CN100595780C - Handwriting digital automatic identification method based on module neural network SN9701 rectangular array - Google Patents

Abstract

The invention discloses a handwritten digital automatic identifying method based on module neural network SN9701 matrix, which comprises the following four steps: (1) pre-treating handwritten digitalimage; (2) dividing class space of training set using K means clustering algorithm or affinity propagation algorithm; (3) designing classifier using module neural network matrix; (4) ensembling classifier using improved means integrative method. The classifier of the invention is used for dividing the class space and matrix module neural network learning for task decomposition, thereby greatly improving learning speed for the classifier and the precision of handwritten digital classifying.

Description

A kind of handwriting digital automatic identification method based on modular neural network

Technical field

The present invention relates to the handwriting digital automatic recognition system, particularly a kind of handwriting digital automatic identification method based on modular neural network.

Background technology

Handwritten Digital Recognition research to as if: how to utilize robot calculator to recognize that automatically the people is handwritten in the arabic numeral on the paper.The application of Handwritten Digital Recognition system mainly comprises: postcode, statistical report form, financial statement, bank money or the like.China's " three gold medals " engineering of beginning to widely popularize will rely on the input of data message to a great extent over the years, if can realize the automatic typing of information by the Handwritten Digit Recognition technology, can promote the progress of this cause undoubtedly.Therefore, the Study of recognition of handwritten numeral has major and immediate significance, uses in case study successfully also to drop into, and will produce huge social and economic benefit.

In actual applications, to the requirement of digital identification form word recognition correct rate than the literal harshness many.This be because, numeral does not have context relation, the identification of each individual character is all concerning important, and financial accounting, its severity of financial field that digit recognition often relates to are self-evident especially.In addition, the standard exercise storehouse of some common handwriting digitals such as American National Standard and Technical Board NIST database, United States postal service database (USPS) etc. all include a large amount of training samples at present, data processing in enormous quantities has suitable requirement again to systematic learning speed, and is many very perfect in theory but the low excessively method of speed is impracticable.Therefore, studying high performance Handwritten Digit Recognition algorithm is a challenging task.

In in the past 40 years, people have found out the key feature that a lot of ways are obtained hand-written character.These means are divided two big classes: global analysis and structure analysis.To the former, can use technology such as template matches, PEL (picture element) density, square, unique point, mathematic(al) manipulation.The feature of this class is usually used together with the statistical classification method.To the latter, need extract the essential characteristic of character shape mostly from the profile of character or skeleton, comprising: circle, end points, node, arc, projection, depression, stroke or the like, the sorting technique of the sentence structure often that is used with these architectural features.The years of researches practice shows, for complete hard-core handwritten numeral, almost can affirm: do not have a kind of simple proposal can reach very high discrimination and accuracy of identification.Therefore, the effort of this respect develops towards more ripe, complicated, comprehensive direction recently.On the one hand, the research worker makes great efforts new method is applied to pre-service, and in the middle of the feature extraction, and on the other hand, the research worker makes great efforts to design new and effective sorter, comes ten classifications of numerical character are classified.The present invention designs at the latter just.

As a kind of very important sorter in the pattern-recognition, nerual network technique is because characteristics such as its high precision, concurrency, self study, self-adaptation have been widely used in the Handwritten Digital Recognition system.When neural network face be extensive training sample set the time, be subjected to the restriction of hardware advances, its pace of learning and popularization ability often can not be satisfactory." divide and rule problem " exactly in an important channel of head it off, many scholars have proposed to solve complicated classification problem based on the modular neural network model of various task decomposition methods.Modular neural network resolves into better simply a series of subtask with a complex task, processing is finished with a neural network (submodule) in each subtask, finish the not necessarily serial of subtask cooperation along order, have it concurrently and may be that serial is parallel.Therefore, modular neural network has pace of learning faster than single neural network, and the scale of single network can be very not big yet in the system.One-against-others[N.J.Nilsson, Learning Machines:Foundations of Trainable Pattern-Classifying Systems.New York:McGraw-Hill, 1965] be the task decomposition method of the K class problem of standard, it is that K is individual than the simple two class subtasks of former classification task with a K class PROBLEM DECOMPOSITION, and each subtask complexity is opened certain classification with other all K-1 class discrimination.Therefore, need K two class sorters of design, be respectively f ₁..., f _K, the output of each sorter all is the one dimension value in the scope [0,1].For a new samples, calculate all sorter f ₁..., f _KTo the output valve of this sample, the sorter sequence number that output valve is the highest is exactly whole modular neural network to the class categories of this sample number.This modular neural network can be obtained classification and the generalization ability better than single neural network.

Yet, this simple task decomposition method fundamentally can not reduce the training sample number of each module classification device, because though the classification number that each sorter need solve is few, but the sample number that they need be learnt still all is all original sample numbers, and, if the training set classification of K class problem originally is symmetry (sample number that comprises of all categories about equally) relatively, after then using this task decomposition method, all two class subtasks all become asymmetric problem, the K value is big more, and asymmetric degree is big more.Classification is asymmetric then to be the thorny problem that another one causes the network convergence difficulty easily.Though Anand etc. are at document [R.Anand, K.Mehrotra﹠amp; C.K.Mohan etc, " Efficient Classification forMulticlass Problems Using Modular Neural Networks; " IEEE Trans.Neural Networks, vol.6 (1), pp.117-124,1995] a kind of the improving one's methods of BP algorithm proposition at general perceptron in solved the classification asymmetric problem that occurs in this modular neural network, and experimental results show that it can accelerate pace of learning, but this solution has specific specific aim, only at the BP algorithm of perceptron, and inapplicable to other learning algorithm of the neural network of other type or perceptron, so it has very strong limitation.Thereby, thoroughly solve this asymmetric problem, must start with from the task decomposition method, avoid the appearance of asymmetric classification problem in the task decomposable process.The present invention will start with from the task decomposition method and fundamentally solve the problem of using neural network.

Summary of the invention: the objective of the invention is: propose a kind of handwriting digital automatic identification method based on modular neural network, this method can provide a kind of efficient sorter at big training sample set in the Handwritten Digital Recognition, realizes study fast and high precision identification.

Technical scheme of the present invention is: a kind of handwriting digital automatic identification method based on modular neural network, particularly: the pre-service of handwriting digital image, obtain the handwriting digital sample from ccd image sensor, obtain original data with CPLD as the control center of image capturing system, with DSP as graphics processing unit, realize the automatic acquisition process of image, the quick collection and the storage of completion rate image are carried out series of preprocessing to digital picture then;

Described CCD is linear array type imageing sensor Charge Coupled Device, and described CPLD is CPLD Complex Programmable Logic Device, and described DSP is digital signal processor Digital Signal Processor;

To digital picture carry out series of preprocessing be to image carry out binaryzation, level and smooth, cut apart, standardize, obtain input signal, after obtaining the gray level image of character 16*16 pixel, all that obtain are trained with stretching 256 dimensional vectors that become of gray level image matrix, form training and gather χ with input vector _k, k=1 wherein, 2 ..., ten category labels of 10 expressions;

Do you judge that the input sample is training sample or test sample book? training sample then carries out the space-like division and the classifier design of training set successively in this way;

If test sample book, then carry out categorised decision, sorter comprises two ingredients: a neural network SN 9701 rectangular array and an integrated computer, wherein neural network SN 9701 rectangular array is to be made of the SN9701 chip, the manufacturer of described SN9701 chip is a Siemens, model is the SN9701 with self-learning function, it is the neural network integration modules of the single output of many inputs, and its inside mainly becomes circuit, feature weights to adjust circuit, performance index decision circuitry and function by the Chebyshev polynomials graupel to form circuit and form;

Integrated computer is made of a totalizer and a divider, and input vector produces a network output matrix through neural network SN 9701 rectangular array, produces final categorised decision according to this network output matrix by integrated computer again.

As a further improvement of existing technologies, the space-like of training set is divided, and is that input vector is gathered χ _kBe divided into D _kIndividual submanifold χ _d ^(k), d=1,2 ..., D _k, all submanifolds are formed

Individual submanifold is right

Each submanifold is to importing as the training of one two class sorter.

The mean value of formula (1) estimates to import the posterior probability that sample x belongs to di submanifold in the i class below totalizer in the integrated computer and divider calculating

Wherein M is a neural network SN 9701 rectangular array,

It is the neural network element

The output valve of certain SN9701 chip is by the 7th port OUT output of SN9701 chip, and then final classification judgement is according to being: ign x → c _{K '},

if

i＝1，2，...，K，di＝1，2，...，Di(2)。

In the design of sorter, neural network SN 9701 rectangular array is expressed as formula (3) form:

$M=\left(\begin{array}{ccccc}\mathrm{\Φ}& M_{12}& M_{13}& ......& M_{1K}\\ M_{21}& \mathrm{\Φ}& M_{23}& ......& M_{2K}\\ & ......& & ......& \\ M_{K1}& M_{K2}& ......& M_{K(K-1)}& \mathrm{\Φ}\end{array}\right),---\left(3\right)$

Wherein $M_{\mathrm{ij}}=\left(\begin{array}{cccc}{P}_{1^{\left(i\right)}{1}^{\left(j\right)}}& P_{1^{\left(i\right)}{2}^{\left(j\right)}}& ......& P_{1^{\left(i\right)}{D_j}^{\left(j\right)}}\\ P_{2^{\left(i\right)}{1}^{\left(j\right)}}& P_{2^{\left(i\right)}{2}^{\left(j\right)}}& ......& P_{2^{\left(i\right)}{D_j}^{\left(j\right)}}\\ & ......& ......& \\ P_{{D_i}^{\left(i\right)}{1}^{\left(j\right)}}& P_{{D_i}^{\left(i\right)}{2}^{\left(j\right)}}& ......& P_{{D_i}^{\left(i\right)}{D_j}^{\left(j\right)}}\end{array}\right)=\left(P_{{\mathrm{di}}^{\left(i\right)}{\mathrm{dj}}^{\left(j\right)}}\right)---\left(4\right)$

Wherein K is the classification number, i, and j=1,2 ..., K, i ≠ j, di=1,2 ..., D _i, dj=1,2 ..., D _j, neural network Be single output, its training sample set is expressed as formula (5)

$S_{{\mathrm{di}}^{\left(i\right)}{\mathrm{dj}}^{\left(j\right)}}=\{{(X_l^{{\mathrm{di}}^{\left(i\right)}},1)}_{l=1}^{N_{\mathrm{di}}^{\left(i\right)}}\∪{(X_l^{{\mathrm{dj}}^{\left(j\right)}},0)}_{l=1}^{N_{\mathrm{dj}}^{\left(j\right)}}\},---\left(5\right)$

In the design of sorter, each neural network SN 9701 rectangular array is the single output of a multilayer perceptron, each perceptron is a SN9701, dynamically construct its hidden unit number, and adopt back propagation learning algorithm study, the dynamic structure and the learning process of each perceptron is as follows in the neural network SN 9701 rectangular array: a, initialization S1=0;

B, perceptron that comprises S1 hidden unit of structure, wherein S1=0 represents that this perceptron is a double-layer structure, otherwise is three-decker.

The invention has the beneficial effects as follows:

Single neural network solves the handwriting digital automatic recognition problem, because training sample is larger, cause e-learning to be difficult to convergence, and pace of learning is slow.And traditional one-against-others task decomposition method still can not reduce the training sample scale, and causes the classification of each network training collection asymmetric easily, has increased the difficulty of study more.

The present invention proposes a kind of new neural network SN 9701 rectangular array, with K mean cluster method and affinity propagation clustering method, being used for space-like division and task decomposes, avoided the appearance of asymmetric classification problem in the task decomposable process, the training sample scale of each neural network after the decomposition reduces greatly, has simplified study greatly.The last integrated method of a kind of improved average that proposes again comes the result of integrated each module to obtain final categorised decision.

Test result on large scale database has proved the validity of this disaggregated model, the learning time of this model is 24% (as Fig. 5) of single neural network, is 20% (as Fig. 5) with modular neural network learning time of one-against-others task decomposition method; The recognition accuracy of this model can improve 1.6% than single perceptron, improves 0.7% (as Fig. 6) than the modular neural network with traditional one-against-others task decomposition method.

In addition, K mean cluster method or affinity propagation clustering method are used for the matrix module neural network that the space-like is divided and task is decomposed, be used for than general random method that the space-like is divided and the learning time of the matrix module neural network of task decomposition shortens (as Fig. 5) greatly, nicety of grading also increase (as Fig. 6).

Description of drawings:

Fig. 1 is a method flow diagram of the present invention.

Fig. 2 is a matrix module neural network classifier structural representation.

Fig. 3 is the neural network module that the SN9701 chip is imported single output more.

Fig. 4 is a pretreated example of handwriting digital image.

Fig. 5 is the learning time comparative result figure of sorter of the present invention and other sorter.

Fig. 6 is the nicety of grading comparative result figure of sorter of the present invention and other sorter.

Embodiment:

For a more detailed description to the present invention for example:

Fig. 1 is a method flow diagram of the present invention.In Fig. 1, at first obtain the handwritten numeral sample from CCD, CCD is as imageing sensor, with CPLD as image acquisition control center, with DSP as the primary image processing unit, realize the automatic acquisition process of image, finished the quick collection and the storage of image, then digital picture has been carried out series of preprocessing;

After the pre-service, enter the training set space-like respectively and divide and square modular neural network sorter.Wherein, the training set space-like is divided, input vector set χ _kBe divided into D _kIndividual submanifold χ _d ^(k), d=1,2 ..., D _kAll submanifolds are formed

Individual submanifold is right Each submanifold is to can be used as the training input of one two class sorter.Related work adopts K mean cluster method to divide and the propagation clustering method of making a concerted effort is divided.

Fig. 2 is a matrix module neural network structure synoptic diagram.This figure has illustrated the element of a matrix module nerve network system, and has described the flow direction of data.It mainly comprises two ingredients: a neural network SN 9701 rectangular array and an integrated computer.Wherein neural network SN 9701 rectangular array is by the SN9701 chip structure, and integrated computer is made of a totalizer and a divider.Input vector produces a neural network output matrix through neural network SN 9701 rectangular array, produces final categorised decision according to this neural network output matrix by integrated computer again.

Fig. 3 is the intensive through mixed-media network modules mixed-media of the single output of input more than, adopts the SN9701 chip to realize, its pin is arranged as shown in the figure.Its inside mainly becomes circuit, feature weights to adjust circuit, performance index decision circuitry and function by the Chebyshev polynomials graupel to form circuit etc. and form.

In Fig. 3, the SN9701 chip pin is arranged as: the first port SS is the first port sample learning input end, for sample set xi, di}, analog quantity di hold input thus;

The second port IN: the sample learning input end, to sample set xi, di}, analog quantity xi hold input thus; For the neural network that has trained, input variable is end input thus also;

The 3rd port DIS: sample training end mark end, low level is effective, the light emitting diode indication;

The 4th port GND: power supply ground;

Five-port ε: the performance index input end, ε is a voltage for little just simulation between appointing, and can be obtained by two resistance in series dividing potential drops;

The 6th port ST: start neural network learning input merchant end, pulse is effective;

The 7th port OUT: neural network output terminal;

The 8th port Vcc: power positive end, Vcc are the 10-30V power supply.

Fig. 4 is a pretreated example of handwriting digital image.The digital image (a) that obtains by imageing sensor through binaryzation, level and smooth, cut apart and standardization processing obtains image (b).

Binary conversion treatment is that the gray level image that will obtain is converted into the bi-level digital image, and this method adopts the global threshold method to carry out binary conversion treatment.Smoothly be to use a kind of simple and effective smoothing method---median filter method.Cut apart is to be cutting apart of image to tell the digital picture district of needs identification and useless background image district from image acquisition in the resulting entire image.Here adopt the method for searching based on square frame.Standardization is used based on the location specification method of center of gravity the dot array data after handling and is carried out standard, makes it to become 16 * 16 dot array data form.All that obtain are trained with stretching 256 dimensional vectors that become of gray level image matrix, form training and gather with input vector

K=1 wherein, 2 ..., ten category labels of 10 expressions, N _kIt is input sample number of all categories.

Fig. 5 is sorter of the present invention and and the learning time comparative result figure of other sorter.In Fig. 5, the learning time of adopting two kinds of matrix module neural networks of cluster dividing in K mean cluster method and the affinity propagation clustering method respectively than other sorter as: single perceptron, one-against-others modular neural network and adopt the matrix module neural network etc. of the method for random division bunch to expend still less.

Fig. 6 is sorter of the present invention and and the nicety of grading comparative result figure of other sorter.In Fig. 6, adopt respectively two kinds of matrix module neural networks of method that K average and affinity propagate cluster dividing to the recognition accuracy of test set than other sorter as single perceptron, one-against-others modular neural network and adopt the matrix module neural network etc. of method of random division bunch higher to the recognition accuracy of test set.

Embodiment:

United States postal service database (USPS) picks up from actual hand-written postcode, and sample is the normalized gray level image of 16*16.Data centralization comprises 7291 training samples and 2007 test sample books altogether.All samples are represented numeral 0 to 9 respectively from 10 classifications.The number of training that each classification comprised sees Table 4.5, and they are unbalanced.USPS is a relatively classification task of difficulty, because people's naked eyes reach 2.5% to the identification error rate of these data.In the experiment,, all Digital Character Image are converted into the gray level image of 8*8 according to formula (6) in order to simplify study.

$I_{64}(i,j)=\frac{I_{256}(2i-\mathrm{1,2}j-1)+I_{256}(2i-\mathrm{1,2}j)+I_{256}(2i,2j-1)+I_{256}(2i,2j)}{4},---\left(6\right)$

I wherein, j=1,2 ..., 8

The USPS training set

Numerical character	0	1	2	3	4	5	6	7	8	9
											Number of training	1194	1005	731	658	652	556	664	645	542	644

A. training set is divided

With affinity propagation clustering method each classification in the training set is carried out a bunch division, wherein parameter a is set to-1,0,0.5 respectively.Simultaneously, also with K mean cluster method each classification in the training set is carried out the division of identical submanifold number with random approach respectively.

When different parameters a value being set with the affinity propagation clustering, the number of clusters order that is divided of all categories

B. classifier design

Adopt the sorter that designs among the present invention, respectively at three different divided rank, design comprises the matrix module neural network of 4031,12035 and 40736 SN9701 modules respectively.In the matrix module neural network, each SN9701 chip module all adopts dynamic structure and the learning algorithm that designs among the present invention.Simultaneously, in order relatively needing, also to have constructed single three layers of perceptron respectively and be used for the USPS classification with the one-against-others modular neural network that comprises 10 SN9701 chip modules.

The neural network number that comprises different hidden unit numbers in each sorter.Wherein ' NH ' represents hidden unit number in the perceptron, and NH=0 means that this network is two-layer perceptron, otherwise is three layers of perceptron.

D. categorised decision

Adopt the integrated sorting algorithm among the present invention integrated, and carry out categorised decision the network output matrix in the matrix module neural network.From Fig. 5 and Fig. 6 result as can be seen, more single three layers of perceptron, the one-against-others modular neural network, and the matrix module neural network of dividing submanifold with random approach, the learning time of sorter can not only be shortened among the present invention with the matrix module neural network of K mean cluster and affinity Law of Communication division submanifold greatly, the nicety of grading of sorter can also be improved.When training set was divided, submanifold was divided thin more, and the sorter precision is high more; Though the learning time of sorter can increase to some extent, still can save a lot of learning times than single three layers of perceptron and one-against-others modular neural network.

Claims (4)

1, a kind of handwriting digital automatic identification method based on modular neural network is characterized in that:

Obtain the handwriting digital sample from ccd image sensor, obtain original data with CPLD as the control center of image capturing system, with DSP as graphics processing unit, realize the automatic acquisition process of image, finish image and gathered fast and storage, then digital picture has been carried out series of preprocessing;

Described to digital picture carry out series of preprocessing be to image carry out binaryzation, level and smooth, cut apart, standardize, obtain input signal, behind the gray level image that obtains character 16*16 pixel, all that obtain are trained with stretching 256 dimensional vectors that become of gray level image matrix, form training and gather χ with input vector _k, k=1 wherein, 2 ..., ten category labels of 10 expressions;

Do you judge that the input sample is training sample or test sample book? training sample then carries out the space-like division and the classifier design of training set successively in this way, and the space-like of described training set is divided, and is that input vector is gathered χ _kBe divided into D _kIndividual submanifold χ _d ^(k), d=1,2 ..., D _k, all submanifolds are formed

Individual submanifold is right Each submanifold is to importing as the training of one two class sorter;

If test sample book is then carried out categorised decision, sorter comprises two ingredients:

A neural network SN 9701 rectangular array and an integrated computer, wherein neural network SN 9701 rectangular array is made of a plurality of SN9701 chips, it is the neural network integration modules of the single output of many inputs, and its inside mainly becomes circuit, feature weights to adjust circuit, performance index decision circuitry and function by the Chebyshev polynomials graupel to form circuit and form;

Described integrated computer is made of a totalizer and a divider, and input vector produces a neural network output matrix through neural network SN 9701 rectangular array, produces final categorised decision according to this neural network output matrix by integrated computer again;

Totalizer in the described integrated computer and divider are used to calculate the mean value of following formula (1), estimate to import the posterior probability that sample x belongs to di submanifold in the i class

$=\frac{1}{D-D_i}\underset{\underset{j\≠i}{j=1}}{\overset{K}{\mathrm{\Σ}}}\underset{\mathrm{dj}=1}{\overset{D_j}{\mathrm{\Σ}}}{o}_{{\mathrm{di}}^{\left(i\right)}{\mathrm{dj}}^{\left(j\right)}}/\underset{k=1}{\overset{K}{\mathrm{\Σ}}}\underset{\mathrm{dk}=1}{\overset{D_k}{\mathrm{\Σ}}}\left(\frac{1}{{D-D}_k}\underset{\underset{j\≠i}{j=1}}{\overset{K}{\mathrm{\Σ}}}\underset{\mathrm{dj}=1}{\overset{D_j}{\mathrm{\Σ}}}{o}_{{\mathrm{dk}}^{\left(k\right)}{\mathrm{dj}}^{\left(j\right)}}\right)---\left(1\right)$

Wherein M is a neural network SN 9701 rectangular array,

It is the neural network element

The output valve of SN9701 chip is by the 7th port OUT output of SN9701 chip, and then final classification judgement is according to being: ign x → c _{K '},

2, a kind of handwriting digital automatic identification method based on modular neural network according to claim 1 is characterized in that: in the design of described sorter, neural network SN 9701 rectangular array is expressed as formula (3) form:

$M=\left(\begin{array}{ccccc}\mathrm{\Φ}& M_{12}& M_{13}& ......& M_{1K}\\ M_{21}& \mathrm{\Φ}& M_{23}& ......& M_{2K}\\ & ......& & ......& \\ M_{K1}& M_{K2}& ......& M_{K(K-1)}& \mathrm{\Φ}\end{array}\right),---\left(3\right)$

Wherein $M_{\mathrm{ij}}=\left(\begin{array}{cccc}{P}_{1^{\left(i\right)}{1}^{\left(j\right)}}& P_{1^{\left(i\right)}{2}^{\left(j\right)}}& ......& P_{1^{\left(i\right)}{D_j}^{\left(j\right)}}\\ P_{2^{\left(i\right)}{1}^{\left(j\right)}}& P_{2^{\left(i\right)}{2}^{\left(j\right)}}& ......& P_{2^{\left(i\right)}{D_j}^{\left(j\right)}}\\ & ......& ......& \\ P_{{D_i}^{\left(i\right)}{1}^{\left(j\right)}}& P_{{D_i}^{\left(i\right)}{2}^{\left(j\right)}}& ......& P_{{D_i}^{\left(i\right)}{D_j}^{\left(j\right)}}\end{array}\right)=\left(P_{{\mathrm{di}}^{\left(i\right)}{\mathrm{dj}}^{\left(j\right)}}\right)---\left(4\right)$

Wherein K is the classification number, i, and j=1,2 ..., K, i ≠ j, di=1,2 ..., D _i, dj=1,2 ..., D _j, neural network

Be single output, its training sample set is expressed as formula (5)

$S_{{\mathrm{di}}^{\left(i\right)}{\mathrm{dj}}^{\left(j\right)}}=\{{(X_l^{{\mathrm{di}}^{\left(i\right)}},1)}_{l=1}^{N_{\mathrm{di}}^{\left(i\right)}}\∪{(X_l^{{\mathrm{dj}}^{\left(j\right)}},0)}_{l=1}^{N_{\mathrm{dj}}^{\left(j\right)}}\},---\left(5\right)$

3, a kind of handwriting digital automatic identification method according to claim 1 based on modular neural network, it is characterized in that: in the described classifier design, neural network is the single output of a multilayer perceptron, each perceptron is a SN9701 chip, dynamically construct its hidden unit number, and adopt back propagation learning algorithm study, the dynamic structure and the learning process of each perceptron is as follows in the neural network:

A, initialization S1=0;

B, perceptron that comprises S1 hidden unit of structure, wherein S1=0 represents that this perceptron is a double-layer structure, otherwise is three-decker;

C, employing learning algorithm are trained this perceptron, in certain big iterations, if end condition can satisfy, and then whole dynamic structure and study end; Otherwise put S1=S1+1, change b.

4, a kind of handwriting digital automatic identification method based on modular neural network according to claim 1 is characterized in that: it is with K mean cluster method or affinity propagation clustering method that the space-like of described training set is divided.

Images