CN103544506B - A kind of image classification method and device based on convolutional neural networks - Google Patents
A kind of image classification method and device based on convolutional neural networks Download PDFInfo
- Publication number
- CN103544506B CN103544506B CN201310477564.XA CN201310477564A CN103544506B CN 103544506 B CN103544506 B CN 103544506B CN 201310477564 A CN201310477564 A CN 201310477564A CN 103544506 B CN103544506 B CN 103544506B
- Authority
- CN
- China
- Prior art keywords
- classification
- image
- layer
- sample data
- convolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of image classification method and device based on convolutional neural networks, methods described calculates the corresponding neural network weight of image of each classification by the image pattern of multiple classifications of reception input；The corresponding neural network weight of multiple classification images is distributed using hierarchy, each layer forms corresponding learning database；The category of test image sample data progress of input is handled and obtains corresponding one-dimensional characteristic description, the corresponding one-dimensional characteristic description of the category of test image sample data is subjected to feedforward study with the neural network weight in the learning database, so as to judge the category of test whether in the classification image learnt；Traditional convolutional neural networks are solved in quantitative limitation of classifying by the structure of layer distributed, prevent the mistake problem concerning study of convolutional neural networks, expand the classification capacity of convolutional neural networks in itself, and improving the accuracy of classification so that image classification algorithms have higher robustness in a new environment.
Description
Technical field
The present invention relates to image classification method technical field, more particularly to a kind of image based on convolutional neural networks
Sorting technique and device.
Background technology
Existing image classification method, conventional sorting technique supervised learning method for example neutral net and supporting vector
Machine etc., unsupervised learning method for example K mean cluster and nearest neighbour method etc..Traditional neural network belongs to supervised learning method, is logical
Cross and the neural network weight feature that learned object is got on the basis of existing sample learning is described, and according to knowing for being acquired
The classification for knowing the storehouse learnt to external world in environment makes a distinction judgement, but limited by the characteristics of objects knowledge that is learnt
Property, for the test object in extraneous changing environment, it is possible to can be beyond original existing learning knowledge scope so as to being learned
Identification classifying quality of the object under new complex environment is bad.
Further, due to the asymmetric mapping matrix that existing mode is used, the dimension of the matrix is determined based on the knot
The quantity that the convolutional neural networks of structure can be classified, when during classifying for more image category often
Neutral net is caused in itself because overabundance of data, some classifications learned can not be learnt and be recognized, causes nerve occur
It is common in network learning procedure " over fitting " problems, i.e., problem concerning study.
Therefore, prior art has yet to be improved and developed.
The content of the invention
The technical problem to be solved in the present invention is that the drawbacks described above for prior art is refreshing based on convolution there is provided one kind
Image classification method and device through network, it is intended to solve existing neural network image sorting technique classifying quality it is bad, hold
The problem of being also easy to produce study phenomenon.
The technical proposal for solving the technical problem of the invention is as follows：
A kind of image classification method based on convolutional neural networks, wherein, comprise the following steps：
A, the image pattern for receiving the multiple classifications inputted, return to the image sample data of each classification of input
One changes, and the image sample data after normalization is carried out into convolution, then using predetermined asymmetric mapping matrix to the figure after convolution
Decent notebook data is mapped, and the image sample data after mapping arrange is obtained corresponding one-dimensional characteristic and described, root
Describe to calculate the corresponding neural network weight of image of each classification according to the one-dimensional characteristic；
B, using hierarchy the corresponding neural network weight of multiple classification images is distributed, and each layer of distribution
Class number is the maximum differentiation class number determined according to the asymmetric mapping matrix, by multiple classification images in multiple layers
It is sequentially distributed, each layer forms corresponding learning database；
C, the category of test image sample data progress to input, which are handled, obtains corresponding one-dimensional characteristic description, is surveyed described
Try the corresponding one-dimensional characteristic description of classification image sample data and carry out feedforward study with the neural network weight in the learning database,
So as to judge the category of test whether in the classification image learnt.
The described image classification method based on convolutional neural networks, wherein, the step A is specifically included：
A1, the image pattern for receiving the multiple classifications inputted, the image sample data of each classification of input is carried out
Normalization；
A2, convolution is carried out to normalized image sample data, obtain corresponding multiple convolution results；
A3, to multiple convolution results carry out half it is down-sampled；
A4, the convolution results after down-sampled are carried out with convolution again, and using predetermined asymmetric mapping matrix to convolution
Image sample data afterwards is mapped；
A5, the down-sampled of half is carried out to the image sample data after mapping, and down-sampled result is carried out to arrange
One-dimensional characteristic to the image is described；
A6, to the one-dimensional characteristic describe subscribed it is down-sampled reduced after characteristic vector；
A7, the characteristic vector inputted into neutral net, into the learning process of neutral net, carry out feedforward study and BP
Feedback system is adjusted, and obtains corresponding neural network weight.
The described image classification method based on convolutional neural networks, wherein, the step B is specifically included：
B1, the image pattern to N number of classification, the maximum differentiation class number M determined according to the asymmetric mapping matrix；
B2, NUM layers of study classification are set up, NUM is（N/(M-1)）And be（N/(M-1)）Integer in orientation；
B3, for each layer, distinguish M classification of class number to build using maximum, complete N in NUM layers successively
The distribution of individual classification；
B4, each layer is learnt by convolutional neural networks and the learning database corresponding to each layer is formed.
The described image classification method based on convolutional neural networks, wherein, the step B also includes：
B5, when there is new classification to add, be put into last component layers and study re-started to this layer and form new
Learning database.
The described image classification method based on convolutional neural networks, wherein, the step C is specifically included：
C1, the category of test image sample data progress to input, which are handled, obtains corresponding one-dimensional characteristic description；
C2, into i-th layer, wherein i=1,2,3 ... NUM, i is initialized as 1, i.e., by category of test image sample data pair
The one-dimensional characteristic description answered carries out feedforward study with the neural network weight in i-th layer of corresponding learning database, if can distinguish
Go out classification and out then provide result, and exit the layer；If can not classify, then it is assumed that the category of test is not learnt in this layer
Classification；
C3, into next layer, i.e. i+1 layer repeats feedforward study, knot is provided if it can distinguish classification out
Really, and the layer is exited；If can not classify, then it is assumed that the category of test is the classification not learnt in this layer；Until NUM layers,
If not also being identified finally, the classification tested is not among the classification learnt.
A kind of image classification device based on convolutional neural networks, wherein, including：
Neural network weight acquisition module, the image pattern of multiple classifications for receiving input, to each of input
The image sample data of classification is normalized, and the image sample data after normalization is carried out into convolution, then using predetermined non-
Symmetrical mapping matrix is mapped the image sample data after convolution, and the image sample data progress after mapping is arranged
To the description of corresponding one-dimensional characteristic, described to calculate the corresponding neutral net of image of each classification according to the one-dimensional characteristic
Weights；
Hierarchical block, for being distributed using hierarchy to the corresponding neural network weight of multiple classification images, and
The class number of each layer of distribution is the maximum differentiation class number determined according to the asymmetric mapping matrix, by multiple classifications
Image is sequentially distributed in multiple layers, and each layer forms corresponding learning database；
Classification judge module, handles for the category of test image sample data progress to input and obtains corresponding one-dimensional spy
Description is levied, the corresponding one-dimensional characteristic description of the category of test image sample data is weighed with the neutral net in the learning database
Whether value carries out feedforward study, so as to judge the category of test in the classification image learnt.
The described image classification device based on convolutional neural networks, wherein, the neural network weight acquisition module tool
Body includes：
Normalization unit, the image pattern of multiple classifications for receiving input, by the image of each classification of input
Sample data is normalized；
First convolution unit, for carrying out convolution to normalized image sample data, obtains corresponding multiple convolution knots
Really；
First down-sampled unit, for carrying out the down-sampled of half to multiple convolution results；
Second convolution unit, for carrying out convolution again to the convolution results after down-sampled；
Map unit, for being mapped using predetermined asymmetric mapping matrix the image sample data after convolution；
Second down-sampled unit, for carrying out the down-sampled of half to the image sample data after mapping；
One-dimensional characteristic describes acquiring unit, and the one-dimensional characteristic that the image is obtained for arranging down-sampled result progress is retouched
State；
3rd down-sampled unit, for the one-dimensional characteristic describe subscribed it is down-sampled reduced after feature
Vector；
Unit, for the characteristic vector to be inputted into neutral net, into the learning process of neutral net, before progress
Feedback study and the regulation of BP feedback systems, obtain corresponding neural network weight.
The described image classification device based on convolutional neural networks, wherein, the hierarchical block is specifically included：
Number of plies determining unit, for the image pattern to N number of classification, the maximum determined according to the asymmetric mapping matrix
Distinguish class number M；NUM layers of study classification are set up, NUM is（N/(M-1)）And be（N/(M-1)）Integer in orientation；
Layer distributed unit, for for each layer, distinguishing M classification of class number to build using maximum, existing successively
The distribution of N number of classification is completed in NUM layers；
Learning database generation unit is right for being learnt and being formed each layer of institute to each layer by convolutional neural networks
The learning database answered.
The described image classification device based on convolutional neural networks, wherein, the hierarchical block also includes：
Updating block, for when there is new classification to add, being put into last component layers and re-starting to the layer
Habit forms new learning database.
A kind of image classification method and device based on convolutional neural networks provided by the present invention, are efficiently solved existing
Some neural network image sorting technique classifying qualities are bad, easy the problem of produced study, and its method, which passes through, receives input
Multiple classifications image pattern, calculate the corresponding neural network weight of image of each classification；Using hierarchy pair
The corresponding neural network weight of multiple classification images is distributed；Category of test image sample data progress to input is handled
To the description of corresponding one-dimensional characteristic, by the corresponding one-dimensional characteristic description of the category of test image sample data and the learning database
In neural network weight carry out feedforward study, so as to judge the category of test whether in the classification image learnt,
Design is improved by this body structure to convolutional neural networks, its original classification capacity is improved so that the class classified
It is not far longer than traditional convolutional neural networks classification capacities itself so that convolutional neural networks adapt to energy for extraneous change
Power more preferably, is solved the problem of the ability of image classification identification caused by new circumstances not known change declines and for volume well
Mistake problem concerning study of the product neutral net produced by during a large amount of classification sample learnings, its implementation is simple, passes through software
Realize, cost is relatively low.
Brief description of the drawings
The flow chart for the image classification method preferred embodiment based on convolutional neural networks that Fig. 1 provides for the present invention.
Nerve net is obtained in the image classification method preferred embodiment based on convolutional neural networks that Fig. 2 provides for the present invention
The neural network structure schematic diagram of network weights.
Hierarchy in the image classification method preferred embodiment based on convolutional neural networks that Fig. 3 provides for the present invention
Structural representation.
Category of test in the image classification method preferred embodiment based on convolutional neural networks that Fig. 4 provides for the present invention
Flow chart.
Increase study class in the image classification method preferred embodiment based on convolutional neural networks that Fig. 5 provides for the present invention
Other structural representation.
Hierarchy in the image classification method Application Example based on convolutional neural networks that Fig. 6 provides for the present invention
Structural representation.
The structured flowchart for the image classification device preferred embodiment based on convolutional neural networks that Fig. 7 provides for the present invention.
Embodiment
The present invention provides a kind of image classification method and device based on convolutional neural networks, for make the purpose of the present invention,
Technical scheme and advantage are clearer, clear and definite, and the present invention is described in more detail for the embodiment that develops simultaneously referring to the drawings.Should
Understand, the specific embodiments described herein are merely illustrative of the present invention, is not intended to limit the present invention.
Referring to Fig. 1, the image classification method preferred embodiment based on convolutional neural networks that Fig. 1 provides for the present invention
Flow chart, described image sorting technique comprises the following steps：
Step S100, the image pattern for receiving the multiple classifications inputted, to the image pattern number of each classification of input
According to being normalized, the image sample data after normalization is subjected to convolution, then using predetermined asymmetric mapping matrix to volume
Image sample data after product is mapped, and the image sample data after mapping carried out to arrange and obtain corresponding one-dimensional characteristic
Description, describes to calculate the corresponding neural network weight of image of each classification according to the one-dimensional characteristic；
Step S200, using hierarchy the corresponding neural network weight of multiple classification images is distributed, and it is each
The class number of layer distribution is that the maximum determined according to the asymmetric mapping matrix distinguishes class number, by multiple classification images
It is sequentially distributed in multiple layers, each layer forms corresponding learning database；
Step S300, the category of test image sample data progress to input, which are handled, obtains corresponding one-dimensional characteristic description,
The corresponding one-dimensional characteristic description of the category of test image sample data is carried out with the neural network weight in the learning database
Feedforward study, so as to judge the category of test whether in the classification image learnt.
Above-mentioned steps are described in detail below in conjunction with specific embodiment.
In the step s 100, neural study is carried out to the image pattern of multiple classifications, so as to draw the figure of each classification
As corresponding neural network weight.Specifically, the step S100 includes：S110, the image for receiving the multiple classifications inputted
Sample, the image sample data of each classification of input is normalized；S120, normalized image sample data is entered
Row convolution, obtains corresponding multiple convolution results；S130, to multiple convolution results carry out half it is down-sampled；S140, to drop adopt
Convolution results after sample carry out convolution again, and the image sample data after convolution is entered using predetermined asymmetric mapping matrix
Row mapping；S150, the down-sampled of half is carried out to the image sample data after mapping, and down-sampled result is carried out to arrange
One-dimensional characteristic to the image is described；S160, to the one-dimensional characteristic describe subscribed it is down-sampled reduced after spy
Levy vector；S170, the characteristic vector inputted into neutral net, into the learning process of neutral net, carry out feedforward study and
BP feedback systems are adjusted, and obtain corresponding neural network weight.
In practical application, referring to Fig. 2, the image classification method based on convolutional neural networks that Fig. 2 provides for the present invention
The neural network structure schematic diagram of neural network weight is obtained in preferred embodiment, as shown in Fig. 2 the figure of a certain class to input
Decent notebook data, that is, actual conditions are got it is corresponding study graphical representation be：,
WhereinThe feature description of sample image is represented,It is a two-dimensional matrix；It isCorresponding label, it is different for distinguishing；N is positive integer, represents the sum of sample.First to the data to inputIt is normalized, is normalized to [- 11], then to returning
One change after image carry out convolution using corresponding parameter, the purpose of convolution is to more preferably embody the feature of data, subtract
Few data processing total amount.
Specifically, the image of convolution can be 4 to may also be 6 can also be 8, need exist for according to reality
Situation is set to the number of convolution, can be set to C_num, and please continue to refer to Fig. 2, this sentences the figure obtained after four convolution
Illustrated as exemplified by, use size to enter for kernel_num*kernel_num convolution kernel to the image after normalization here
Row convolution.In practical application, kernel_num=3 here, can also be big to the convolution kernel of image according to specific practical application
It is small to be configured, but to ensure that the width of core is high equal, the convolution that the numerical value inside convolution kernel is constituted for the number of random generation 0 to 1
Core carries out convolution, because being the image of four convolution, needs to generate four groups of different random number kernel_num*kernel_
Num convolution kernel.The convolution kernel be a row and column identical matrix, and the numeral of each in matrix be between zero and one with
The numeral of machine generation.
Then, this four convolution results C1_1, C1_2, C1_3 and C1_4 are carried out with the down-sampled of half respectively to obtain pair
S1_1, S1_2, S1_3 and the S1_4 answered.If setting convolution number to be C_num, then corresponding convolution results are C1_1, C1_2,
C1_3 ... C1_C_num, and the down-sampled result that correspondence carries out half is S1_1, S1_2, S1_3 ... ... S1_C_num.
Please continue to refer to Fig. 2, between S1 layers to C2 layers, the data to S1 layers carry out convolution so as to obtain C2 layers again
Result, but the size of selected convolution kernel is less than being equal to kernel_num, while the data in convolution kernel at this
Also the number of 0 to 1 be randomly generated, that is to say, that the convolution block that C2 layers of convolution block will be with respect to C1 layers is small, and S1 and C2 layers
Between a predetermined asymmetrical mapping matrix is additionally used to be mapped.
The down-sampled of half is carried out to the image sample data after mapping again, that is, one is carried out to C2 layers of view data
Partly down-sampled, so as to obtain S2 layers.Image after down-sampled to each in S2 layers again, they according to image from a left side to
The right side, from top to bottom, arranges from first to last progress and obtains an one-dimensional characteristic of the image and describe S_feature,
The one-dimensional vector size be S2_num*S2_width*S2_high, wherein, S2_num be S2 layers in down-sampled rear image number
Measure, S2_width is the down-sampled rear pixel wide per pictures, S2_high is the down-sampled rear pixels tall per pictures.
Characteristic vector feature after suitably down-sampled is reduced is carried out to this feature S_feature vectors formed, wherein
Feature dimension is the 90% of S_feature dimensions, primarily to retaining the principal character of each convolved image.
Again using obtained one-dimensional vector feature as neutral net input, into the process of neutral net, the god
Result through network can be realized using traditional three-layer neural network, simply when the output layer of neutral net is built
Ensure that the dimension of output is consistent with the characteristic of image classification to be carried out.Used simultaneously for output layer in its place exported
RBF functions realize that this is neutral net common structure, be no longer described in detail herein.God is carried out in the image pattern to a certain classification
During through study in addition to carrying out feedforward study, BP is similarly carried out（Back Propagation, backpropagation）Instead
The study of feedback mode, so as to can guarantee that the last neural network weight learnt can well be weighed to all this progress that imitate
Value adjustment and description.
The present invention just multiple classification images can be carried out with feedforward study by above-mentioned steps and BP feedback systems are adjusted, and obtains
The corresponding neural network weight of each classification image, but be based on because the dimension of the predetermined asymmetric mapping matrix is determined
The quantity that the convolutional neural networks of the structure can be classified, that is, maximum differentiation class number.When to more image
During classification is classified, often cause neutral net in itself because overabundance of data, can not to some classifications learned
Learnt and recognized, be common mistake problem concerning study during neural network learning, therefore, the present invention is by multiple classification figures
Picture and corresponding neural network weight layer distributed, to solve the mistake problem concerning study under method.
Specifically, the step S200 includes：S210, the image pattern to N number of classification, according to the asymmetric mapping
The maximum differentiation class number M that matrix is determined；S220, NUM layers of study classification are set up, NUM is（N/(M-1)）And be（N/(M-
1)）Integer in orientation；S230, for each layer, distinguish M classification of class number to build using maximum, successively at NUM layers
In complete the distribution of N number of classification；S240, each layer is learnt by convolutional neural networks and formed each layer institute it is right
The learning database answered.Wherein, N, M, NUM are positive integer.
In practical application, the classification that the unsymmetrical matrix, which determines the convolutional network, to be distinguished distinguishes classification to be maximum
Number M, and the classification number learnt to the image pattern of multiple classifications is N, if N≤M, shows that the class number of study is small
Class number is distinguished in maximum；If N>M, shows that the image separating capacity of current convolutional network only has M, and has N-M image
Classification can not be learnt and be recognized.
N number of classification is being kept ensuring the mistake problem concerning study of neutral net not in the case that learning sample is constant in order to realize
Produce, referring to Fig. 3, dividing in the image classification method preferred embodiment based on convolutional neural networks that Fig. 3 provides for the present invention
The structural representation of Rotating fields, as shown in figure 3, the multiple classification images of layer distributed of the present invention and corresponding neutral net power
Value, sets up NUM layers of study classification, and NUM is（N/(M-1)）And be（N/(M-1)）Integer in orientation, mould is set up it is determined that having descended
The number of plies of formula classification is NUM layers；For each layer, built using M classification, complete N number of classification in NUM layers successively
Distribution, as shown in Figure 3；For each layer, using convolutional neural networks the study to complete each layer and form each layer
Corresponding learning database, adjusts especially by feedforward study and BP feedback systems, obtains the corresponding nerve net of each classification image
Network weights, the learning database is different classes of in this layer and its corresponding neural network weight, and for example first layer passes through volume
Accumulate the method for neutral net to train obtained learning database for W_1, W_2 ... ... W_M, NUM layers pass through convolutional neural networks
Learning method is to train obtained learning database：W_ [(NUM-1) * M+1], W_ [(NUM-1) * M+1] ... ... W_ [N].
Then in step S300, to the category of test image sample data of input by step S100 handling equally
Should be specifically the processing by step S110 ~ S150 so as to obtain category of test figure to the description of corresponding one-dimensional characteristic
Described and the study as the description of corresponding one-dimensional characteristic, then by the corresponding one-dimensional characteristic of the category of test image sample data
Whether the neural network weight in storehouse carries out feedforward study, so as to judge the category of test in the classification image learnt
In.
Referring to Fig. 4, in the image classification method preferred embodiment based on convolutional neural networks that Fig. 4 provides for the present invention
The flow chart of category of test, as illustrated, specifically, the step S300 includes：S310, the category of test image to input
Sample data progress, which is handled, obtains corresponding one-dimensional characteristic description；S320, into i-th layer, wherein i=1,2,3 ... NUM, at the beginning of i
Beginning turns to 1, i.e., by the corresponding one-dimensional characteristic of category of test image sample data describe with i-th layer of corresponding learning database
Neural network weight carries out feedforward study, result is provided if it can distinguish classification out, and exit the layer；If can not classify,
It is the classification not learnt in this layer then to think the category of test；S330, into next layer, i.e. i+1 layer repeats feedforward
Study, provides result, and exit the layer if it can distinguish classification out；If can not classify, then it is assumed that the category of test is
The classification not learnt in this layer；Until NUM layers, if not also being identified finally, the classification tested is not in institute
Among the classification of study.
In practical application, when there is category of test to need differentiation, being introduced into i-th layer（I=1,2, wherein 3 ... NUM, i
It is initialized as 1）, result is provided if it can distinguish classification out, and exit the layer, if can not classify, then it is assumed that in being the layer
The classification not learnt, into next layer（i=i+1）, until the class of last layer has been recognized, if not also being identified finally
Come, then just illustrate tested classification not among the class learnt.
The present invention is using the convolutional neural networks with hierarchy, and the technology is relative to traditional neutral net logarithm
According to more preferable abstract analysis ability, and for traditional classification, deeper information can be got in the extraction of feature.
The present invention is by carrying out Curve guide impeller on convolutional neural networks this body structures, improving its original classification capacity so that volume
Product neutral net more preferably, can realize a certain degree of on-line study, reduce conduct itself for extraneous change adaptability
The limitation of supervised learning, improves the robustness in the classification ability and assorting process to image classification, while also improving
Adaptability of this method in circumstances not known, has more accurately separating capacity for unknown situation.The figure that the present invention is provided
As sorting algorithm has good robustness in a new environment, cause while introducing Stratified Strategy in the class with identical characteristic
Not middle discrimination is increased, it is therefore prevented that " the over fitting " of neutral net（Cross study）Problem.
Preferably, referring to Fig. 5, Fig. 5 is preferable for the image classification method based on convolutional neural networks that the present invention is provided
The structural representation of increase study classification in embodiment, as illustrated, the step S200 also includes：S250, when there is new class
When not adding, be put into last component layers and to this layer re-start study form new learning database.Specifically, it is new when having
Classification add when, be directly placed into the sample of last component layers and only to the sample of this layer and re-start study, i.e., it is right again
The sample of this layer carries out feedforward study and the regulation of BP feedback systems, obtains the corresponding god of each classification image in last component layers
Through network weight, and in one group of new learning database of this layer formation；A new classification, total classification are such as added in NUM layers
For N+1, corresponding new neural network weight has been obtained by the study of the convolutional neural networks to NUM layers：W’_[(NUM-
1) * M+1], W ' _ [(NUM-1) * M+1] ... ... W ' _ [N], W ' _ [N+1].Wherein W ' is to discriminate between the nerve net with being learnt before
Network weights W.
Below with a specific Application Example, the detailed process of the present invention is described in detail.
Also referring to Fig. 1, Fig. 2 and Fig. 6, wherein, the image based on convolutional neural networks point that Fig. 6 provides for the present invention
The structural representation of hierarchy in class method Application Example, specifically, the following is the image point provided using the present invention
Class method is identified to commodity Logo icons in external environment.By being sampled to commodity Logo, 45 kinds are got
Corresponding commodity Logo icons, wherein the quantity of each icon is 1000, this 1000 include Logo under various circumstances
Image.Differentiation is marked using numeral to this 45 kinds of labels, 1,2,3 is respectively labeled as ... 45, is expressed as,Its
Middle i=1,2 ... ..45 represents species, j=1,2,3 ... 1000 represent be a certain class image number.WhereinRepresent the i-th class
In j-th of image,Represent the label of the i-th class（=1,2,3 ... 1000）.Then workIt is incoming for sample
Learnt in convolutional neural networks and obtain corresponding learning database, i.e. neural network weight.Image pattern is big during practical application
Small is 54*131.First incoming sample data is normalized [- 1,1], sets C1_num=4 pair image to carry out convolution, obtains
Get image C1_1, C1_2, C1_3 and C1_4.C1_num=4 are obtained by experiment and practical application.Kernel_ is set again
Num=3, that is, use size to carry out convolution for the convolution kernel that the number that the numerical value inside 3*3 is random generation 0 to 1 is constituted.Because being
Four data, so the 3*3 convolution kernels of four different random numbers of generation.Kernel_num=3, being should in experiment and reality
Analyze what is obtained with middle.Again this four convolution results C1_1, C1_2, C1_3, C1_4 are carried out with the down-sampled of half respectively to obtain
To corresponding S1_1, S1_2, S1_3, S1_4.Then convolution is carried out to S1 layers using same convolution method and obtains C2 layers, also
It is to produce different random number to constitute convolution kernel but convolution kernel size is smaller than the convolution kernel of first time convolution, can for example uses
2*2 convolution kernels, and S1 to C2 layers also map generation using a kind of asymmetric mapping matrix, and the asymmetric mapping matrix is
Prior art, is illustrated below.
It is for unsymmetrical matrix as shown in table 1, here and combine Fig. 2 and understand：
Table 1
It can see from table 1, S1_1, S1_2 that the leftmost row of form correspond in Fig. 2 labeled as S1_1-S1_4,
X tables are drawn between S1 and C2 in S1_3, S1_4, and the C2 that C2_1 to C2_11 above form then corresponds in Fig. 2 that row, table
It is connection to show corresponding ranks, such as S1_1 is corresponding with C2_1 ranks in table get off to have an X if represent S1_1 in Fig. 2 and
There is mapping to connect between C2_1, do not draw X and then represent to be not connected to.
Then, S2 layers are the down-sampled obtained results for being carried out on the basis of C2 layers half, for each in S2 layers
It is individual it is down-sampled after image, they according to image from left to right, from top to bottom, arranged from first to last progress
An one-dimensional characteristic to the image describes S_feature, the one-dimensional vector size be S2_num*S2_width*S2_high=
11*30*14=4620, due to S2 layers it is down-sampled after obtained image be 11, a width of the 30 of every image, a high position 14, so always
Number is 4620.To the feature that is formed by down-sampled its 90% data of reservation, the characteristic vector after being reduced is：4158,
Then the image thinks that 4158 input in neutral net as feature, carries out feedforward study, so as to obtain corresponding neutral net power
Value.Also utilize existing sample label, pass through BP（Back Propagation）Feedback system is to obtained neutral net
Weights are adjusted, so as to can guarantee that the last neural network weight learnt can well be weighed to all this progress that imitate
Value adjustment and description, so as to ensure that the sample learning obtained under obtained final neural network weight is exported and actual given defeated
It is consistent to go out, and this is prior art, is no longer described in detail herein.Specifically, exactly obtained by certain pictures under a certain class image
To a neural network weight, feedback regulation is then carried out by the neural network weight of other pictures under such image, so that
Obtain most showing the neural network weight of such image.Ensure neutral net during neural network structure is set simultaneously
Actual study classification sum is output as, but because the limitation of network structure in itself is so that the maximal dimension of output is 30,30
By being determined using matrix mapping method, specifically, the dimension of the asymmetric mapping matrix is determined based on the knot
The quantity that the convolutional neural networks of structure can be classified.
Because the asymmetric mapping matrix determines the classification capacity of the convolutional neural networks, at most in 30 classifications
Make a distinction, in order to solve traditional convolutional neural networks in quantitative limitation of classifying, please continue to refer to Fig. 6, for given
45 classifications, the classification that convolutional network can be distinguished under asymmetric mapping matrix be M=30, N number of classification is classified, N
=45, set up NUM study classification：NUM=（N/(M-1)）=（45/(30-1)）=1.5517 numbers that round up are：2.Such as Fig. 6 institutes
Show, 45 classifications are layered according to the identification classification of convolutional neural networks in itself, are divided into two layers, first layer is 30 classes
Not, the second layer is 15 classifications, obtains corresponding learning database by every layer of the study of convolutional neural networks, is：First layer
Practising storehouse is：W_1, W_2 ... ... W_30；The learning database of the second layer is：W_31, W_32 ... ... W_45；When there is category of test entrance
When, using the method described in Fig. 4, the category of test image sample data progress of input is handled and obtains corresponding one-dimensional characteristic and retouches
State, the corresponding one-dimensional characteristic description of the category of test image sample data is entered with the neural network weight in the learning database
Row feedforward study, if identifying the category in first layer, provides result, if can not recognize, into the second layer, if the second layer
It can recognize that and also provide result, if can not provide, illustrate the category of test not among the class learnt.
Traditional method can only often be identified during business icon is recognized for a small amount of sample class,
The reason for often should be convolutional neural networks itself when sample size is increased fails to be enlarged in classification.The present invention is logical
Cross and introduce a kind of new frame mode, be i.e. the structure of layer distributed solves traditional convolutional neural networks in quantitative office of classifying
Sex-limited, structure of the invention expands the classification capacity of convolutional neural networks in itself, and improves the accuracy of classification, realizes
To the study of the samples of a large amount of classifications under convolutional neural networks, and it can also keep original knowledge well in practical situations both
Other effect.
Based on the above-mentioned image classification method based on convolutional neural networks, convolutional Neural is based on present invention also offers one kind
The image classification device of network, referring to Fig. 7, the image classification device based on convolutional neural networks that Fig. 7 provides for the present invention
The structured flowchart of preferred embodiment, as shown in fig. 7, described device includes：
Neural network weight acquisition module 10, the image pattern of multiple classifications for receiving input, to each of input
The image sample data of individual classification is normalized, and the image sample data after normalization is carried out into convolution, then using predetermined
Asymmetric mapping matrix is mapped the image sample data after convolution, and the image sample data after mapping is arranged
Corresponding one-dimensional characteristic description is obtained, is described to calculate the corresponding nerve net of image of each classification according to the one-dimensional characteristic
Network weights；Specifically as described in step S100.
Hierarchical block 20, for being distributed using hierarchy to the corresponding neural network weight of multiple classification images,
And the class number of each layer of distribution is the maximum differentiation class number determined according to the asymmetric mapping matrix, by multiple classes
Other image is sequentially distributed in multiple layers, and each layer forms corresponding learning database；Specifically as described in step S200.
Classification judge module 30, handles for the category of test image sample data progress to input and obtains corresponding one-dimensional
Feature is described, by the corresponding one-dimensional characteristic description of the category of test image sample data and the neutral net in the learning database
Whether weights carry out feedforward study, so as to judge the category of test in the classification image learnt；Specific such as step
Described in S300.
Preferably, the neural network weight acquisition module 10 is specifically included：
Normalization unit, the image pattern of multiple classifications for receiving input, by the image of each classification of input
Sample data is normalized；
First convolution unit, for carrying out convolution to normalized image sample data, obtains corresponding multiple convolution knots
Really；
First down-sampled unit, for carrying out the down-sampled of half to multiple convolution results；
Second convolution unit, for carrying out convolution again to the convolution results after down-sampled；
Map unit, for being mapped using predetermined asymmetric mapping matrix the image sample data after convolution；
Second down-sampled unit, for carrying out the down-sampled of half to the image sample data after mapping；
One-dimensional characteristic describes acquiring unit, and the one-dimensional characteristic that the image is obtained for arranging down-sampled result progress is retouched
State；
3rd down-sampled unit, for the one-dimensional characteristic describe subscribed it is down-sampled reduced after feature
Vector；
Unit, for the characteristic vector to be inputted into neutral net, into the learning process of neutral net, before progress
Feedback study and the regulation of BP feedback systems, obtain corresponding neural network weight.
Preferably, the hierarchical block 20 is specifically included：
Number of plies determining unit, for the image pattern to N number of classification, the maximum determined according to the asymmetric mapping matrix
Distinguish class number M；NUM layers of study classification are set up, NUM is（N/(M-1)）And be（N/(M-1)）Integer in orientation；
Layer distributed unit, for for each layer, distinguishing M classification of class number to build using maximum, existing successively
The distribution of N number of classification is completed in NUM layers；
Learning database generation unit is right for being learnt and being formed each layer of institute to each layer by convolutional neural networks
The learning database answered.
Preferably, the hierarchical block 20 also includes：
Updating block, for when there is new classification to add, being put into last component layers and re-starting to the layer
Habit forms new learning database.
In summary, the present invention is provided a kind of image classification method and device based on convolutional neural networks, the side
Method calculates the corresponding neural network weight of image of each classification by the image pattern of multiple classifications of reception input；
The corresponding neural network weight of multiple classification images is distributed using hierarchy, each layer forms corresponding learning database；
The category of test image sample data progress of input is handled and obtains corresponding one-dimensional characteristic description, by the category of test image
The corresponding one-dimensional characteristic description of sample data carries out feedforward study with the neural network weight in the learning database, so as to judge
Whether the category of test is in the classification image learnt；Traditional convolutional neural networks are solved by the structure of layer distributed
In the quantitative limitation of classifying, it is therefore prevented that the mistake problem concerning study of convolutional neural networks, convolutional neural networks are expanded in itself
Classification capacity, and improve the accuracy of classification so that image classification algorithms have higher robustness in a new environment, real
The study under convolutional neural networks to the sample of a large amount of classifications is showed, and also can keep original well in practical situations both
Recognition effect, its implementation is simple, is realized by software, cost is relatively low, can be widely applied to digital household appliances, advertisement and push away
Give, the application field such as data mining or image classification.
It should be appreciated that the application of the present invention is not limited to above-mentioned citing, for those of ordinary skills, can
To be improved or converted according to the above description, all these modifications and variations should all belong to the guarantor of appended claims of the present invention
Protect scope.
Claims (9)
1. a kind of image classification method based on convolutional neural networks, it is characterised in that comprise the following steps：
A, the image pattern for receiving the multiple classifications inputted, normalizing is carried out to the image sample data of each classification of input
Change, the image sample data after normalization is subjected to convolution, then using predetermined asymmetric mapping matrix to the image after convolution
Sample data is mapped, and the image sample data after mapping arrange is obtained corresponding one-dimensional characteristic and described, according to
The one-dimensional characteristic description calculates the corresponding neural network weight of image of each classification；
B, using hierarchy the corresponding neural network weight of multiple classification images is distributed, and the classification of each layer of distribution
Number be according to the asymmetric mapping matrix determine it is maximum distinguish class number, by multiple classification images in multiple layers successively
Distribution, each layer forms corresponding learning database；
C, the category of test image sample data progress to input, which are handled, obtains corresponding one-dimensional characteristic description, by the test class
Neural network weight in the other corresponding one-dimensional characteristic description of image sample data and the learning database carries out feedforward study, so that
Judge the category of test whether in the classification image learnt.
2. the image classification method according to claim 1 based on convolutional neural networks, it is characterised in that the step A
Specifically include：
A1, the image pattern for receiving the multiple classifications inputted, normalizing is carried out by the image sample data of each classification of input
Change；
A2, convolution is carried out to normalized image sample data, obtain corresponding multiple convolution results；
A3, to multiple convolution results carry out half it is down-sampled；
A4, the convolution results after down-sampled are carried out with convolution again, and using predetermined asymmetric mapping matrix to convolution after
Image sample data is mapped；
A5, the down-sampled of half is carried out to the image sample data after mapping, and down-sampled result is carried out to arrange this
The one-dimensional characteristic description of image；
A6, the one-dimensional characteristic is described to carry out it is predetermined it is down-sampled reduced after characteristic vector；
A7, the characteristic vector inputted into neutral net, into the learning process of neutral net, carry out feedforward study and BP feedbacks
Mode is adjusted, and obtains corresponding neural network weight.
3. the image classification method according to claim 1 based on convolutional neural networks, it is characterised in that the step B
Specifically include：
B1, the image pattern to N number of classification, the maximum differentiation class number M determined according to the asymmetric mapping matrix；
B2, NUM layers of study classification are set up, NUM is（N/(M-1)）Integer in orientation；
B3, for each layer, distinguish M classification of class number to build using maximum, complete N number of class in NUM layers successively
Other distribution；
B4, each layer is learnt by convolutional neural networks and the learning database corresponding to each layer is formed.
4. the image classification method according to claim 1 based on convolutional neural networks, it is characterised in that the step B
Also include：
B5, when there is new classification to add, be put into last component layers and study re-started to this layer and form new study
Storehouse.
5. the image classification method according to claim 3 based on convolutional neural networks, it is characterised in that the step C
Specifically include：
C1, the category of test image sample data progress to input, which are handled, obtains corresponding one-dimensional characteristic description；
C2, into i-th layer, wherein i=1,2,3 ... NUM, i is initialized as 1, i.e., category of test image sample data is corresponding
One-dimensional characteristic describes to carry out feedforward study with the neural network weight in i-th layer of corresponding learning database, if class can be distinguished
Result Chu Lai not be then provided, and exits the layer；If can not classify, then it is assumed that the category of test is the class not learnt in this layer
Not；
C3, into next layer, i.e. i+1 layer repeats feedforward study, result is provided if it can distinguish classification out, and
Exit the layer；If can not classify, then it is assumed that the category of test is the classification not learnt in this layer；Until NUM layers, if most
Also it is not identified afterwards, then the classification tested is not among the classification learnt.
6. a kind of image classification device based on convolutional neural networks, it is characterised in that including：
Neural network weight acquisition module, the image pattern of multiple classifications for receiving input, to each classification of input
Image sample data be normalized, the image sample data after normalization is subjected to convolution, then using predetermined asymmetric
Mapping matrix is mapped the image sample data after convolution, and the image sample data after mapping arrange to obtain pair
The one-dimensional characteristic description answered, is weighed according to the corresponding neutral net of image that the one-dimensional characteristic describes to calculate each classification
Value；
Hierarchical block, for being distributed using hierarchy to the corresponding neural network weight of multiple classification images, and it is each
The class number of layer distribution is that the maximum determined according to the asymmetric mapping matrix distinguishes class number, by multiple classification images
It is sequentially distributed in multiple layers, each layer forms corresponding learning database；
Classify judge module, handled for the category of test image sample data progress to input and obtain corresponding one-dimensional characteristic and retouch
State, the corresponding one-dimensional characteristic description of the category of test image sample data is entered with the neural network weight in the learning database
Row feedforward study, so as to judge the category of test whether in the classification image learnt.
7. the image classification device according to claim 6 based on convolutional neural networks, it is characterised in that the nerve net
Network weights acquisition module is specifically included：
Normalization unit, the image pattern of multiple classifications for receiving input, by the image pattern of each classification of input
Data are normalized；
First convolution unit, for carrying out convolution to normalized image sample data, obtains corresponding multiple convolution results；
First down-sampled unit, for carrying out the down-sampled of half to multiple convolution results；
Second convolution unit, for carrying out convolution again to the convolution results after down-sampled；
Map unit, for being mapped using predetermined asymmetric mapping matrix the image sample data after convolution；
Second down-sampled unit, for carrying out the down-sampled of half to the image sample data after mapping；
One-dimensional characteristic describes acquiring unit, for carrying out arranging the one-dimensional characteristic for obtaining image description to down-sampled result；
3rd down-sampled unit, for the one-dimensional characteristic is described to carry out it is predetermined it is down-sampled reduced after feature to
Amount；
Unit, for the characteristic vector to be inputted into neutral net, into the learning process of neutral net, carries out feedforward
Practise and the regulation of BP feedback systems, obtain corresponding neural network weight.
8. the image classification device according to claim 6 based on convolutional neural networks, it is characterised in that the layering mould
Block is specifically included：
Number of plies determining unit, for the image pattern to N number of classification, the maximum differentiation determined according to the asymmetric mapping matrix
Class number M；NUM layers of study classification are set up, NUM is（N/(M-1)）Integer in orientation；
Layer distributed unit, for for each layer, distinguishing M classification of class number to build using maximum, successively at NUM layers
In complete the distribution of N number of classification；
Learning database generation unit, for each layer to be learnt and formed corresponding to each layer by convolutional neural networks
Learning database.
9. the image classification device according to claim 6 based on convolutional neural networks, it is characterised in that the layering mould
Block also includes：
Updating block, for when there is new classification to add, being put into last component layers and re-starting study shape to the layer
Cheng Xin learning database.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310477564.XA CN103544506B (en) | 2013-10-12 | 2013-10-12 | A kind of image classification method and device based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310477564.XA CN103544506B (en) | 2013-10-12 | 2013-10-12 | A kind of image classification method and device based on convolutional neural networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103544506A CN103544506A (en) | 2014-01-29 |
| CN103544506B true CN103544506B (en) | 2017-08-08 |
Family
ID=49967941
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310477564.XA Active CN103544506B (en) | 2013-10-12 | 2013-10-12 | A kind of image classification method and device based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103544506B (en) |
Families Citing this family (31)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103984959B (en) * | 2014-05-26 | 2017-07-21 | 中国科学院自动化研究所 | A kind of image classification method based on data and task-driven |
| CN105224963B (en) * | 2014-06-04 | 2019-06-07 | 华为技术有限公司 | The method and terminal of changeable deep learning network structure |
| CN104036474B (en) * | 2014-06-12 | 2017-12-19 | 厦门美图之家科技有限公司 | A kind of Automatic adjustment method of brightness of image and contrast |
| CN104182735A (en) * | 2014-08-18 | 2014-12-03 | 厦门美图之家科技有限公司 | Training optimization pornographic picture or video detection method based on convolutional neural network |
| CN104268521A (en) * | 2014-09-23 | 2015-01-07 | 朱毅 | Image recognition method based on convolutional neural network in non-finite category |
| CN104244113B (en) * | 2014-10-08 | 2017-09-22 | 中国科学院自动化研究所 | A kind of video abstraction generating method based on depth learning technology |
| US10387773B2 (en) * | 2014-10-27 | 2019-08-20 | Ebay Inc. | Hierarchical deep convolutional neural network for image classification |
| CN104463243B (en) * | 2014-12-01 | 2017-09-29 | 中科创达软件股份有限公司 | Sex-screening method based on average face feature |
| WO2016090520A1 (en) * | 2014-12-10 | 2016-06-16 | Xiaogang Wang | A method and a system for image classification |
| CN105843816A (en) * | 2015-01-15 | 2016-08-10 | 阿里巴巴集团控股有限公司 | Method and device for determining display information of picture |
| CN105989330A (en) * | 2015-02-03 | 2016-10-05 | 阿里巴巴集团控股有限公司 | Picture detection method and apparatus |
| CN104732249A (en) * | 2015-03-25 | 2015-06-24 | 武汉大学 | Deep learning image classification method based on popular learning and chaotic particle swarms |
| CN106156846B (en) * | 2015-03-30 | 2018-12-25 | 日本电气株式会社 | The treating method and apparatus of convolutional neural networks feature |
| CN105068644A (en) * | 2015-07-24 | 2015-11-18 | 山东大学 | Method for detecting P300 electroencephalogram based on convolutional neural network |
| CN105512676A (en) * | 2015-11-30 | 2016-04-20 | 华南理工大学 | Food recognition method at intelligent terminal |
| CN105512681A (en) * | 2015-12-07 | 2016-04-20 | 北京信息科技大学 | Method and system for acquiring target category picture |
| CN105426883B (en) * | 2015-12-25 | 2019-03-01 | 中国科学院深圳先进技术研究院 | The method and device that visual classification quickly identifies |
| CN105821538B (en) * | 2016-04-20 | 2018-07-17 | 广州视源电子科技股份有限公司 | Spun yarn breakage detection method and system |
| CN106203327B (en) * | 2016-07-08 | 2019-04-19 | 清华大学 | Lung tumor identification system and method based on convolutional neural networks |
| CN106529564B (en) * | 2016-09-26 | 2019-05-31 | 浙江工业大学 | A kind of food image automatic classification method based on convolutional neural networks |
| CN106548201B (en) * | 2016-10-31 | 2020-07-21 | 北京小米移动软件有限公司 | Training method of convolutional neural network, image recognition method and device |
| EP3610240A4 (en) * | 2017-04-13 | 2020-04-22 | Siemens Healthcare Diagnostics Inc. | Methods and apparatus for hiln characterization using convolutional neural network |
| CN107123027B (en) * | 2017-04-28 | 2021-06-01 | 广东工业大学 | Deep learning-based cosmetic recommendation method and system |
| FR3067496B1 (en) * | 2017-06-12 | 2021-04-30 | Inst Mines Telecom | DESCRIPTOR LEARNING PROCESS FOR THE DETECTION AND LOCATION OF OBJECTS IN A VIDEO |
| CN107451996A (en) * | 2017-07-26 | 2017-12-08 | 广州慧扬健康科技有限公司 | Deep learning training system applied to cutaneum carcinoma identification |
| CN108765035A (en) * | 2018-06-19 | 2018-11-06 | 北京奇艺世纪科技有限公司 | A kind of advertising image feature extracting method, device and electronic equipment |
| CN111368852A (en) * | 2018-12-26 | 2020-07-03 | 沈阳新松机器人自动化股份有限公司 | Article identification and pre-sorting system and method based on deep learning and robot |
| CN111401396B (en) * | 2019-01-03 | 2023-04-18 | 阿里巴巴集团控股有限公司 | Image recognition method and device |
| TWI688276B (en) * | 2019-03-06 | 2020-03-11 | 明新科技大學 | A method of characterizing a color image sensor by probabilistic neural network |
| CN109978058B (en) * | 2019-03-28 | 2023-08-04 | 腾讯科技（深圳）有限公司 | Method, device, terminal and storage medium for determining image classification |
| CN111950723B (en) * | 2019-05-16 | 2024-05-21 | 武汉Tcl集团工业研究院有限公司 | Neural network model training method, image processing method, device and terminal equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101681449A (en) * | 2007-06-15 | 2010-03-24 | 佳能株式会社 | Calculation processing apparatus and method |
| CN103279759A (en) * | 2013-06-09 | 2013-09-04 | 大连理工大学 | Vehicle front trafficability analyzing method based on convolution nerve network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8504361B2 (en) * | 2008-02-07 | 2013-08-06 | Nec Laboratories America, Inc. | Deep neural networks and methods for using same |
| US8582807B2 (en) * | 2010-03-15 | 2013-11-12 | Nec Laboratories America, Inc. | Systems and methods for determining personal characteristics |
-
2013
- 2013-10-12 CN CN201310477564.XA patent/CN103544506B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101681449A (en) * | 2007-06-15 | 2010-03-24 | 佳能株式会社 | Calculation processing apparatus and method |
| CN103279759A (en) * | 2013-06-09 | 2013-09-04 | 大连理工大学 | Vehicle front trafficability analyzing method based on convolution nerve network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103544506A (en) | 2014-01-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103544506B (en) | A kind of image classification method and device based on convolutional neural networks | |
| CN111368896B (en) | Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network | |
| CN109977943A (en) | A kind of images steganalysis method, system and storage medium based on YOLO | |
| CN104484681B (en) | Hyperspectral Remote Sensing Imagery Classification method based on spatial information and integrated study | |
| CN107341518A (en) | A kind of image classification method based on convolutional neural networks | |
| CN106203327A (en) | Lung tumor identification system and method based on convolutional neural networks | |
| CN107683469A (en) | A kind of product classification method and device based on deep learning | |
| CN109583483A (en) | A kind of object detection method and system based on convolutional neural networks | |
| CN111523521A (en) | Remote sensing image classification method for double-branch fusion multi-scale attention neural network | |
| CN110210492B (en) | Stereo image visual saliency detection method based on deep learning | |
| CN102385592B (en) | Image concept detection method and device | |
| CN107209942A (en) | Method for checking object and image indexing system | |
| CN108960260B (en) | Classification model generation method, medical image classification method and medical image classification device | |
| CN110211127B (en) | Image partition method based on bicoherence network | |
| CN109711401A (en) | A kind of Method for text detection in natural scene image based on Faster Rcnn | |
| US11694428B1 (en) | Method for detecting Ophiocephalus argus cantor under intra-class occulusion based on cross-scale layered feature fusion | |
| CN112163520B (en) | MDSSD face detection method based on improved loss function | |
| CN106844518A (en) | A kind of imperfect cross-module state search method based on sub-space learning | |
| CN111931867B (en) | New coronary pneumonia X-ray image classification method and system based on lightweight model | |
| CN109086753A (en) | Traffic sign recognition method, device based on binary channels convolutional neural networks | |
| CN101980251A (en) | Remote sensing classification method for binary tree multi-category support vector machines | |
| CN110245747A (en) | Image processing method and device based on full convolutional neural networks | |
| CN111739037B (en) | Semantic segmentation method for indoor scene RGB-D image | |
| CN113642445A (en) | Hyperspectral image classification method based on full convolution neural network | |
| CN111414930B (en) | Deep learning model training method and device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |