model, the method further includes：
The parameter that image sample data input is treated into training network median filter model via updated described first and update
Described second afterwards treats the image classification model that the parameter of training network median filter model determines；
Obtain each probability value of described image disaggregated model output；Wherein, each probability value represents described image sample respectively
The probability size of each data generic in data；
According to each probability value, the classification for meeting Probability Condition is chosen, described image sample is identified according to selected classification
Data.
7. a kind of processing unit of image classification model, which is characterized in that described device includes：Initialization module, determining module,
First training module, extraction module, the second training module and update module；Wherein,
The initialization module, for initialisation image disaggregated model；
The determining module, for determining that sequentially chosen from described image disaggregated model the first of filter model waits to train
Network and determine that including the filter model of sequence selection and the second of sorter model treats training network；
For not having markd image pattern according to acquisition, training network is treated to described first for first training module
The training of unsupervised mode is carried out, to update the described first parameter for treating training network median filter model；
The extraction module treats that training network extracts characteristics of image from described image sample for being based on described second；
Second training module, for based on the characteristics of image extracted and the label of respective image sample, to described second
Treat that the sorter model in training network carries out the training for having monitor mode；
The update module, for the error based on the sorter model output result after training, update described second is waited to train
The parameter of network median filter model.
8. the processing unit of image classification model according to claim 7, which is characterized in that the determining module, specifically
For：
Target filter model to be trained, and selected target filter mould are sequentially chosen from described image disaggregated model
When type is first filter model in described image disaggregated model,
The input layer and hidden layer for choosing the target filter model treat training network as described first；
The input layer, hidden layer and the sorter model for choosing the target filter model are waited to instruct as described second
Practice network.
9. the processing unit of image classification model according to claim 7, which is characterized in that the determining module, specifically
For：
Target filter model to be trained, and selected target filter mould are sequentially chosen from described image disaggregated model
When type is not first filter model in described image disaggregated model,
Choose the previous of input layer, hidden layer and the target filter model connection of the target filter model
The pond layer of filter model treats training network as described first；
Choose the input layer, hidden layer, the previous filter being connect with the target filter model of the target filter model
The pond layer and the sorter model of wave device model treat training network as described second.
10. the processing unit of image classification model according to claim 7, which is characterized in that first training module,
It is specifically used for：
Initialize the described first parameter for treating the filter model that training network includes；
According to distance between the feature for not having markd image pattern, the feature of image pattern is clustered, to obtain
It obtains and the newer parameter of filter model that training network includes is treated to described first.
11. the processing unit of image classification model according to claim 7, which is characterized in that second training module,
It is specifically used for：
Build the generation for being with the label of the parameter of the sorter model, the characteristics of image extracted and respective image sample
Valency function；
Determine that the cost function meets the updated value of the parameter of the sorter model during condition of convergence；
The update module, is specifically used for：
According to the error of the sorter model output result after training, the ginseng for treating training network median filter model to described first
Number is adjusted, to adjust to meeting the condition of convergence, to obtain the described second update for treating filter model in training network
Parameter.
12. the processing unit of image classification model according to claim 7, which is characterized in that described device further includes：It is defeated
Enter module, after treating the parameter of training network median filter model in update module update described second, by image
The parameter and updated described that sample data input treats training network median filter model via updated described first
Two treat the image classification model that the parameter of training network median filter model determines；
Acquisition module, for obtaining each probability value of described image disaggregated model output；Wherein, each probability value represents respectively
The probability size of each data generic in described image sample data；
Module is chosen, for according to each probability value, choosing the classification for meeting Probability Condition；
Identification module, for identifying described image sample data according to selected classification.
13. a kind of storage medium, is stored thereon with executable program, which is characterized in that the executable code processor is held
The step of processing method such as claim 1 to 6 any one of them image classification model is realized during row.
14. a kind of processing unit of image classification model, including memory, processor and storage on a memory and can be by institute
State the executable program of processor operation, which is characterized in that the processor performs such as right when running the executable program
It is required that the step of processing method of 1 to 6 any one of them image classification model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810087876.2A CN108197666A (en) | 2018-01-30 | 2018-01-30 | A kind of processing method, device and the storage medium of image classification model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810087876.2A CN108197666A (en) | 2018-01-30 | 2018-01-30 | A kind of processing method, device and the storage medium of image classification model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108197666A true CN108197666A (en) | 2018-06-22 |
Family
ID=62591772
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810087876.2A Pending CN108197666A (en) | 2018-01-30 | 2018-01-30 | A kind of processing method, device and the storage medium of image classification model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108197666A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109934281A (en) * | 2019-03-08 | 2019-06-25 | 电子科技大学 | A kind of unsupervised training method of two sorter networks |
| CN110276283A (en) * | 2019-06-10 | 2019-09-24 | 腾讯科技（深圳）有限公司 | Image identification method, Model of Target Recognition training method and device |
| CN110399918A (en) * | 2019-07-24 | 2019-11-01 | 深兰科技（上海）有限公司 | A kind of method and apparatus of target identification |
| CN110909820A (en) * | 2019-12-02 | 2020-03-24 | 齐鲁工业大学 | Image classification method and system based on self-supervision learning |
| CN111476282A (en) * | 2020-03-27 | 2020-07-31 | 东软集团股份有限公司 | Data classification method and device, storage medium and electronic equipment |
| CN111553267A (en) * | 2020-04-27 | 2020-08-18 | 腾讯科技（深圳）有限公司 | Image processing method, image processing model training method and device |
| CN111753863A (en) * | 2019-04-12 | 2020-10-09 | 北京京东尚科信息技术有限公司 | Image classification method and device, electronic equipment and storage medium |
| CN111860582A (en) * | 2020-06-11 | 2020-10-30 | 北京市威富安防科技有限公司 | Image classification model construction method and device, computer equipment and storage medium |
| CN112800880A (en) * | 2021-01-15 | 2021-05-14 | 天津普达软件技术有限公司 | Method for estimating weight of hazardous waste |
| WO2022028147A1 (en) * | 2020-08-06 | 2022-02-10 | 腾讯科技（深圳）有限公司 | Image classification model training method and apparatus, computer device, and storage medium |
| CN114341951A (en) * | 2019-08-20 | 2022-04-12 | 阿卡蒂科有限公司 | Method and system for diagnosing diseases, physiological changes or other internal conditions of crustaceans by non-invasive means |
| CN115346076A (en) * | 2022-10-18 | 2022-11-15 | 安翰科技(武汉)股份有限公司 | Pathological image recognition method, model training method and system thereof, and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102208037A (en) * | 2011-06-10 | 2011-10-05 | 西安电子科技大学 | Hyper-spectral image classification method based on Gaussian process classifier collaborative training algorithm |
| CN104767692A (en) * | 2015-04-15 | 2015-07-08 | 中国电力科学研究院 | Network traffic classification method |
| CN105956612A (en) * | 2016-04-25 | 2016-09-21 | 西安电子科技大学 | Hyperspectral image classification method based on active learning and semi-supervised learning |
| CN105989375A (en) * | 2015-01-30 | 2016-10-05 | 富士通株式会社 | Classifier, classification device and classification method for classifying handwritten character images |
| CN107563428A (en) * | 2017-08-25 | 2018-01-09 | 西安电子科技大学 | Classification of Polarimetric SAR Image method based on generation confrontation network |
-
2018
- 2018-01-30 CN CN201810087876.2A patent/CN108197666A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102208037A (en) * | 2011-06-10 | 2011-10-05 | 西安电子科技大学 | Hyper-spectral image classification method based on Gaussian process classifier collaborative training algorithm |
| CN105989375A (en) * | 2015-01-30 | 2016-10-05 | 富士通株式会社 | Classifier, classification device and classification method for classifying handwritten character images |
| CN104767692A (en) * | 2015-04-15 | 2015-07-08 | 中国电力科学研究院 | Network traffic classification method |
| CN105956612A (en) * | 2016-04-25 | 2016-09-21 | 西安电子科技大学 | Hyperspectral image classification method based on active learning and semi-supervised learning |
| CN107563428A (en) * | 2017-08-25 | 2018-01-09 | 西安电子科技大学 | Classification of Polarimetric SAR Image method based on generation confrontation network |
Non-Patent Citations (1)
| Title |
|---|
| 韦鹏程、冉维、段昂: "《大数据巨量分析与机器学习的整合开发》", 31 May 2017 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109934281A (en) * | 2019-03-08 | 2019-06-25 | 电子科技大学 | A kind of unsupervised training method of two sorter networks |
| CN109934281B (en) * | 2019-03-08 | 2021-01-26 | 电子科技大学 | Unsupervised training method of two-class network |
| CN111753863A (en) * | 2019-04-12 | 2020-10-09 | 北京京东尚科信息技术有限公司 | Image classification method and device, electronic equipment and storage medium |
| CN110276283A (en) * | 2019-06-10 | 2019-09-24 | 腾讯科技（深圳）有限公司 | Image identification method, Model of Target Recognition training method and device |
| CN110399918B (en) * | 2019-07-24 | 2021-11-19 | 深兰科技（上海）有限公司 | Target identification method and device |
| CN110399918A (en) * | 2019-07-24 | 2019-11-01 | 深兰科技（上海）有限公司 | A kind of method and apparatus of target identification |
| CN114341951A (en) * | 2019-08-20 | 2022-04-12 | 阿卡蒂科有限公司 | Method and system for diagnosing diseases, physiological changes or other internal conditions of crustaceans by non-invasive means |
| CN110909820A (en) * | 2019-12-02 | 2020-03-24 | 齐鲁工业大学 | Image classification method and system based on self-supervision learning |
| CN110909820B (en) * | 2019-12-02 | 2023-04-07 | 齐鲁工业大学 | Image classification method and system based on self-supervision learning |
| CN111476282A (en) * | 2020-03-27 | 2020-07-31 | 东软集团股份有限公司 | Data classification method and device, storage medium and electronic equipment |
| CN111553267A (en) * | 2020-04-27 | 2020-08-18 | 腾讯科技（深圳）有限公司 | Image processing method, image processing model training method and device |
| CN111553267B (en) * | 2020-04-27 | 2023-12-01 | 腾讯科技（深圳）有限公司 | Image processing method, image processing model training method and device |
| CN111860582B (en) * | 2020-06-11 | 2021-05-11 | 北京市威富安防科技有限公司 | Image classification model construction method and device, computer equipment and storage medium |
| CN111860582A (en) * | 2020-06-11 | 2020-10-30 | 北京市威富安防科技有限公司 | Image classification model construction method and device, computer equipment and storage medium |
| WO2022028147A1 (en) * | 2020-08-06 | 2022-02-10 | 腾讯科技（深圳）有限公司 | Image classification model training method and apparatus, computer device, and storage medium |
| CN112800880A (en) * | 2021-01-15 | 2021-05-14 | 天津普达软件技术有限公司 | Method for estimating weight of hazardous waste |
| CN112800880B (en) * | 2021-01-15 | 2022-05-20 | 天津普达软件技术有限公司 | Method for estimating weight of hazardous waste |
| CN115346076A (en) * | 2022-10-18 | 2022-11-15 | 安翰科技(武汉)股份有限公司 | Pathological image recognition method, model training method and system thereof, and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108197666A (en) | A kind of processing method, device and the storage medium of image classification model | |
| Pan et al. | Recent progress on generative adversarial networks (GANs): A survey | |
| Hossain et al. | Improving consumer satisfaction in smart cities using edge computing and caching: A case study of date fruits classification | |
| CN111079639B (en) | Method, device, equipment and storage medium for constructing garbage image classification model | |
| Li et al. | Deepsaliency: Multi-task deep neural network model for salient object detection | |
| Lin et al. | Exploring explicit domain supervision for latent space disentanglement in unpaired image-to-image translation | |
| Bi et al. | A survey on evolutionary computation for computer vision and image analysis: Past, present, and future trends | |
| Jia et al. | Deep multi-view learning using neuron-wise correlation-maximizing regularizers | |
| Jiang et al. | Blind image quality measurement by exploiting high-order statistics with deep dictionary encoding network | |
| CN110276741B (en) | Method and device for nodule detection and model training thereof and electronic equipment | |
| Wu et al. | Feedback weight convolutional neural network for gait recognition | |
| CN109063719A (en) | A kind of image classification method of co-ordinative construction similitude and category information | |
| Wu et al. | Enhancing TripleGAN for semi-supervised conditional instance synthesis and classification | |
| Krishnapriya et al. | Pre-trained deep learning models for brain MRI image classification | |
| Li et al. | Semi-supervised learning using adversarial training with good and bad samples | |
| Wang et al. | DBLCNN: Dependency-based lightweight convolutional neural network for multi-classification of breast histopathology images | |
| Goswami | Impact of deep learning in image processing and computer vision | |
| Li et al. | Breast cancer histopathological image classification based on deep second-order pooling network | |
| Kishorjit Singh et al. | Image classification using SLIC superpixel and FAAGKFCM image segmentation | |
| CN109522970A (en) | Image classification method, apparatus and system | |
| Hada et al. | Sparse oblique decision trees: A tool to understand and manipulate neural net features | |
| Lu et al. | Combining context, consistency, and diversity cues for interactive image categorization | |
| Peng et al. | Motion boundary emphasised optical flow method for human action recognition | |
| Wang et al. | Fusion network for face-based age estimation | |
| Wang et al. | An improved simplified PCNN model for salient region detection |
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: 20180622 |
|
| RJ01 | Rejection of invention patent application after publication |