CN106919980A - A kind of increment type target identification system based on neuromere differentiation - Google Patents
A kind of increment type target identification system based on neuromere differentiation Download PDFInfo
- Publication number
- CN106919980A CN106919980A CN201710055128.1A CN201710055128A CN106919980A CN 106919980 A CN106919980 A CN 106919980A CN 201710055128 A CN201710055128 A CN 201710055128A CN 106919980 A CN106919980 A CN 106919980A
- Authority
- CN
- China
- Prior art keywords
- neuromere
- sample
- hidden layer
- ganglionic
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000004069 differentiation Effects 0.000 title claims abstract description 31
- 230000000574 ganglionic effect Effects 0.000 claims abstract description 37
- 230000004913 activation Effects 0.000 claims abstract description 15
- 230000006907 apoptotic process Effects 0.000 claims abstract description 7
- 230000000875 corresponding effect Effects 0.000 claims abstract description 3
- 230000007935 neutral effect Effects 0.000 claims abstract description 3
- 239000000284 extract Substances 0.000 claims abstract 2
- 238000000034 method Methods 0.000 claims description 36
- 238000012549 training Methods 0.000 claims description 31
- 238000010606 normalization Methods 0.000 claims description 11
- 238000004422 calculation algorithm Methods 0.000 claims description 10
- 230000010354 integration Effects 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 7
- 230000008859 change Effects 0.000 claims description 6
- 230000000694 effects Effects 0.000 claims description 6
- 230000001174 ascending effect Effects 0.000 claims description 5
- 230000000903 blocking effect Effects 0.000 claims description 5
- 210000005036 nerve Anatomy 0.000 claims description 5
- 238000013135 deep learning Methods 0.000 claims description 4
- 230000001965 increasing effect Effects 0.000 claims description 4
- 241001465754 Metazoa Species 0.000 claims description 3
- 230000006978 adaptation Effects 0.000 claims description 3
- 238000010586 diagram Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 2
- 230000001537 neural effect Effects 0.000 claims 1
- 230000003252 repetitive effect Effects 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 239000011159 matrix material Substances 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000007711 solidification Methods 0.000 description 2
- 230000008023 solidification Effects 0.000 description 2
- 241000282327 Felis silvestris Species 0.000 description 1
- 241000983594 Litocranius walleri Species 0.000 description 1
- 241000238413 Octopus Species 0.000 description 1
- 241000289371 Ornithorhynchus anatinus Species 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000005056 compaction Methods 0.000 description 1
- 238000010924 continuous production Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 239000010931 gold Substances 0.000 description 1
- 229910052737 gold Inorganic materials 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000002203 pretreatment Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
- 230000002087 whitening effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- 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
- G06F18/211—Selection of the most significant subset of features
- G06F18/2111—Selection of the most significant subset of features by using evolutionary computational techniques, e.g. genetic algorithms
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Physiology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
A kind of increment type target identification system based on neuromere differentiation, is made up of low hidden layer, periopticon, hidden layer high and grader, and periopticon is located between low hidden layer and hidden layer high;The neuromere of periopticon extracts sample characteristics rule, by ganglionic activation and differentiation, corresponding independent feature set, the feature memory of shape paired samples is formed in hidden layer high;The variable amounts of the feature set of hidden layer high;Newly-increased neuromere forms new memory to new samples, adaptively updates the feature set of hidden layer high, realizes the target identification of increment type;Described neuromere refers to that, when input sample characteristic parameter is more than activation threshold, neuromere is activated, and the neuromere forms independent feature set in hidden layer high for characterizing one group of neutral net node of the similar sample of the regularity of distribution;When the ganglionic activity being activated is less than threshold value, the neuromere is by apoptosis;The different sample of feature activates different neuromeres, and newly-increased neuromere.
Description
Technical field
The invention belongs to field of target recognition, particularly a kind of increment type target identification system based on neuromere differentiation.
Background technology
The deep neural network of early stage is that end to end, network structure keeps constant in training process, and training process is to learn
Practise the process of network parameter.Training set is complete and during enough network depths, and arbitrarily complicated original number can be expressed in theory
According to.And in many practical applications, can not disposably obtain complete data set, therefore learning algorithm is also required to be one to hold
Continuous process.For example to scheme to search figure, in the software products such as shopping of taking pictures, existing commodity image collection is obtained in advance, by depth god
Through network for input picture finds immediate preceding some items as recommendation results.But when input is new
During product image, although the training algorithm of legacy network, proposed algorithm are very perfect, can neither suitably be recommended, also cannot
Data set is improved as new samples.
Traditional neural network often receives a new samples and is required for updating global parameter:If new samples are noises,
There is network to be fitted noise without distinction, so as to destroy existing network;If new samples belong to new type, and (example is as in the previous
Sample is all animal, and new samples are vehicles), then need bottom-up change all-network parameter, the internal memory of this global adaptation
Consumption and time loss are typically huge.In order to avoid the repetitive learning under mass data, network in newly-increased sample simultaneously
All of knowledge need not be refreshed, but only update the local parameter as caused by increasing sample newly.
Incremental Learning Algorithm can make the ability that system possesses above-mentioned dynamic learning, and various forms of Incremental Learning Algorithms
Being studied under each field different scenes.Most classical clustering algorithm has following two inherent shortcomings:The network number of plies or
The parameters such as network node number need pre-defined;Network is to noise-sensitive.On delta algorithm in the non-supervisory task of image
Research work is relatively fewer, and major reason is that the information content high redundancy of initial data is full so as to be difficult to be reached in cluster
The accuracy rate of meaning.
The content of the invention
The present invention seeks to, solve following technical problem, existing recognition methods exist network structure solidification, be only applicable to it is quiet
Repetitive learning, the situation of anti-noise ability difference are carried out under state data set, mass data environment.Propose it is a kind of based on neuromere differentiation
Increment type target identification system.
Technical solution of the present invention, a kind of increment type target identification system based on neuromere differentiation, by low hidden layer, nerve
Ganglionic layer, hidden layer high and grader are constituted, and periopticon is located between low hidden layer and hidden layer high;The neuromere of nervous layer is carried
Sampling eigen rule, by ganglionic activation and differentiation, corresponding independent feature set is formed in hidden layer high, and it is right to be formed
The feature memory of sample;The variable amounts of the feature set of hidden layer high, the quantity of feature set and the neuromere of differentiation in nervous layer
Quantity is corresponding;Newly-increased neuromere can form new memory to new samples, adaptively update the feature set of hidden layer high, real
The target identification of existing increment type;
Described neuromere refers to for characterizing one group of neutral net node of the similar sample of the regularity of distribution, when input sample
When eigen parameter is more than activation threshold, neuromere is activated, and the neuromere forms independent feature set in hidden layer high;Swashed
When ganglionic activity living is less than threshold value, apoptosis is not formed feature set by the neuromere in hidden layer high;The different sample of feature
The different neuromere of this activation, and newly-increased neuromere, adaptively update the feature set of hidden layer high;
Ganglionic activation refers to that the distribution of sample meets ganglionic internal characteristicses, participates in the corresponding spy of neuromere
The training of collection;Described differentiation refers to that self adaptation increases ganglionic process newly, and some samples incorporate neuromere, increased the god
The activity of warp knuckle;
Ganglionic apoptosis refer in certain scope (certain hour, quantity or other factors) without new sample
The neuromere is activated, the relative activity of the sample declines, is removed in periopticon.Neuromere is activated and not necessarily break up,
May apoptosis;
The periopticon replaces initial data as input with the feature diagram data after compressed encoding:
If training the low hidden layer of dried layer on data set to obtain local filter, and general local feature is acted on into sample
This, coding obtains the characteristic pattern of each sample, is input into periopticon;Remember L layers altogether of low hidden layer, parameter is respectively W1、W2、…
WL, remembering that pretreated operable data is V, then V is via the first hidden layer W1H is obtained after coding1, H1By space pond
To pH1,pH1Via the second hidden layer W2H is obtained after coding2, until obtaining pHLAs the input data of periopticon.
Further, determine that the method for neuromere number comprises the following steps:
Note neuromere GeCharacterize one group of similar sample set of the regularity of distribution, wherein subscript e ∈ [1,2 ..., | G |], | G | tables
Show neuromere number;Calculate the significance γ of each sample pointi, wherein i ∈ [1,2 ..., | V |], | V | expression number of samples, and
It is sorted in ascending order, finds boundary value τ of the significance far above other sample points, sample point of the significance higher than boundary value τ is god
Through node, and remember that ganglionic number is N.
Further, the periopticon activity is specifically included:
The sample characteristics figure regularity of distribution is extracted, input sample is stimulated different neuromeres and is processed respectively;If new samples
Certain any known neuromere is activated, then the known neuromere parameter is only updated, without updating all-network parameter;If
Noise, then new samples periopticon be marked as noise without enter any neuromere, to protect existing network;If category
In unknown neuromere, then break up new neuromere and the sample is participated in the training of the corresponding feature set of the neuromere.
Further, whether judgement sample point activates known ganglionic method and comprises the following steps:
Note LgRepresent the label of neuromere g, g ∈ [G1,G2,…,GN], N is neuromere number;If there is at least one
Neuromere g so that current sample point v is less than to its distance and blocks apart from dc, then sample point v belongs to known neuromere, and god
Warp knuckle label LvFor:
Wherein operatorThe value of independent variable x, d when expression makes function f (x) obtain minimum valuevgRepresent
The distance between sample point v and neuromere g, dcTo block distance.
Further, whether judgement sample point is that the method for noise comprises the following steps:
If current sample point is all higher than blocking apart from d to the distance of any known ganglionic center pointc, then judge current
Sample point is noise.
Further, whether judgement sample point belongs to unknown ganglionic method and comprises the following steps:
If current sample point is noise, but significance γiIt is more than boundary value τ or notable higher than non-neurode
Degree, then current sample point belongs to unknown neuromere, and updates boundary value, differentiates new neuromere.
A kind of increment type target identification system based on neuromere differentiation according to claim 1, it is characterised in that
Each ganglionic feature set is integrated, all samples is encoded with the feature set after unified integration, the characteristic pattern and sample for obtaining
Label is input into as grader, completes the study of sorter model and sample to be identified input grader is completed into prediction.
Further, the integration method of wave filter is:
The low hiding layer data pH of sampleLNo longer need to experience periopticon, but obtained in the training of all neuromeres
Respective feature WL+1 cThe W of splicingL+1 new=[WL+1 1,WL+1 2,…,WL+1 N] on map out the unified height of dimension and hide layer data HL+1,
And subsequent grader is input into, wherein subscript L+1 represents a floor height hidden layer, and subscript represents ganglionic label.
Input data is activated different neuromeres and processed respectively by the present invention:If new samples are noises, new samples exist
Weak cluster layer is marked as noise without entering any neuromere, also will not continue to send into high-rise training, so as to protect
There is network;If new samples activate unknown neuromere (example sample as in the previous is all animal, and new samples are vehicles), then automatic point
Change new neuromere, and form independent characteristic collection;If the known neuromere of new samples activation, it is not required that refreshing is all of to be known
Know, but only update the ganglionic local parameter of correspondence.Methods described specifically includes following steps:
Step 1, sample preprocessing:All samples that data are concentrated are carried out with denoising, albefaction and normalization pretreatment (also may be used
Including some image enhancement processings), generate operable data;Normalized therein is essential operation, and other can suitably be selected
Select treatment.
The original sample for remembering input is X, XijRepresent the i-th row jth column data value of image.Normalization pretreatment is divided into two
Step, carries out subtraction normalization first, namely:Yij=Xij-∑pqwpq·Xi+p,j+q, wherein wpqIt is the pth row q of Gauss weighting windows
Row weights, and meet condition ∑pqwpq=1;Division normalization is carried out to Yij again, namely:Wherein
Step 2, general local feature training:If the low hidden layer of dried layer is trained on data set, to obtain local filter
(also can be using the good local feature of off-line training);
Step 3, feature graph code:General local feature is acted on into sample, coding obtains the characteristic pattern of each sample;
Step 4, neuromere differentiation specifically includes following steps:
The distance between each sample point d is calculated firstij.Sample involved in the present invention is arbitrary dimension, and is different from tradition
Fixed dimension in recognition methods, and general distance metric method requirement dimension is consistent, therefore devise a kind of space gold word
Tower basin algorithm has solved above-mentioned contradiction.
Remember that the data after L hidden layers coding areWhereinReal number space is represented, R, C, T are represented respectively
The length and width of the data, port number;And remember pyramid block size for n, then piecemeal window length and width are respectively
Piecemeal window transverse and longitudinal direction moving step length isWhereinRepresent respectively downward rounding operation and upwards
Rounding operation.By piecemeal window in pHLEach passage on mobile and computing pool value, then be fixed the output of dimensionNotice that the dimension of Λ is only determined by pyramid block size and port number, and with the length and width size of sample data without
Close, the sample data of such different dimensions has been converted into identical dimensional, then can easily calculate distance.
Secondly neuromere number is determined.The significance γ of all sample points is calculated respectivelyi, and be sorted in ascending order, find aobvious
Far above the boundary value τ of other sample points, sample point of the significance higher than boundary value τ is ganglionic center point to work degree, and note ought
Preceding branch's number is N number of.
Then the attribute of new sample point is judged.Note LgRepresent the label of neuromere g, g ∈ [G1,G2,…,GN], N is nerve
Section number;If there is at least one neuromere g so that current sample point v is less than to its distance and blocks apart from dc, then sample
Point v belongs to known neuromere, and neuromere label LvFor:Wherein operatorExpression makes function f
The value of independent variable x when () obtains minimum value x.
If current sample point is all higher than blocking apart from d to any ganglionic distancec, then current sample point is noise.
If current sample point is noise, but significance γiMore than boundary value τ, or higher than the significance of non-neurode, then current sample
This point belongs to unknown neuromere, updates boundary value, breaks up new neuromere and trains.
Step 5, each neuromere high level hidden layer training, determines the ownership of each sample point in step 4, and differentiation is obtained
Different neuromeres respectively in respective sample set train high-level characteristic collection.
Hidden layer in step 2 and step 5 can be trained using convolution strategy by deep learning, the chi of convolution kernel
Very little and number can decide as circumstances require, but convolution kernel in step 2 is unsuitable excessive, and the convolution kernel of step 5 Zhong Ge branches should not mistake
It is many.
Step 6, classifier training and prediction:Each ganglionic feature set that integration step 5 is obtained first, by all samples
Encoded with the feature set after unified integration, the characteristic pattern and its sample label for obtaining are input into as grader, complete grader
Sample to be identified input grader is simultaneously completed prediction by the study of model.
Feature graph code is specially in step 3 and step 6:Remember L layers altogether of low hidden layer, parameter is respectively W1、W2、…WL,
Pretreated operable data is remembered for V, then V is via the first hidden layer W1H is obtained after coding1, H1By nonlinear activation, sky
Between the conversion such as pond obtain pH1,pH1Via the second hidden layer W2H is obtained after coding2, until obtaining pHLAs the input of cluster layer
Data.
Hidden layer high amounts to one layer, namely low hiding layer data pHLNo longer need to experience clustering phase, but in all god
The respective feature W that warp knuckle training is obtainedL+1 c(subscript L+1 represents a floor height hidden layer, and subscript c represents c-th neuromere) splices
WL+1 new=[WL+1 1,WL+1 2,…,WL+1 N] on map out the unified height of dimension and hide layer data HL+1, and it is input into subsequent classification
Device.
The invention provides a kind of increment type target identification system, its on-line study in identification, and with unsupervised side
Formula realizes the growth of network structure.Carried out instant invention overcomes the solidification of existing recognition methods network structure, under mass data environment
The defect of repetitive learning, it is adaptable to recognize various targets in time, especially emerging unknown classification target.
Specifically, sample is pre-processed with enhancing technology and method for normalizing, to improve discrimination;Use compaction table
Hiding layer data after levying replaces high redundancy data input periopticon, to ensure the incremental nature of new network;Use band convolution
Each hidden layer of deep learning network training of characteristic, local wave filter is applied to obtain;Final each sample is by integrating
The unified wave filter of rear dimension is encoded and sends into grader, to provide the result of decision.
Beneficial effect, this structure can learn in identification, can incrementally determine the structure and local parameter of network.
In face of the database that dynamic updates, this structure automatic marking noise sample, and the knowledge for having learnt can be protected;This method is adaptive
The affiliated neuromere of sample is should determine that, and can only update local parameter, it is to avoid repetitive learning;This method can detect unknown object class
Type, and when training data is identical, Detection results are better than other related legacy networks.Further, the pyramid of this method
Space pondization designs the system of causing and is applied to the arbitrary target data set of size.
Brief description of the drawings
Fig. 1 is schematic structural view of the invention;
Fig. 2 is target pre-treatment step sub-process figure in the present invention;
Fig. 3 is local feature training step schematic diagram in the present invention.
Specific embodiment
To make the present invention apparent understandable, hereby with specific embodiment, and accompanying drawing is coordinated to be described below in detail.
As shown in figure 1, the invention discloses a kind of increment type target identification system based on neuromere differentiation, the identification of its side
Side on-line study, and the growth of network structure is realized in unsupervised mode.The method is comprised the following steps:
Step 1, image preprocessing:All images that data are concentrated are carried out with gray processing, normalization and whitening pretreatment, such as
Shown in Fig. 2.
First by coloured image I by original RGB triple channels according to [0.2989,0.5870,0.1140] proportion weighted ash
Degreeization obtains image X.
Followed by subtraction normalization, namely:Yij=Xij-∑pqwpq·Xi+p,j+q, wherein wpqIt is Gauss weighting windows
Pth row q row weights, and meet condition ∑pqwpq=1;Division normalization is carried out to Yij again, namely:Wherein
Finally done by transform to frequency domain and with notch filter H do after inner product Fourier inversion obtain but layer V, V most
The first hidden layer of input is subsequently trained eventually.
Hidden layer in step 2 and step 5 is trained using convolution strategy by deep learning, tight after every layer of convolutional layer
Adjacent pond layer, pond yardstick is 2 × 2.Low hidden layer amounts to two-layer:The convolution kernel size of the first hidden layer is 10 × 10,
Number is 48;The convolution kernel size of the second hidden layer is 8 × 8, and number is 96.Hidden layer high each neuromere only has one
Layer, convolution kernel size is 20 × 20, and number is 20.
Remember that l hidden layers are output as pHl, dimension is M × N × K, and base vector is g;L+1 hidden layers are Hl+1, dimension
It is P × Q × R, base vector is b;Wave filter between the two is Wl+1, dimension is Nw×Nw×K×R.By lower level to higher level
Coded system is:
Wherein σ represents S- type activation primitives, and * represents convolution operator, and the tilde above matrix is represented matrix elder generation water
Flip vertical and other dimensions keep constant again for flat upset, and Φ (x) represents that pondization is operated.
It is corresponding, be to lower level decoding process by higher level:
Wherein Ψ (x) represents and is sampled according to certain normal distribution that I represents the unit matrix of identical dimensional.
The feature graph code of image is specially in step 3 and step 6:Remember that low hiding layer parameter is respectively W1、W2, the pre- place of note
Visual layer data after reason is V, then V is via the first hidden layer W1H is obtained after coding1, H1PH is obtained by pond1,pH1Via
Two hidden layer W2H is obtained after coding2, the pH of Chi Huahou2As the input data of cluster layer.Hidden layer high amounts to one layer, namely low
Hide layer data pH2No longer need to experience clustering phase, but the respective feature W obtained in the training of all neuromeres3 c(subscript 3
Represent the 3rd hidden layer, subscript c represents c-th neuromere) splicing W3 new=[W3 1,W3 2,…,W3 N] on map out dimension system
One height hides layer data H3, and it is input into subsequent grader.
The pond of pyramid space described in step 4 method is specially:
Remember that the data after the second hidden layer coding areWhereinRepresent real number space, R, C, 96 difference tables
Show length and width, the port number of the data;And remember pyramid block size for n, then piecemeal window length and width are respectively
Piecemeal window transverse and longitudinal direction moving step length isWhereinRepresent respectively downward rounding operation and upwards
Rounding operation.By piecemeal window in pH2Each passage on mobile and computing pool value, then be fixed the output of dimensionNotice that the dimension of Λ is only determined by pyramid block size and port number, and with the length and width size of sample data without
Close, the sample data of such different dimensions has been converted into identical dimensional, then can easily calculate distance.Pyramid piecemeal
Size n × n is four class values 9 × 9,7 × 7,5 × 5,3 × 3, therefore the vector dimension for finally giving is 164 × 96, unified dimensional
Afterwards apart from computational algorithm use CW-SSIW indexes.
Determine that feature map space divides a counting method and is specially described in step 4:
The local density ρ of all sample points is calculated respectively:
Wherein dijRepresent the distance between i-th sample point and j-th sample point, dcTo block distance (value in experiment
For 0.007), i, j represent the sequence number of sample point, | V | represents sample points, and ξ represents switch function.
The local radius r of all sample points:
The significance γ of each sample pointi=ρiri, and be sorted in ascending order, significance is found far above other sample points
Boundary value τ, sample point of the significance higher than boundary value τ is ganglionic center point, and remembers that current branch number is N number of.
Then the attribute of new sample point is judged.Note LgRepresent the label of neuromere g, g ∈ [G1,G2,…,GN], N is nerve
Section number;If there is at least one neuromere g so that current sample point v is less than to its distance and blocks apart from dc, then sample
Point v belongs to known neuromere, and neuromere label LvFor:Wherein operatorExpression makes function f
The value of independent variable x when () obtains minimum value x.If current sample point is all higher than blocking distance to any ganglionic distance
dc, then current sample point is noise.If current sample point is noise, but significance γiMore than boundary value τ, or higher than non-nerve
The significance of node, then current sample point belong to unknown neuromere, update boundary value, break up new neuromere and simultaneously train.
Step 5, each branch's high level hidden layer training, determines the ownership of each sample point in step 4, and differentiation is obtained
Different neuromeres train high-rise wave filter in respective sample set respectively.
Step 6, classifier training and the prediction for having supervision:The wave filter of each branch that integration step 5 is obtained first, by institute
There is sample to be encoded with the wave filter after unified integration, the characteristics of image figure and its label for obtaining complete softmax as input
The parameter learning of grader, and images to be recognized input softmax graders are completed into prediction.
For convenience, the input data of note softmax graders is made up of m marked sample:{(x1,y1) ...,
(xm,ym), whereinRepresent that characteristic vector x is the real number of N+1 dimensions, x0=1 correspondence intercept, tag entry yi∈{1,
2 ..., K } represent actually total K classification.
For the corresponding input x of test sample for giving, it is assumed that function hθX () is:
WhereinIt is the parameter of sorter model.
Further, cost function is:
Value is 1 when wherein function g (x) and if only if x expression formulas are true, is 0 in the case of other.For the minimum of J (θ)
Change problem is there is presently no enclosed solution, therefore we using L-BFGS iteratives, wherein gradient formula is:
Wherein symbolItself it is a vector, its l-th elementIt is J (θ) to θjL-th component
Partial derivative.
The present invention is compared in Caltech-101, ORL Face common data sets with other method.Caltech-
The test process simulation of bulking power described in step 4 is as follows in 101 data sets:In 1-150 sequences, input sample is taken from random
Ice skate, candlestick, binoculars, Garfield and gerenuk;Since the 151st sequence, test environment changes, input
Sample takes from platypus, wildcat, octopus, scissors and Snoopy at random;Since the 301st sequence, test environment changes again, with
This analogizes.Wherein, 10% image in each environment takes from remaining data collection as ambient noise at random.
Performance of the structure of of the invention and correlation technique in the identification mission of Caltech-101 storehouses, table are compared in following table
Conv represents convolutional layer in lattice, and bracket inner digital represents the size of convolution kernel, number, wherein the volume of the 3rd convolutional layer of the invention
Product core number is adaptive to be should determine that.The convolution number of plies of IDNN systems amounts to 3 layers in experiment, therefore be compared for respectively in form last column
Individual network is in the Conv3 layers of recognition accuracy of (SP2 layers of correspondence in LCC networks).
The performance of the present invention and network of relation in ORL database Non-surveillance clustering tasks is compared in following table, with normalizing
It is evaluation criterion to change mutual information NMI.
A kind of increment type target identification system based on neuromere differentiation disclosed by the invention, the above is only the present invention
Preferred embodiment, the method and approach for implementing the technical scheme be a lot, under the premise without departing from the principles of the invention
Some improvements and modifications can also be made.Therefore, protection scope of the present invention is worked as and is defined depending on those as defined in claim.
Claims (10)
1. it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that by low hidden layer, periopticon, height
Hidden layer and grader are constituted, and periopticon is located between low hidden layer and hidden layer high;The neuromere of periopticon extracts sample
Eigen rule, by ganglionic activation and differentiation, corresponding independent feature set, shape paired samples is formed in hidden layer high
Feature memory;The variable amounts of the feature set of hidden layer high, the quantity of feature set and the neuromere quantity of differentiation in nervous layer
It is corresponding;Newly-increased neuromere can form new memory to new samples, adaptively update the feature set of hidden layer high, realize increasing
The target identification of amount formula;
Described neuromere refers to for characterizing one group of neutral net node of the similar sample of the regularity of distribution, when input sample is special
When levying parameter more than activation threshold, neuromere is activated, and the neuromere forms independent feature set in hidden layer high;It is activated
When ganglionic activity is less than threshold value, apoptosis is not formed feature set by the neuromere in hidden layer high;The different sample of feature swashs
The different neuromere of work, and newly-increased neuromere, adaptively update the feature set of hidden layer high;
Ganglionic activation refers to that the distribution of sample meets ganglionic internal characteristicses, participates in the corresponding feature set of neuromere
Training;Described differentiation refers to that self adaptation increases ganglionic process newly, and some samples incorporate neuromere, increased the neuromere
Activity;
Ganglionic apoptosis refers to that (certain hour, quantity or other factors) are activated without new sample in certain scope
The neuromere, the relative activity of the sample declines, and is removed in periopticon;Neuromere is activated and not necessarily break up, it is also possible to
Can apoptosis;
The periopticon replaces initial data as input with the feature diagram data after compressed encoding:
If training the low hidden layer of dried layer on data set to obtain local filter, and general local feature is acted on into sample,
Coding obtains the characteristic pattern of each sample, is input into periopticon;Remember L layers altogether of low hidden layer, parameter is respectively W1、W2、…WL,
Pretreated operable data is remembered for V, then V is via the first hidden layer W1H is obtained after coding1, H1Obtained by space pond
pH1,pH1Via the second hidden layer W2H is obtained after coding2, until obtaining pHLAs the input data of periopticon.
2. it is according to claim 1 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that really
The method following steps of warp knuckle of composing oneself number:
Note neuromere GeOne group of similar sample set of the regularity of distribution is characterized, wherein subscript e ∈ [1,2 ..., | G |], | G | represents nerve
Section number;Calculate the significance γ of each sample pointi, wherein i ∈ [1,2 ..., | V |], | V | expression number of samples, and by ascending order
Sequence, finds boundary value τ of the significance far above other sample points, and sample point of the significance higher than boundary value τ is neurode,
And remember that ganglionic number is N.
3. it is according to claim 1 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that institute
Periopticon activity is stated to specifically include:
The sample characteristics figure regularity of distribution is extracted, input sample is stimulated different neuromeres and is processed respectively;If new samples are activated
Certain any known neuromere, then only update the known neuromere parameter, without updating all-network parameter;If making an uproar
Sound, then new samples periopticon be marked as noise without enter any neuromere, to protect existing network;If belonged to
Unknown neuromere, then break up new neuromere and the sample is participated in the training of the corresponding feature set of the neuromere.
4. it is according to claim 3 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that sentence
Whether disconnected sample point activates known ganglionic method comprises the following steps:
Note LgRepresent the label of neuromere g, g ∈ [G1,G2,…,GN], N is neuromere number;If there is at least one neuromere
G so that current sample point v is less than to its distance and blocks apart from dc, then sample point v belongs to known neuromere, and neural feast-brand mark
Sign LvFor:
Wherein operatorThe value of independent variable x, d when expression makes function f (x) obtain minimum valuevgRepresent sample point
The distance between v and neuromere g, dcTo block distance.
5. it is according to claim 3 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that sentence
Whether disconnected sample point is that the method for noise comprises the following steps:
If current sample point is all higher than blocking apart from d to the distance of any known ganglionic center pointc, then current sample is judged
Point is noise.
6. it is according to claim 3 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that sentence
Whether disconnected sample point belongs to unknown ganglionic method comprises the following steps:
If current sample point is noise, but significance γiMore than boundary value τ, or higher than the significance of non-neurode, then
Current sample point belongs to unknown neuromere, and updates boundary value, differentiates new neuromere.
7. it is according to claim 1 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that it is whole
Each ganglionic feature set is closed, all samples is encoded with the feature set after unified integration, the characteristic pattern and sample mark for obtaining
Sign and be input into as grader, complete the study of sorter model and sample to be identified input grader is completed into prediction.
8. it is according to claim 1 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that filter
The integration method of ripple device is:
The low hiding layer data pH of sampleLNo longer need to experience periopticon, but the respective spy obtained in the training of all neuromeres
Levy WL+1 cThe W of splicingL+1 new=[WL+1 1,WL+1 2,…,WL+1 N] on map out the unified height of dimension and hide layer data HL+1, and be input into
Subsequent grader, wherein subscript L+1 represent a floor height hidden layer, and subscript represents ganglionic label.
9. it is according to claim 1 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that it is low
Hide layer data pHLNo longer need to experience clustering phase, but the respective feature W obtained in the training of all neuromeresL+1 c(subscript L
+ 1 represents a floor height hidden layer, and subscript c represents c-th neuromere) W of splicingL+1 new=[WL+1 1,WL+1 2,…,WL+1 N] on map
Go out the unified height of dimension and hide layer data HL+1, and it is input into subsequent grader.
10. it is according to claim 1 it is a kind of based on neuromere differentiation increment type target identification system, it is characterised in that
Input data is activated into different neuromeres and is processed respectively:If new samples are noises, new samples are labeled in weak cluster layer
Be noise without enter any neuromere, also will not continue to send into high-rise training, so as to protect existing network;If new sample
The unknown neuromere of this activation (example sample as in the previous is all animal, and new samples are vehicles), then break up new neuromere automatically, and
Form independent characteristic collection;If the known neuromere of new samples activation, it is not required that refresh all of knowledge, but only update correspondence
Ganglionic local parameter;Methods described specifically includes following steps:
Step 1, sample preprocessing:All samples that data are concentrated are carried out with denoising, albefaction and normalization pretreatment (to may also comprise
Some image enhancement processings), generate operable data;Normalized therein is essential operation, and other can suitably at selection
Reason;
The original sample for remembering input is X, XijRepresent the i-th row jth column data value of image;Normalization pretreatment is divided into two steps, first
Subtraction normalization is first carried out, namely:Yij=Xij-∑pqwpq·Xi+p,j+q, wherein wpqFor the pth row q of Gauss weighting windows arranges power
Value, and meet condition ∑pqwpq=1;Again to YijDivision normalization is carried out, namely:Zij=Yij/max(c,θij), whereinC=mean (θij);
Step 2, general local feature training:If training the low hidden layer of dried layer on data set, (also may be used with obtaining local filter
Using the good local feature of off-line training);
Step 3, feature graph code:General local feature is acted on into sample, coding obtains the characteristic pattern of each sample;
Step 4, neuromere differentiation specifically includes following steps:
The distance between each sample point d is calculated firstij;Sample involved in the present invention is arbitrary dimension, and is different from tional identification
Fixed dimension in method, and general distance metric method requirement dimension is consistent, therefore devise a kind of spatial pyramid pond
Change algorithm and solve above-mentioned contradiction;
Remember that the data after L hidden layers coding areWhereinReal number space is represented, R, C, T represent the number respectively
According to length and width, port number;And remember pyramid block size for n, then piecemeal window length and width are respectivelyPiecemeal
Window transverse and longitudinal direction moving step length isWhereinDownward rounding operation is represented respectively and is rounded up
Computing;By piecemeal window in pHLEach passage on mobile and computing pool value, then be fixed the output of dimension
Notice that the dimension of Λ is only determined by pyramid block size and port number, and it is unrelated with the length and width size of sample data, so not
Identical dimensional has been converted into the sample data of dimension, distance then can have easily been calculated;
Secondly neuromere number is determined;The significance γ of all sample points is calculated respectivelyi, and be sorted in ascending order, find significance remote
Higher than the boundary value τ of other sample points, sample point of the significance higher than boundary value τ is ganglionic center point, and remembers current branch
Number is N number of;
Then the attribute of new sample point is judged;Note LgRepresent the label of neuromere g, g ∈ [G1,G2,…,GN], N is neuromere
Number;If there is at least one neuromere g so that current sample point v is less than to its distance and blocks apart from dc, then sample point v
Belong to known neuromere, and neuromere label LvFor:Wherein operatorExpression makes function f (x)
The value of independent variable x when obtaining minimum value;
If current sample point is all higher than blocking apart from d to any ganglionic distancec, then current sample point is noise;If worked as
Preceding sample point is noise, but significance γiMore than boundary value τ, or higher than the significance of non-neurode, then current sample point category
In unknown neuromere, boundary value is updated, break up new neuromere and train;
Step 5, each neuromere high level hidden layer training, determines the ownership of each sample point in step 4, and differentiation is obtained not
High-level characteristic collection is trained in respective sample set respectively with neuromere;
Hidden layer in step 2 and step 5 can be trained using convolution strategy by deep learning, the size of convolution kernel and
Number can decide as circumstances require, but convolution kernel in step 2 is unsuitable excessive, and the convolution kernel of step 5 Zhong Ge branches should not be excessive;
Step 6, classifier training and prediction:Each ganglionic feature set that integration step 5 is obtained first, by all samples system
Feature set coding after one integration, the characteristic pattern for obtaining and its sample label are input into as grader, complete sorter model
Study and by sample to be identified input grader complete prediction;
Feature graph code is specially in step 3 and step 6:Remember L layers altogether of low hidden layer, parameter is respectively W1、W2、…WL, note is in advance
Operable data after treatment is V, then V is via the first hidden layer W1H is obtained after coding1, H1By nonlinear activation, space pond
The conversion such as change obtains pH1,pH1Via the second hidden layer W2H is obtained after coding2, until obtaining pHLAs the input number of cluster layer
According to.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710055128.1A CN106919980B (en) | 2017-01-24 | 2017-01-24 | Incremental target identification system based on ganglion differentiation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710055128.1A CN106919980B (en) | 2017-01-24 | 2017-01-24 | Incremental target identification system based on ganglion differentiation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106919980A true CN106919980A (en) | 2017-07-04 |
CN106919980B CN106919980B (en) | 2020-02-07 |
Family
ID=59453429
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710055128.1A Active CN106919980B (en) | 2017-01-24 | 2017-01-24 | Incremental target identification system based on ganglion differentiation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106919980B (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107622283A (en) * | 2017-09-28 | 2018-01-23 | 上海理工大学 | A kind of increment type object identification method based on deep learning |
CN109064454A (en) * | 2018-07-12 | 2018-12-21 | 上海蝶鱼智能科技有限公司 | Product defects detection method and system |
WO2019066718A1 (en) * | 2017-09-28 | 2019-04-04 | Agency For Science, Technology And Research | Self-assessing deep representational units |
WO2019085379A1 (en) * | 2017-10-30 | 2019-05-09 | 北京深鉴智能科技有限公司 | Hardware realization circuit of deep learning softmax classifier and method for controlling same |
CN109766954A (en) * | 2019-01-31 | 2019-05-17 | 北京市商汤科技开发有限公司 | A kind of target object processing method, device, electronic equipment and storage medium |
CN110390355A (en) * | 2019-07-01 | 2019-10-29 | 东北大学 | The new defect identification method of pipeline based on evolution maximum fuzzy minimum neural network |
CN110443363A (en) * | 2018-05-04 | 2019-11-12 | 北京市商汤科技开发有限公司 | Characteristics of image learning method and device |
CN111667069A (en) * | 2020-06-10 | 2020-09-15 | 中国工商银行股份有限公司 | Pre-training model compression method and device and electronic equipment |
CN112396084A (en) * | 2019-08-19 | 2021-02-23 | ***通信有限公司研究院 | Data processing method, device, equipment and storage medium |
CN112988082A (en) * | 2021-05-18 | 2021-06-18 | 南京优存科技有限公司 | Chip system for AI calculation based on NVM and operation method thereof |
CN113960553A (en) * | 2021-10-14 | 2022-01-21 | 电子科技大学 | Gaussian weight distribution feature extraction method in one-dimensional image recognition |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7313550B2 (en) * | 2002-03-27 | 2007-12-25 | Council Of Scientific & Industrial Research | Performance of artificial neural network models in the presence of instrumental noise and measurement errors |
CN101826166A (en) * | 2010-04-27 | 2010-09-08 | 青岛大学 | Novel recognition method of neural network patterns |
CN104778448A (en) * | 2015-03-24 | 2015-07-15 | 孙建德 | Structure adaptive CNN (Convolutional Neural Network)-based face recognition method |
CN105046272A (en) * | 2015-06-29 | 2015-11-11 | 电子科技大学 | Image classification method based on concise unsupervised convolutional network |
CN105426919A (en) * | 2015-11-23 | 2016-03-23 | 河海大学 | Significant guidance and unsupervised feature learning based image classification method |
CN105469142A (en) * | 2015-11-13 | 2016-04-06 | 燕山大学 | Neural network increment-type feedforward algorithm based on sample increment driving |
CN106326899A (en) * | 2016-08-18 | 2017-01-11 | 郑州大学 | Tobacco leaf grading method based on hyperspectral image and deep learning algorithm |
-
2017
- 2017-01-24 CN CN201710055128.1A patent/CN106919980B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7313550B2 (en) * | 2002-03-27 | 2007-12-25 | Council Of Scientific & Industrial Research | Performance of artificial neural network models in the presence of instrumental noise and measurement errors |
CN101826166A (en) * | 2010-04-27 | 2010-09-08 | 青岛大学 | Novel recognition method of neural network patterns |
CN104778448A (en) * | 2015-03-24 | 2015-07-15 | 孙建德 | Structure adaptive CNN (Convolutional Neural Network)-based face recognition method |
CN105046272A (en) * | 2015-06-29 | 2015-11-11 | 电子科技大学 | Image classification method based on concise unsupervised convolutional network |
CN105469142A (en) * | 2015-11-13 | 2016-04-06 | 燕山大学 | Neural network increment-type feedforward algorithm based on sample increment driving |
CN105426919A (en) * | 2015-11-23 | 2016-03-23 | 河海大学 | Significant guidance and unsupervised feature learning based image classification method |
CN106326899A (en) * | 2016-08-18 | 2017-01-11 | 郑州大学 | Tobacco leaf grading method based on hyperspectral image and deep learning algorithm |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107622283A (en) * | 2017-09-28 | 2018-01-23 | 上海理工大学 | A kind of increment type object identification method based on deep learning |
US11657270B2 (en) | 2017-09-28 | 2023-05-23 | Agency For Science, Technology And Research | Self-assessing deep representational units |
WO2019066718A1 (en) * | 2017-09-28 | 2019-04-04 | Agency For Science, Technology And Research | Self-assessing deep representational units |
WO2019085379A1 (en) * | 2017-10-30 | 2019-05-09 | 北京深鉴智能科技有限公司 | Hardware realization circuit of deep learning softmax classifier and method for controlling same |
CN110443363A (en) * | 2018-05-04 | 2019-11-12 | 北京市商汤科技开发有限公司 | Characteristics of image learning method and device |
CN109064454A (en) * | 2018-07-12 | 2018-12-21 | 上海蝶鱼智能科技有限公司 | Product defects detection method and system |
CN109766954A (en) * | 2019-01-31 | 2019-05-17 | 北京市商汤科技开发有限公司 | A kind of target object processing method, device, electronic equipment and storage medium |
US11403489B2 (en) | 2019-01-31 | 2022-08-02 | Beijing Sensetime Technology Development Co., Ltd. | Target object processing method and apparatus, electronic device, and storage medium |
CN110390355A (en) * | 2019-07-01 | 2019-10-29 | 东北大学 | The new defect identification method of pipeline based on evolution maximum fuzzy minimum neural network |
CN112396084A (en) * | 2019-08-19 | 2021-02-23 | ***通信有限公司研究院 | Data processing method, device, equipment and storage medium |
CN111667069A (en) * | 2020-06-10 | 2020-09-15 | 中国工商银行股份有限公司 | Pre-training model compression method and device and electronic equipment |
CN111667069B (en) * | 2020-06-10 | 2023-08-04 | 中国工商银行股份有限公司 | Pre-training model compression method and device and electronic equipment |
CN112988082A (en) * | 2021-05-18 | 2021-06-18 | 南京优存科技有限公司 | Chip system for AI calculation based on NVM and operation method thereof |
CN112988082B (en) * | 2021-05-18 | 2021-08-03 | 南京优存科技有限公司 | Chip system for AI calculation based on NVM and operation method thereof |
CN113960553A (en) * | 2021-10-14 | 2022-01-21 | 电子科技大学 | Gaussian weight distribution feature extraction method in one-dimensional image recognition |
Also Published As
Publication number | Publication date |
---|---|
CN106919980B (en) | 2020-02-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106919980A (en) | A kind of increment type target identification system based on neuromere differentiation | |
CN110458844B (en) | Semantic segmentation method for low-illumination scene | |
CN107798349B (en) | Transfer learning method based on depth sparse self-coding machine | |
CN109271522A (en) | Comment sensibility classification method and system based on depth mixed model transfer learning | |
CN112364779A (en) | Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion | |
CN107885853A (en) | A kind of combined type file classification method based on deep learning | |
CN109934158B (en) | Video emotion recognition method based on local enhanced motion history map and recursive convolutional neural network | |
CN107871136A (en) | The image-recognizing method of convolutional neural networks based on openness random pool | |
CN102622515B (en) | A kind of weather prediction method | |
Zhang et al. | Bilateral sensitivity analysis: a better understanding of a neural network | |
CN108171318A (en) | One kind is based on the convolutional neural networks integrated approach of simulated annealing-Gaussian function | |
CN110570035B (en) | People flow prediction system for simultaneously modeling space-time dependency and daily flow dependency | |
CN109871892A (en) | A kind of robot vision cognitive system based on small sample metric learning | |
CN112560948B (en) | Fundus image classification method and imaging method under data deviation | |
CN114548591B (en) | Sequential data prediction method and system based on mixed deep learning model and Stacking | |
CN112733602B (en) | Relation-guided pedestrian attribute identification method | |
CN108073978A (en) | A kind of constructive method of the ultra-deep learning model of artificial intelligence | |
CN114330541A (en) | Road traffic accident risk prediction deep learning algorithm | |
CN108073985A (en) | A kind of importing ultra-deep study method for voice recognition of artificial intelligence | |
CN108073979A (en) | A kind of ultra-deep study of importing artificial intelligence knows method for distinguishing for image | |
CN105718858A (en) | Pedestrian recognition method based on positive-negative generalized max-pooling | |
Wang et al. | Kernel-based deep learning for intelligent data analysis | |
Li et al. | Diffusion Probabilistic Model Based End-to-End Latent Fingerprint Synthesis | |
Tomka et al. | Practical aspects of forming training/test samples for convolutional neural networks | |
CN117648646B (en) | Drilling and production cost prediction method based on feature selection and stacked heterogeneous integrated learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |