CN105184226A - Digital identification method, digital identification device, neural network training method and neural network training device - Google Patents

Abstract

The invention discloses a digital identification method, a digital identification device, a neural network training method and a neural network training device. The digital identification method comprises the steps of acquiring a digital image sample to be identified, and identifying a digital category corresponding to the digital image sample to be identified through a neural network after training. The neural network after training is trained by the steps as follows: acquiring digital detection samples; detecting a preset neural network by using the digital detection samples, wherein the preset neural network is a neural network trained by digital training samples; judging whether the accuracy rate of the detection results reaches a preset threshold; and when judging that the accuracy rate of the detection results does not reach the preset threshold, adjusting the number of nodes output by the preset neural network, and using an error sample in the digital detection samples to retrain the preset neural network to obtain the neural network after training. The technical problem in the prior art that the precision of digital identification is low is solved.

Description

Digit recognition method and device and neural network training method and device

Technical field

The present invention relates to digital identification field, in particular to a kind of digit recognition method and device and neural network training method and device.

Background technology

Exam paper marking electronization is one of form evolution trend of going over examination papers now, uses the instrument such as computing machine, can improve the automaticity of the process of going over examination papers, thus reaches raising and to go over examination papers speed, reduces the object of manpower consumption.In middle and primary schools' daily teaching practice process, paper remains the main media of paper, and how the process of going over examination papers of papery paper and computer technology to be combined, be the problem that Many researchers is paid close attention to.Particularly subjective item part, needs the mark of digital form, instead of simple to mistake.Prior art adopts man-machine interaction mode usually, namely the person that do not go over examination papers inputs the mark that should provide on computers, or on paper, arrange mark full-filling block, adopts the form of filling block of answering to obtain mark in conjunction with machine-read card.Although said method all has certain practicality, but all need artificial intervention, compare with traditional mode of going over examination papers, efficiency of going over examination papers significantly does not promote, and because same tradition custom of going over examination papers differs greatly, the easily efficiency because the uncomfortable reduction of operation of teacher is goed over examination papers.

In order to improve efficiency of going over examination papers, prior art adopts the hand-written mark of teacher in digital identification techniques identification papery paper usually, but digital accuracy of identification of the prior art is low, and tolerance is poor, reduces the accuracy of paper fractional statistics.

For the problem that accuracy of identification digital in prior art is low, at present effective solution is not yet proposed.

Summary of the invention

Embodiments provide a kind of digit recognition method and device and neural network training method and device, at least to solve the technical matters that in prior art, digital accuracy of identification is low.

According to an aspect of the embodiment of the present invention, provide a kind of digit recognition method, comprising: obtain digital picture sample to be identified; And the digital classification that the neural network recognization digital picture to be identified sample passing through to have trained is corresponding, wherein, the neural network of having trained is trained in the following manner and is obtained: obtain Digital Detecting sample; Utilize Digital Detecting sample to detect default neural network, wherein, default neural network is train by digital training sample the neural network obtained; Judge whether the accuracy of testing result reaches predetermined threshold value; And when judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, using the sample re-training of makeing mistakes in Digital Detecting sample to preset neural network, obtaining the neural network of having trained.

Further, the number of the default neural network output node of adjustment comprises: judge whether the first sample size reaches first threshold, wherein, the first sample size is the quantity that output node numerical value is less than the sample of makeing mistakes of the first numerical value; When judgement first sample size reaches first threshold, newly-built output node, and the positive sample of sample as newly-built output node of makeing mistakes output node numerical value being less than the first numerical value.

Further, the number that neural network output node is preset in adjustment comprises: judge whether the second sample size reaches Second Threshold, wherein, the second sample size is the quantity that the first output node numerical value and the second output node numerical value are all more than or equal to the sample of makeing mistakes of the first numerical value; When judgement second sample size reaches Second Threshold, first output node and the second output node are merged into the 3rd output node, and the first output node numerical value and the second output node numerical value are all more than or equal to the positive sample of sample as the 3rd output node of makeing mistakes of the first numerical value, wherein, the classification that the first output node is corresponding and classification corresponding to the second output node are the subclass of classification corresponding to the 3rd output node.

Further, the number of the default neural network output node of adjustment comprises: judge whether the 3rd sample size is less than the 3rd threshold value, wherein, the 3rd sample size is the quantity that output node numerical value is more than or equal to the sample of makeing mistakes of the first numerical value; When judgement the 3rd sample size is less than the 3rd threshold value, delete the 4th output node, wherein, the 4th output node is that output node numerical value is more than or equal to output node corresponding to the sample of makeing mistakes of the first numerical value.

Further, there is corresponding relation in default neural network output node and digital classification, a corresponding digital classification of output node.

Further, comprised by digital training sample neural network training: digital training sample is carried out stochastic transformation, generate multiple sub-figure training samples of digital training sample, wherein, the digital classification that multiple sub-figure training sample the is corresponding digital classification corresponding with digital training sample is identical; By multiple sub-figure training sample neural network training, obtain default neural network.

Further, sample of makeing mistakes comprises digital picture exceptional sample, after the neural network obtaining having trained, method also comprises: digital picture exceptional sample is inputed to the neural network of having trained, obtain the first array, wherein, the first array is the array of the output node numerical value composition that digital picture exceptional sample is corresponding; Using the input quantity of the first array as decision tree, obtain the pretreatment mode of digital image abnormity sample, wherein, decision tree is used for the pretreatment mode according to the first array determination digital picture exceptional sample; According to pretreatment mode, pre-service is carried out to digital image abnormity sample, obtain the normal sample of digital picture; And the digital classification that the normal sample of neural network recognization digital picture passing through to have trained is corresponding.

Further, pretreatment mode at least comprise following any one: picture traverse adjust; Picture altitude adjusts; Image top is blocked; Image bottom is blocked.

According to the another aspect of the embodiment of the present invention, additionally provide a kind of digital recognition apparatus, comprising: acquisition module, for obtaining digital picture sample to be identified; And identification module, for the digital classification that the neural network recognization digital picture to be identified sample by having trained is corresponding, wherein, the neural network of having trained is trained in the following manner and is obtained: obtain Digital Detecting sample; Utilize Digital Detecting sample to detect default neural network, wherein, default neural network is train by digital training sample the neural network obtained; Judge whether the accuracy of testing result reaches predetermined threshold value; When judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, uses the sample re-training of makeing mistakes in Digital Detecting sample to preset neural network, obtains the neural network of having trained.

According to the another aspect of the embodiment of the present invention, additionally provide a kind of neural network training method, comprising: obtain Digital Detecting sample; Utilize Digital Detecting sample to detect default neural network, wherein, default neural network is train by digital training sample the neural network obtained; Judge whether the accuracy of testing result reaches predetermined threshold value; And when judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, the sample re-training of makeing mistakes in Digital Detecting sample is used to preset neural network.

Further, the number of the default neural network output node of adjustment comprises: judge whether the first sample size reaches first threshold, wherein, the first sample size is the quantity that output node numerical value is less than the sample of makeing mistakes of the first numerical value; When judgement first sample size reaches first threshold, newly-built output node, and the positive sample of sample as newly-built output node of makeing mistakes output node numerical value being less than the first numerical value.

Further, the number that neural network output node is preset in adjustment comprises: judge whether the second sample size reaches Second Threshold, wherein, the second sample size is the quantity that the first output node numerical value and the second output node numerical value are all more than or equal to the sample of makeing mistakes of the first numerical value; When judgement second sample size reaches Second Threshold, first output node and the second output node are merged into the 3rd output node, and the first output node numerical value and the second output node numerical value are all more than or equal to the positive sample of sample as the 3rd output node of makeing mistakes of the first numerical value, wherein, the classification that the first output node is corresponding and classification corresponding to the second output node are the subclass of classification corresponding to the 3rd output node.

Further, the number of the default neural network output node of adjustment comprises: judge whether the 3rd sample size is less than the 3rd threshold value, wherein, the 3rd sample size is the quantity that output node numerical value is more than or equal to the sample of makeing mistakes of the first numerical value; When judgement the 3rd sample size is less than the 3rd threshold value, delete the 4th output node, wherein, the 4th output node is that output node numerical value is more than or equal to output node corresponding to the sample of makeing mistakes of the first numerical value.

According to the another aspect of the embodiment of the present invention, additionally provide a kind of neural metwork training device, comprising: acquisition module, for obtaining Digital Detecting sample; Detection module, for utilizing Digital Detecting sample to detect default neural network, wherein, default neural network is train by digital training sample the neural network obtained; Judge module, for judging whether the accuracy of testing result reaches predetermined threshold value; And adjusting module, for when judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, uses the sample re-training of makeing mistakes in Digital Detecting sample to preset neural network.

In embodiments of the present invention, adopt the digital classification that deep neural network identification digital picture to be identified sample is corresponding, solve the technical matters that in prior art, digital accuracy of identification is low, reach the object improving the accuracy rate that numeral identifies, thus achieve raising and to go over examination papers efficiency, the technique effect of the mark accuracy that ensures to go over examination papers.

Accompanying drawing explanation

Accompanying drawing described herein is used to provide a further understanding of the present invention, and form a application's part, schematic description and description of the present invention, for explaining the present invention, does not form inappropriate limitation of the present invention.In the accompanying drawings:

Fig. 1 is the process flow diagram of the neural network training method according to the embodiment of the present invention;

Fig. 2 is the neural network annexation schematic diagram according to the embodiment of the present invention;

Fig. 3 is the schematic diagram of the neural metwork training device according to the embodiment of the present invention;

Fig. 4 is the process flow diagram of the digit recognition method according to the embodiment of the present invention; And

Fig. 5 is the schematic diagram of the digital recognition apparatus according to the embodiment of the present invention.

Embodiment

The present invention program is understood better in order to make those skilled in the art person, below in conjunction with the accompanying drawing in the embodiment of the present invention, technical scheme in the embodiment of the present invention is clearly and completely described, obviously, described embodiment is only the embodiment of a part of the present invention, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, should belong to the scope of protection of the invention.

It should be noted that, term " first ", " second " etc. in instructions of the present invention and claims and above-mentioned accompanying drawing are for distinguishing similar object, and need not be used for describing specific order or precedence.Should be appreciated that the data used like this can be exchanged in the appropriate case, so as embodiments of the invention described herein can with except here diagram or describe those except order implement.In addition, term " comprises " and " having " and their any distortion, intention is to cover not exclusive comprising, such as, contain those steps or unit that the process of series of steps or unit, method, system, product or equipment is not necessarily limited to clearly list, but can comprise clearly do not list or for intrinsic other step of these processes, method, product or equipment or unit.

According to the embodiment of the present invention, provide a kind of embodiment of the method for neural network training method, it should be noted that, can perform in the computer system of such as one group of computer executable instructions in the step shown in the process flow diagram of accompanying drawing, and, although show logical order in flow charts, in some cases, can be different from the step shown or described by order execution herein.

Fig. 1 is the process flow diagram of the neural network training method according to the embodiment of the present invention, and as shown in Figure 1, this neural network training method comprises the steps:

Step S102, obtains Digital Detecting sample;

Step S104, utilizes Digital Detecting sample to detect default neural network, and wherein, default neural network is train by digital training sample the neural network obtained;

Step S106, judges whether the accuracy of testing result reaches predetermined threshold value;

Step S108, when judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, uses the sample re-training of makeing mistakes in Digital Detecting sample to preset neural network.

By above-mentioned steps, can realize improving neural network to the precision of numeral identification, and then solve the problem that digital accuracy of identification is low in prior art, reach raising and to go over examination papers efficiency, improve the technique effect of precision of going over examination papers.

Alternatively, Digital Detecting sample in this embodiment and digital training sample are the numeral sample in numeral sample storehouse, wherein, store a large amount of numeral samples in numeral sample storehouse, this numeral sample can be handwritten numeral sample, also can be non-handwritten numeral sample.It should be noted that, the embodiment of the present invention is described for handwritten numeral sample, but does not represent the embodiment of the present invention and be not suitable for unscripted numeral sample.

Numeral sample storehouse is that the numeral sample storehouse obtained by existing disclosed digital picture Sample Storehouse and sampling carries out merging or rejecting operates the Sample Storehouse obtained.Store a large amount of different writing styles in existing disclosed digital picture Sample Storehouse or write the numeral sample of shape, wherein, be no lack of the numeral sample existing and do not meet normal person's writing style or the hand-written custom of exam paper marking, therefore those numeral samples not meeting normal person's writing style or the hand-written custom of exam paper marking in existing disclosed digital picture Sample Storehouse are carried out rejecting operation when obtaining numeral sample by the embodiment of the present invention.Sample and to extract the Sample Storehouse that the numeral sample that obtains forms in paper that the numeral sample storehouse that obtains read and made comments by teacher from true examination.

Alternatively, the numeral sample storehouse set up based on the hand-written custom of teacher can in the following manner:

Mode 1: obtain the paper that teacher read and made comments from true examination, mark teacher's handwritten numeral region, extracts the digital cutting image in numeric area, to obtain numeral sample;

Mode 2: design handwritten numeral capture card, obtains the cutting image of volunteer's handwritten numeral, to obtain numeral sample.

Wherein, from true examination, obtain the paper that teacher read and made comments in mode 1, mark teacher's handwritten numeral region, extract the cutting image in numeric area, can comprise the following steps with the detailed process obtaining numeral sample:

Step S1, the paper that teacher is read and made comments as original image, from original image identification full-filling area identification frame.

Paper comprises the handwritten numeral region of hand-written mark when one or more is read and made comments for teacher, and this handwritten numeral region can as the full-filling area identification frame in original image.From original image, identify that full-filling area identification frame can use the threshold value pre-established that original image is carried out binaryzation operation, change original image into black white image, then use the rectangle in Hough transform method identification binary image, identify that the rectangle in binary image is because the handwritten numeral region for hand-written mark in paper is generally rectangle.Due to may multiple rectangle be comprised in the original image that paper is corresponding, but not all rectangle is full-filling area identification frame in this multiple rectangle, also be likely the figure in paper in geometry exercise question, so, after identifying the rectangle in binary image, need the size according to the full-filling area identification frame preset, the rectangle that size conforms to the size of the full-filling area identification frame preset is selected from multiple rectangle, it can be used as the full-filling area identification frame in original image, and export its co-ordinate position information in original image.Original image correspondence has coordinate system for representing position relationship, and this coordinate system can using the top left corner apex of original image as true origin, using the horizontal edge of original image and vertical edge as abscissa axis and axis of ordinates.It should be noted that, choosing of coordinate system can correspondingly adjust according to the actual requirements.

Step S2, generates hand-written cutting image based on color and contrast.

Handwritten numeral in original image is usually different with contrast from the color of other parts, original image is cut into a lot of minute cells, from a lot of smile unit, obtain cutting image corresponding to handwritten numeral.

Step S3, the cutting image corresponding to handwritten numeral carries out region segmentation, union operation, the interfering picture in filtering cutting image.Due to unnecessary stroke can be there is during handwritten numeral, or the situations such as numeral extends to outside full-filling area identification frame, in order to improve Handwritten Digit Recognition precision, after obtaining cutting image corresponding to handwritten numeral, filtering interfering operation is carried out to cutting image, ensure that the accuracy of numeral sample.

Handwritten numeral capture card is designed in mode 2, obtain the cutting image of volunteer's handwritten numeral, can be described as with the detailed process obtaining numeral sample: design handwritten numeral capture card, the form of handwritten numeral capture card is preferably anchor point and adds form, and often row can fill in 20 numerals, amounts to 10 row, often row fills in identical numeral, the first row is filled in " 0 ", and the second row fills out one writing, by that analogy.The form number that handwritten numeral capture card is arranged can set according to the demand of reality to numeral sample quantity, 10 above-mentioned row, the handwritten numeral capture card that often row 20 is digital is a preferred embodiment of the present invention, and the handwritten numeral capture card the present invention for other types illustrates no longer one by one and describes in detail.Each volunteer writes numeral in a handwritten numeral capture card, and in order to ensure that numeral sample is comprehensive, the embodiment of the present invention can arrange several (such as 50 or 100) volunteer, carries out the collection of numeral sample.

Handwritten numeral capture card fill in complete after, utilize the program that writes in advance to identify hand-written Data Acquisition Card, to generate the digital picture sample of well cutting.Consider that volunteer can exist the situation of clerical error in digital writing process, the program write in advance in the embodiment of the present invention can rewrite accordingly or adjust, to adapt to the situation of digital writing mistake.Concrete operations are draw a fork, four angles of four summit connection table gridirons of fork in the Form Frame at the digital place of wrongly writing.

After acquisition numeral sample storehouse, therefrom selected part numeral sample is as digital training sample, trains neural network, obtains default neural network; Selected part numeral sample, as Digital Detecting sample, detects default neural network, to ensure the accuracy of identification of default neural network.

Alternatively, the default neural network in this embodiment is train by digital training sample the neural network obtained.This embodiment describes in detail with the process obtaining default neural network to by digital training sample neural network training for the neural network with 5 Rotating fields, it should be noted that, the neural network with other structures is equally applicable to the neural network training method of the embodiment of the present invention.

Fig. 2 is the neural network annexation schematic diagram according to the embodiment of the present invention, and as shown in Figure 2, each node in neural network stores a numerical value (shaping or floating type), and every one deck of neural network is made up of several nodes.Each node has annexation with several nodes of last layer, annexation by represent, node and the n-th layer label of meaning to be (n-1)th layer of label be i are the weight connected between the node of j, and value is floating number.

The numerical value of ground floor node is inputted by outside, the nodal values of remainder layer is obtained by the numerical evaluation on the connected node of last layer, first the nodal values of these last layers is weighted summation and obtains a new numerical value y, institute's weighted is the connection weight between corresponding node and lower level node, then y obtains f (y) through a compression function f, and this value is the numerical value of lower floor's corresponding node.

The nodal values of neural network successively calculates, and after whole nodal values of last layer have calculated, carries out the calculating of next node layer numerical value.

Excitation function conventional in nodal values computation process has following several:

1) linear function: f (y)=y

2) two-valued function: $f\left(y\right)=\left\{\begin{array}{c}1,y>=0\\ 0,otherwise\end{array}\right.$

3) linear function blocked is with: $f\left(y\right)=\left\{\begin{array}{c}y,y>=0\\ 0,otherwise\end{array}\right.$

4) S type function: $f\left(y\right)=\frac{1}{1+e^{-y}}$

5) hyperbolic function: $f\left(y\right)=\frac{1-e^{-2y}}{1+e^{-2y}}$

The neural network of this embodiment is made up of 5 node layers, and its connected mode is as follows:

Ground floor is input layer, and network node number is identical with digital picture size, and input quantity is numeral sample image pixel value.Such as, the gray level image size of the numeral sample of input is 29*29, then input layer has 841 nodes altogether.

The second layer is sample level, and network is made up of 6 convolved images, and convolved image is produced on the digital image by specific convolution kernel effect.The corresponding convolved image pixel value of this layer network node, convolution process is realized by the connection of ground floor and second layer network node and weight.A typical node input/output model is y _{c, i, j}for node exports, D represents convolution kernel set of coordinate values, and g (i) is respective coordinates value before convolution, and c is convolved image numbering, and (i, j) represents the coordinate on this convolved image, w _k,lthe weight of convolution kernel in respective coordinates.This layer comprises 6 different convolution kernels, convolution kernel is the matrix of size 5*5, the mode acted on numeral sample image is that interlacing is every row, the numerical value that convolution kernel each position calculation on image obtains is exactly the numerical value of this node layer, the image size obtained due to convolution is 13*13, so amount to (13*13*6) individual node, each node associates with 5*5 node of ground floor respectively, in addition each node also corresponding side-play amount.

Third layer is sample level, and network is made up of 50 convolved images, and convolved image is acted on whole 6 images of the second layer by specific convolution kernel and obtains.The corresponding convolved image pixel value of this layer network node.This layer comprises 50 convolution kernels, convolution kernel is the matrix of size 5*5, the mode of action is that convolution is carried out in the relevant position of the second layer 6 convolved images, obtaining image size is 5*5, this layer amounts to (5*5*50) individual node, each node associates with 5*5 node of the second layer respectively, and a corresponding side-play amount.

4th layer is learning layer, is made up of 100 nodes.Each node and all nodes of third layer, namely 1250 nodes connect.

Layer 5 is output layer, is made up of N number of node, and N represents final digital class number, and namely the interstitial content of this layer is identical with digital class number, and digital classification default number is 10, respectively corresponding 0-9 this be numeral.Each node is connected with 100 nodes of the 4th layer.The initialization values connected is equally distributed random number.

Alternatively, can be described below with the process obtaining default neural network by digital training sample neural network training:

Logarithm word train sample does random perturbation process, and such as logarithm word train sample carries out image translation, rotation etc.; Digital training sample is inputed to the input layer of neural network; The each node layer numerical value of neural network is calculated successively according to formula; According to output layer nodal values, select suitable weight correction ratio, the back-propagation method of random selecting numeral training sample use error revises the connection weight of each interlayer, when the digital number of training selected reaches certain preset value, adjustment weight correction ratio, the back-propagation method continuing use error revises the connection weight of each interlayer, until weight correction ratio global error that is enough little, neural network is less than threshold value or frequency of training reaches higher limit just deconditioning process.

The direction communication process of error refers to according to the error between the numerical value on bottom node and the numerical value of expectation, the process of the weight between each node layer of bottom-up layer-by-layer correction.

Error is the numerical value that last one deck calculates and target value between difference, be designated as , it is defined as follows:

$E_n=\frac{1}{2}\underset{i}{\Σ}{(x_n^i-T_n^i)}^2$

The error of calculation, about the partial derivative of weight, uses gradient descent method adjustment weight, thus reaches the object reducing error.

Between (n-1)th layer and n-th layer, weight partial derivative is calculated as follows:

$\frac{\∂E_n^p}{\∂x_n^i}=x_n^i-T_n^i$

$\frac{\∂E_n^p}{\∂y_n^i}=G\left(x_n^i\right)\·\frac{\∂E_n^p}{\∂x_n^i}$

$\frac{\∂E_n^p}{\∂w_n^{i,j}}=x_{n-1}^j\·\frac{\∂E_n^p}{\∂y_n^i}$

Wherein, G (x) is node the derived function of corresponding excitation function f (y).All the other each layers successively try to achieve the partial derivative of error about weight by following formula:

$\frac{\∂E_{n-1}^p}{\∂x_{n-1}^i}=\underset{i}{\Σ}{w}_n^{i,k}\·\frac{\∂E_n^p}{\∂y_n^i}$

$\frac{\∂E_{n-1}^p}{\∂y_{n-1}^i}=G\left(x_{n-1}^i\right)\·\frac{\∂E_{n-1}^p}{\∂x_{n-1}^i}$

$\frac{\∂E_{n-1}^p}{\∂w_{n-1}^{i,j}}=x_{n-2}^j\·\frac{\∂E_n^p}{\∂y_n^i}$

Then, then by following formula weight is revised:

Passing through digital training sample neural network training, after obtaining default neural network, in order to improve the accuracy of identification of default neural network, the neural network training method of this embodiment is after neural network is preset in acquisition, utilize Digital Detecting sample to detect default neural network, judge whether the accuracy of testing result reaches predetermined threshold value.

Step S104, Digital Detecting sample is utilized to detect default neural network, wherein, Digital Detecting sample is the part number sample in numeral sample storehouse, the quantity of Digital Detecting sample can be one or more, in order to ensure the accuracy of identification of default neural network, preferably, multiple Digital Detecting sample is adopted to detect default neural network.The process nature utilizing Digital Detecting sample to detect default neural network is for using the input quantity of Digital Detecting sample as default neural network, whether the digital classification corresponding with this Digital Detecting sample reality is consistent to detect the output quantity presetting neural network, if detect that the output quantity of the default neural network digital classification actual corresponding with this Digital Detecting sample is inconsistent, be sample of makeing mistakes by this Digital Detecting sample labeling.After multiple Digital Detecting sample detects default neural network successively, calculate the accuracy of testing result, wherein, the accuracy of testing result is that sample of makeing mistakes accounts for the number percent of Digital Detecting sample, and the accuracy of this testing result can as judging the measurement index presetting neural network recognization precision.

Utilizing Digital Detecting sample, default neural network is detected, after obtaining the accuracy of testing result, perform step S106 and judge whether the accuracy of testing result reaches predetermined threshold value, wherein, predetermined threshold value can adjust accordingly according to actual conditions.If when judging that the accuracy of testing result reaches predetermined threshold value, then illustrate that the accuracy of identification presetting neural network meets the demands, and adjusts default neural network without the need to continuing again; If when judging that the accuracy of testing result does not reach predetermined threshold value, then illustrate that the accuracy of identification presetting neural network does not meet the demands, need to adjust default neural network, until the accuracy of identification presetting neural network meets the demands, namely the accuracy of testing result reaches predetermined threshold value.

Particularly, when judging that the accuracy of testing result does not reach predetermined threshold value, be adjusted to step S108 to default neural network, the number of neural network output node is preset in adjustment, uses the sample re-training of makeing mistakes in Digital Detecting sample to preset neural network.

Presetting neural network output layer node number and be defaulted as 10, respectively corresponding 0-9 ten digital classifications, there is corresponding relation in default neural network output node and digital classification, a corresponding digital classification of output node.When judging that the accuracy of testing result does not reach predetermined threshold value, the neural network training method of this embodiment can adjust the number of default neural network output node.

Alternatively, the number of the default neural network output node of adjustment can comprise: judge whether the first sample size reaches first threshold, wherein, the first sample size is the quantity that output node numerical value is less than the sample of makeing mistakes of the first numerical value; When judgement first sample size reaches first threshold, newly-built output node, and the positive sample of sample as newly-built output node of makeing mistakes output node numerical value being less than the first number Γ value.Need newly-built output node, namely other condition of newly-built numeric class is needed to be expressed as: Γ { x|N (x) <0.7}>C, wherein, Γ represents the first sample size, C represents first threshold, N (x) represents output node numerical value, and the first numerical value is 0.7 herein.Newly-built output node will add in the layer 5 output layer of default neural network, all meet in the Digital Detecting sample of above-mentioned condition sample of makeing mistakes will as other positive sample of numeric class corresponding to this newly-built output node.

Alternatively, the number that neural network output node is preset in adjustment can also comprise: judge whether the second sample size reaches Second Threshold, wherein, the second sample size is the quantity that the first output node numerical value and the second output node numerical value are all more than or equal to the sample of makeing mistakes of the first numerical value; When judgement second sample size reaches Second Threshold, first output node and the second output node are merged into the 3rd output node, and the first output node numerical value and the second output node numerical value are all more than or equal to the positive sample of sample as the 3rd output node of makeing mistakes of the first numerical value, wherein, the classification that the first output node is corresponding and classification corresponding to the second output node are the subclass of classification corresponding to the 3rd output node.Need to merge output node, namely needing to merge other condition of numeric class can be expressed as: judge that the condition needing to merge classification is as follows: Γ { x|N ₁(x)>=0.7, N ₂x ()>=0.7}>D, wherein, Γ represents the second sample size, and D represents Second Threshold, N ₁x () represents the first output node numerical value, N ₂x () represents the second output node numerical value, the first numerical value is 0.7 herein.The number of elements that above formula can be expressed as the common factor of the numeral sample collection that degree of membership is higher between two digital classifications merges this two digital classifications when being greater than Second Threshold D, and these two digital classifications are necessary for the subclass of same numeral.Other positive sample of numeric class after merging is the union of the positive sample of these two digital classifications.

Alternatively, the number of the default neural network output node of adjustment can also comprise: judge whether the 3rd sample size is less than the 3rd threshold value, wherein, the 3rd sample size is the quantity that output node numerical value is more than or equal to the sample of makeing mistakes of the first numerical value; When judgement the 3rd sample size is less than the 3rd threshold value, delete the 4th output node, wherein, the 4th output node is that output node numerical value is more than or equal to output node corresponding to the sample of makeing mistakes of the first numerical value.Need to delete output node, namely need to delete other condition of numeric class can be expressed as: Γ { x|N (x) >=0.7}<E, wherein, Γ represents the 3rd sample size, E represents the 3rd threshold value, N (x) represents output node numerical value, and the first numerical value is 0.7 herein.Above formula can be expressed as degree of membership higher numeral sample quantity when being less than the 3rd threshold value E, deletes this digital classification.

Preset neural network to adjust through above-mentioned output node, namely after digital classification adjustment, its output node number may be greater than 10, i.e. same digital classification one or more subclass corresponding.It should be noted that, step S102 in this embodiment to step S108 is a complete circulation, after default neural network output node is correspondingly adjusted, continue to obtain Digital Detecting sample from numeral sample storehouse, then detect default neural network with Digital Detecting sample, the circulation of this detection is until stop when the accuracy of testing result reaches predetermined threshold value.

The neural network training method Digital Detecting sample of this embodiment detects default neural network, when the accuracy of testing result does not reach predetermined threshold value, default neural network output node is adjusted, comprise newly-built output node, merge output node and delete output node, drastically increased accuracy of identification and the tolerance of default neural network by above-mentioned adjustment.

The embodiment of the present invention additionally provides one neural network training method alternatively, this alternatively neural network training method can be described as:

In the process of training at actual acquisition numeral sample and to neural network, due to writing style difference, same class numeral may have several different typical ways of writing, therefore, this embodiment adopts new classification based training method, by automatically adjusting the number of output node in neural network, reach the result of the neural metwork training that becomes more meticulous.

By digital training sample neural network training, be described in detail in the above-described embodiments with the process obtaining default neural network, repeat no more herein.Digital Detecting sample is utilized to be described below the detailed process that default neural network detects:

Whenever training global error to be less than predetermined threshold value, utilize numeral sample to detect neural network, obtain the digital classification that each numeral sample is corresponding, multiple digital classification may belong to same numeral here.

The result that statistical figure sample is made mistakes in neural network detects, carries out statistic of classification according to " actual numbers-detection numeral " classification.Such as, the numeral on numeral sample is " 2 ", but numeric results corresponding to the classification that detects of neural network is " 3 ", then under this pictures being recorded in " 2-3 " classification.

Check the sample size in " actual numbers-detection numeral " classification, if sample size is greater than designated value (such as 100), then add new numeral sample classification, the mode that item name is continued to use " actual numbers-detection numeral ".

Check classification polyisomenism in original sample, belong to the ratio of repeated sample between digital two classifications according to the formula inspection of specifying, if ratio is greater than threshold value, then merge this two classifications.

Specify classification again to each numeral sample, according to neural network output node numerical value, be included into by each numeral sample in the classification of newly having classified, same numeral sample can corresponding multiple classification.Classifying mode is: if " actual numbers " of this numeral sample is identical with " actual numbers " of classification, and the neural network Output rusults in " detecting numeral " classification is greater than specifies threshold value (such as 0.7), then included in corresponding " actual numbers-detection numeral " classification.

The output layer of neural network is rebuild according to the digital classification after adjustment.Training numeral sample is continued by the neural network built.

According to the embodiment of the present invention, additionally provide a kind of device embodiment of neural metwork training device, it should be noted that, the neural metwork training device of this embodiment may be used for performing the neural network training method in the embodiment of the present invention, and the neural network training method in the embodiment of the present invention can perform in the neural metwork training device of this embodiment.

Fig. 3 is the schematic diagram of the neural metwork training device according to the embodiment of the present invention, and as shown in Figure 3, this neural metwork training device comprises: acquisition module 30, detection module 32, judge module 34 and adjusting module 36.

Acquisition module 30, for obtaining Digital Detecting sample.

Detection module 32, for utilizing Digital Detecting sample to detect default neural network, wherein, default neural network is train by digital training sample the neural network obtained.

Judge module 34, for judging whether the accuracy of testing result reaches predetermined threshold value.

Adjusting module 36, for when judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, uses the sample re-training of makeing mistakes in Digital Detecting sample to preset neural network.

Alternatively, the adjusting module 36 in this embodiment can comprise: the first judge module, and for judging whether the first sample size reaches first threshold, wherein, the first sample size is the quantity that output node numerical value is less than the sample of makeing mistakes of the first numerical value; Newly-built module, for when judgement first sample size reaches first threshold, newly-built output node, and the positive sample of sample as newly-built output node of makeing mistakes output node numerical value being less than the first numerical value.

Alternatively, adjusting module 36 in this embodiment can also comprise: the second judge module, for judging whether the second sample size reaches Second Threshold, wherein, the second sample size is the quantity that the first output node numerical value and the second output node numerical value are all more than or equal to the sample of makeing mistakes of the first numerical value; Merge module, for when judgement second sample size reaches Second Threshold, first output node and the second output node are merged into the 3rd output node, and the first output node numerical value and the second output node numerical value are all more than or equal to the positive sample of sample as the 3rd output node of makeing mistakes of the first numerical value, wherein, the classification that the first output node is corresponding and classification corresponding to the second output node are the subclass of classification corresponding to the 3rd output node.

Alternatively, the adjusting module 36 in this embodiment can also comprise: the 3rd judge module, and for judging whether the 3rd sample size is less than the 3rd threshold value, wherein, the 3rd sample size is the quantity that output node numerical value is more than or equal to the sample of makeing mistakes of the first numerical value; Removing module, for when judgement the 3rd sample size is less than the 3rd threshold value, deletes the 4th output node, and wherein, the 4th output node is that output node numerical value is more than or equal to output node corresponding to the sample of makeing mistakes of the first numerical value.

The neural metwork training device of this embodiment obtains Digital Detecting sample by acquisition module 30, Digital Detecting sample is utilized to detect default neural network by detection module 32, wherein, default neural network is train by digital training sample the neural network obtained, judge whether the accuracy of testing result reaches predetermined threshold value by judge module 34, by adjusting module 36 when judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, the sample re-training of makeing mistakes in Digital Detecting sample is used to preset neural network.Solve the low problem of neural network recognization precision by the neural metwork training device of this embodiment, reach the technique effect improving neural network recognization precision and tolerance.

According to the embodiment of the present invention, additionally provide a kind of embodiment of the method for digit recognition method, it should be noted that, can perform in the computer system of such as one group of computer executable instructions in the step shown in the process flow diagram of accompanying drawing, and, although show logical order in flow charts, in some cases, can be different from the step shown or described by order execution herein.

Fig. 4 is the process flow diagram of the digit recognition method according to the embodiment of the present invention, and as shown in Figure 4, this digit recognition method can be applied in paper examination result process, also can be applied to the occasion that other need to identify handwritten numeral.Particularly, the digit recognition method of this embodiment comprises:

Step S402, obtains digital picture sample to be identified;

Step S404, by the digital classification that the neural network recognization digital picture to be identified sample of having trained is corresponding, wherein, the neural network of having trained is trained in the following manner and is obtained: obtain Digital Detecting sample; Utilize Digital Detecting sample to detect default neural network, wherein, default neural network is train by digital training sample the neural network obtained; Judge whether the accuracy of testing result reaches predetermined threshold value; And when judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, using the sample re-training of makeing mistakes in Digital Detecting sample to preset neural network, obtaining the neural network of having trained.

Pass through above-mentioned steps, the accuracy of identification of the neural network that the training obtained completes is greatly improved, and then solve the technical matters that in correlation technique, digital accuracy of identification is low, reach the object improving the accuracy rate that numeral identifies, thus achieve raising and to go over examination papers efficiency, the technique effect of the mark accuracy that ensures to go over examination papers.

It should be noted that, it is identical with the detailed process obtaining numeral sample in the neural network training method of the embodiment of the present invention that step S402 obtains digital picture sample to be identified, the training process of the neural network of having trained in step S404 is identical with the training process of the neural network training method of the embodiment of the present invention, therefore, no longer it is described in detail herein.

Alternatively, can be comprised by digital training sample neural network training: digital training sample is carried out stochastic transformation, generate multiple sub-figure training samples of digital training sample, wherein, the digital classification that multiple sub-figure training sample the is corresponding digital classification corresponding with digital training sample is identical; By multiple sub-figure training sample neural network training, obtain default neural network.Do random perturbation operation by logarithm word train sample, such as image translation, rotation etc., reach the effect improving neural network recognization precision.

Carry out in digital identifying actual, the situation of a lot of digital identification error causes by filling in image cutting position or the proportional jitter of causing lack of standardization, the digit recognition method of this embodiment is introduced decision tree and is identified this kind of situation, and for different Deviation Types, again pre-service is carried out to picture, to reach the object improving digital recognition accuracy further.

Alternatively, the sample of makeing mistakes in this embodiment comprises digital picture exceptional sample, and wherein, digital picture exceptional sample may cause owing to writing the reason such as lack of standardization.After the neural network obtaining having trained, the digit recognition method of this embodiment also comprises: digital picture exceptional sample is inputed to the neural network of having trained, obtain the first array, wherein, the first array is the array of the output node numerical value composition that digital picture exceptional sample is corresponding; Using the input quantity of the first array as decision tree, obtain the pretreatment mode of digital image abnormity sample, wherein, decision tree is used for the pretreatment mode according to the first array determination digital picture exceptional sample; According to pretreatment mode, pre-service is carried out to digital image abnormity sample, obtain the normal sample of digital picture; And the digital classification that the normal sample of neural network recognization digital picture passing through to have trained is corresponding.Wherein, pretreatment mode at least comprise following any one: picture traverse adjust; Picture altitude adjusts; Image top is blocked; Image bottom is blocked.In addition, pretreatment mode can also increase according to the actual requirements accordingly, such as image rotation, image stretch etc.

This embodiment use decision tree reduces the error when neural network of having trained identifies numeral, wherein, being input as of decision tree: numeral sample image is after the neural network recognization of having trained, the array V of all output node numerical value compositions, the output of decision tree is: Image semantic classification mode, comprises picture traverse adjustment; Picture altitude adjusts; Image top is blocked; Image bottom is blocked.

Utilize four kinds of pretreatment modes to train decision tree respectively, train positive sample to be the numeral sample image of the identification error caused by corresponding reason, negative sample is other random numeral sample images.The decision tree trained is assessed for the first time recognition result of the neural network completed the training for numeral identification, decision tree judges that needs carry out pre-service again, then to after numeral sample Image semantic classification, use the neural network of having trained to carry out second time and identify.It should be noted that, may occur that decision tree is judged to need to carry out multiple pretreated situation to numeral sample image, now need to carry out second time to the neural network of having trained through multiple pretreated numeral sample image respectively to identify, choose the recognition result as this numeral sample image that degree of membership is the highest.

Alternatively, this embodiment still provides a kind of utilize decision tree correction to train alternatively after neural network numeral identification error method.It should be noted that, neural network herein refers to the neural network of having trained.

Through the calculating of neural network, the output vector of the corresponding neural network of numeral sample image, the numerical value of this output vector is followed successively by the numerical value on neural network output layer node.This vector is reformed, obtains the new vector that size is 10.Its method is the classification belonging to same " actual numbers " in former vector merged, and merging mode is for getting maximal value.Filter out the numeral sample of classification error, and (Image semantic classification mode, comprises picture traverse adjustment to its one of carrying out in multiple pre-service; Picture altitude adjusts; Image top is blocked; Image bottom is blocked, image rotation, image stretch etc.) again by neural computing, if digital judgement is correct, then record this numeral sample at the reformation vector after neural network and make its judicious mapping mode.According to the mode of " actual numbers-Preliminary detection numeral-pretreatment mode ", all numeral samples that can obtain correct digit through pre-service are classified.Add up the span of neural network output vector numerical value corresponding to " actual numbers " classification in each class, and the span of neural network output vector numerical value corresponding to " Preliminary detection numeral " classification.Add decision tree condition: when numeral sample image is after neural network, when numerical value on its output vector correspondence position meets the span of above-mentioned statistics, according to " pretreatment mode " specified by " actual numbers-Preliminary detection numeral-pretreatment mode ", pre-service is carried out to digital picture, and is again detected by neural network.If " actual numbers " during testing result meets " actual numbers-Preliminary detection numeral-mapping mode " and " Preliminary detection numeral ", then using the testing result of pretreated testing result as digital picture.

The digit recognition method of this embodiment is using the input quantity of the digital picture exceptional sample of makeing mistakes in sample after the neural network recognization by having trained as decision tree, decision tree is utilized to carry out corresponding pre-service to digital image abnormity sample, then utilize the neural network of having trained to carry out numeral to the normal sample of pretreated digital picture to identify, use decision tree can drastically increase the accuracy of numeral identification.

Alternatively, the digit recognition method of this embodiment also provides the detecting warning function that mistake is scribbled, and for the situation of severe sign of erasure, this embodiment uses connective histogram to differentiate it; For the situation of mistake write stroke, this embodiment uses piecemeal detection to differentiate it; For the situation that the information of correcting is crossed the border, this embodiment uses connective monitoring to differentiate it; For the situation of incredible figures, this embodiment uses neural network degree of membership to differentiate it.When there is above-mentioned situation in differentiation, the digit recognition method of this embodiment can carry out corresponding pre-service to it, and provides early warning information, by above-mentioned differentiation, also effectively can improve the accuracy that numeral identifies, and then solve the low problem of prior art numeral accuracy of identification.

The digit recognition method of this embodiment utilizes the higher neural network of a kind of accuracy of identification to identify numeral, in order to ensure digital recognition accuracy, on the basis utilizing deep neural network, use decision tree to carry out pre-service to digital image abnormity sample, the normal sample of digital picture after process is carried out numeral by deep neural network again and identifies.The digit recognition method of this embodiment solves the low problem of prior art numeral accuracy of identification, reaches and improves digital recognition accuracy, and then improve the technique effect of go over examination papers efficiency and accuracy.

According to the embodiment of the present invention, additionally provide a kind of device embodiment of digital recognition apparatus, it should be noted that, the digital recognition apparatus of this embodiment may be used for performing the digit recognition method in the embodiment of the present invention, and the digit recognition method in the embodiment of the present invention can perform in the digital recognition apparatus of this embodiment.

Fig. 5 is the schematic diagram of the digital recognition apparatus according to the embodiment of the present invention, and as shown in Figure 5, this digital recognition apparatus comprises: acquisition module 50 and identification module 52.

Acquisition module 50, for obtaining digital picture sample to be identified.

Identification module 52, for the digital classification that the neural network recognization digital picture to be identified sample by having trained is corresponding, wherein, the neural network of having trained is trained in the following manner and is obtained: obtain Digital Detecting sample; Utilize Digital Detecting sample to detect default neural network, wherein, default neural network is train by digital training sample the neural network obtained; Judge whether the accuracy of testing result reaches predetermined threshold value; When judging that the accuracy of testing result does not reach predetermined threshold value, the number of neural network output node is preset in adjustment, uses the sample re-training of makeing mistakes in Digital Detecting sample to preset neural network, obtains the neural network of having trained.

It should be noted that, the neural network that training in the digital recognition apparatus of this embodiment completes is identical with the neural network that network training device obtains by the god by the embodiment of the present invention, the training of network training device can be obtained by the god of the embodiment of the present invention by the neural network of namely having trained, and no longer to the god of the embodiment of the present invention, network training device is carried out repeated description herein.

The digital recognition apparatus of this embodiment utilizes acquisition module 50 to obtain digital picture sample to be identified, the digital classification that the neural network recognization digital picture to be identified sample utilizing identification module 52 to pass through to have trained is corresponding, to obtain digital classification corresponding to digital picture sample to be identified.The digital recognition apparatus of this embodiment solves the low problem of prior art numeral accuracy of identification, reaches and improves digital recognition accuracy, and then improve the technique effect of go over examination papers efficiency and accuracy.

The present invention devises a kind of neural network training method, can obtain the higher neural network of accuracy of identification by this neural network training method, utilizes this neural network to treat discriminating digit image and carries out numeral identification.In the present invention, in neural network training process, numeral classification does not adopt common 10 class modes, but constantly produce new classification according to training result, merge similar classification, this mode has better adaptability for there being the numeral of multiple ways of writing, also takes into account the different writing styles of the person of going over examination papers, fully takes into account the various factors of environment of going over examination papers; Positive sample produces the mode adopting stochastic sampling and stochastic transformation, improves the diversity of sample on the one hand, the Expired Drugs that another aspect avoids appointment sample and may cause; Select whether to cut digital picture according to principium identification result in differentiation process and identify further, the size shape disturbance of this method to numeral during high writing speeds has stronger error-correcting performance.The invention solves the problem that prior art numeral accuracy of identification is low, effectively improve digital recognition accuracy, and then improve go over examination papers efficiency and accuracy.

The invention described above embodiment sequence number, just to describing, does not represent the quality of embodiment.

In the above embodiment of the present invention, the description of each embodiment is all emphasized particularly on different fields, in certain embodiment, there is no the part described in detail, can see the associated description of other embodiments.

In several embodiments that the application provides, should be understood that, disclosed technology contents, the mode by other realizes.Wherein, device embodiment described above is only schematic, the such as division of described unit, can be that a kind of logic function divides, actual can have other dividing mode when realizing, such as multiple unit or assembly can in conjunction with or another system can be integrated into, or some features can be ignored, or do not perform.Another point, shown or discussed coupling each other or direct-coupling or communication connection can be by some interfaces, and the indirect coupling of unit or module or communication connection can be electrical or other form.

The described unit illustrated as separating component or can may not be and physically separates, and the parts as unit display can be or may not be physical location, namely can be positioned at a place, or also can be distributed on multiple unit.Some or all of unit wherein can be selected according to the actual needs to realize the object of the present embodiment scheme.

In addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, also can be that the independent physics of unit exists, also can two or more unit in a unit integrated.Above-mentioned integrated unit both can adopt the form of hardware to realize, and the form of SFU software functional unit also can be adopted to realize.

If described integrated unit using the form of SFU software functional unit realize and as independently production marketing or use time, can be stored in a computer read/write memory medium.Based on such understanding, the part that technical scheme of the present invention contributes to prior art in essence in other words or all or part of of this technical scheme can embody with the form of software product, this computer software product is stored in a storage medium, comprises all or part of step of some instructions in order to make a computer equipment (can be personal computer, server or the network equipment etc.) perform method described in each embodiment of the present invention.And aforesaid storage medium comprises: USB flash disk, ROM (read-only memory) (ROM, Read-OnlyMemory), random access memory (RAM, RandomAccessMemory), portable hard drive, magnetic disc or CD etc. various can be program code stored medium.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (14)

1. a digit recognition method, is characterized in that, comprising:

Obtain digital picture sample to be identified; And

By the digital classification that digital picture sample to be identified described in the neural network recognization of having trained is corresponding, wherein, the neural network that described training completes is trained in the following manner and is obtained:

Obtain Digital Detecting sample;

Utilize described Digital Detecting sample to detect default neural network, wherein, described default neural network is train by digital training sample the neural network obtained;

Judge whether the accuracy of testing result reaches predetermined threshold value; And

When judging that the accuracy of described testing result does not reach described predetermined threshold value, adjust the number of described default neural network output node, use makeing mistakes described in sample re-training in described Digital Detecting sample to preset neural network, obtain the neural network that described training completes.

2. digit recognition method according to claim 1, is characterized in that, the number adjusting described default neural network output node comprises:

Judge whether the first sample size reaches first threshold, wherein, described first sample size is the quantity that output node numerical value is less than the sample of makeing mistakes of the first numerical value;

When judging that described first sample size reaches described first threshold, newly-built output node, and the positive sample of sample as described newly-built output node of makeing mistakes described output node numerical value being less than the first numerical value.

3. digit recognition method according to claim 1, is characterized in that, the number adjusting described default neural network output node comprises:

Judge whether the second sample size reaches Second Threshold, wherein, described second sample size is the quantity that the first output node numerical value and the second output node numerical value are all more than or equal to the sample of makeing mistakes of the first numerical value;

When judging that described second sample size reaches described Second Threshold, described first output node and described second output node are merged into the 3rd output node, and described first output node numerical value and the second output node numerical value are all more than or equal to the positive sample of sample as described 3rd output node of makeing mistakes of the first numerical value, wherein, the classification that described first output node is corresponding and classification corresponding to described second output node are the subclass of classification corresponding to described 3rd output node.

4. digit recognition method according to claim 1, is characterized in that, the number adjusting described default neural network output node comprises:

Judge whether the 3rd sample size is less than the 3rd threshold value, wherein, described 3rd sample size is the quantity that output node numerical value is more than or equal to the sample of makeing mistakes of the first numerical value;

When judging that described 3rd sample size is less than described 3rd threshold value, delete the 4th output node, wherein, described 4th output node is that described output node numerical value is more than or equal to output node corresponding to the sample of makeing mistakes of the first numerical value.

5. digit recognition method according to any one of claim 1 to 4, is characterized in that, described default neural network output node and described digital classification exist corresponding relation, a corresponding digital classification of output node.

6. digit recognition method according to claim 1, is characterized in that, is comprised by described digital training sample neural network training:

Described digital training sample is carried out stochastic transformation, generates multiple sub-figure training samples of described digital training sample, wherein, the digital classification that described multiple sub-figure training sample the is corresponding digital classification corresponding with described digital training sample is identical;

Train described neural network by described multiple sub-figure training sample, obtain described default neural network.

7. digit recognition method according to claim 1, is characterized in that, described in sample of makeing mistakes comprise digital picture exceptional sample, after obtaining the neural network that described training completes, described method also comprises:

Described digital picture exceptional sample is inputed to the neural network that described training completes, obtain the first array, wherein, described first array is the array of the output node numerical value composition that described digital picture exceptional sample is corresponding;

Using the input quantity of described first array as decision tree, obtain the pretreatment mode of described digital picture exceptional sample, wherein, described decision tree is used for the pretreatment mode determining described digital picture exceptional sample according to described first array;

According to described pretreatment mode, pre-service is carried out to described digital picture exceptional sample, obtain the normal sample of digital picture; And

The digital classification that described in the neural network recognization completed by described training, the normal sample of digital picture is corresponding.

8. digit recognition method according to claim 7, is characterized in that, described pretreatment mode at least comprise following any one:

Picture traverse adjusts; Picture altitude adjusts; Image top is blocked; Image bottom is blocked.

9. a digital recognition apparatus, is characterized in that, comprising:

Acquisition module, for obtaining digital picture sample to be identified; And

Identification module, for the digital classification that digital picture sample to be identified described in the neural network recognization by having trained is corresponding, wherein, the neural network that described training completes is trained in the following manner and is obtained: obtain Digital Detecting sample; Utilize described Digital Detecting sample to detect default neural network, wherein, described default neural network is train by digital training sample the neural network obtained; Judge whether the accuracy of testing result reaches predetermined threshold value; When judging that the accuracy of described testing result does not reach described predetermined threshold value, adjust the number of described default neural network output node, use makeing mistakes described in sample re-training in described Digital Detecting sample to preset neural network, obtain the neural network that described training completes.

10. a neural network training method, is characterized in that, comprising:

Obtain Digital Detecting sample;

Utilize described Digital Detecting sample to detect default neural network, wherein, described default neural network is train by digital training sample the neural network obtained;

Judge whether the accuracy of testing result reaches predetermined threshold value; And

When judging that the accuracy of described testing result does not reach described predetermined threshold value, adjusting the number of described default neural network output node, using makeing mistakes described in sample re-training in described Digital Detecting sample to preset neural network.

11. neural network training method according to claim 10, is characterized in that, the number adjusting described default neural network output node comprises:

Judge whether the first sample size reaches first threshold, wherein, described first sample size is the quantity that output node numerical value is less than the sample of makeing mistakes of the first numerical value;

When judging that described first sample size reaches described first threshold, newly-built output node, and the positive sample of sample as described newly-built output node of makeing mistakes described output node numerical value being less than the first numerical value.

12. neural network training method according to claim 10, is characterized in that, the number adjusting described default neural network output node comprises:

Judge whether the second sample size reaches Second Threshold, wherein, described second sample size is the quantity that the first output node numerical value and the second output node numerical value are all more than or equal to the sample of makeing mistakes of the first numerical value;

When judging that described second sample size reaches described Second Threshold, described first output node and described second output node are merged into the 3rd output node, and described first output node numerical value and the second output node numerical value are all more than or equal to the positive sample of sample as described 3rd output node of makeing mistakes of the first numerical value, wherein, the classification that described first output node is corresponding and classification corresponding to described second output node are the subclass of classification corresponding to described 3rd output node.

13. neural network training method according to claim 10, is characterized in that, the number adjusting described default neural network output node comprises:

Judge whether the 3rd sample size is less than the 3rd threshold value, wherein, described 3rd sample size is the quantity that output node numerical value is more than or equal to the sample of makeing mistakes of the first numerical value;

When judging that described 3rd sample size is less than described 3rd threshold value, delete the 4th output node, wherein, described 4th output node is that described output node numerical value is more than or equal to output node corresponding to the sample of makeing mistakes of the first numerical value.

14. 1 kinds of neural metwork training devices, is characterized in that, comprising:

Acquisition module, for obtaining Digital Detecting sample;

Detection module, for utilizing described Digital Detecting sample to detect default neural network, wherein, described default neural network is train by digital training sample the neural network obtained;

Judge module, for judging whether the accuracy of testing result reaches predetermined threshold value; And

Adjusting module, for when judging that the accuracy of described testing result does not reach described predetermined threshold value, adjusts the number of described default neural network output node, uses makeing mistakes described in sample re-training in described Digital Detecting sample to preset neural network.