CN107590497A - Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks - Google Patents

Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks Download PDF

Info

Publication number
CN107590497A
CN107590497A CN201710855035.7A CN201710855035A CN107590497A CN 107590497 A CN107590497 A CN 107590497A CN 201710855035 A CN201710855035 A CN 201710855035A CN 107590497 A CN107590497 A CN 107590497A
Authority
CN
China
Prior art keywords
mrow
mtd
convolutional neural
neural networks
network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201710855035.7A
Other languages
Chinese (zh)
Inventor
赵辉
王艳美
刘真三
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN201710855035.7A priority Critical patent/CN107590497A/en
Publication of CN107590497A publication Critical patent/CN107590497A/en
Pending legal-status Critical Current

Links

Landscapes

  • Image Analysis (AREA)
  • Character Discrimination (AREA)

Abstract

The present invention proposes the Off-line Handwritten Chinese method using depth convolutional neural networks, specifically, first the handwritten Chinese character of HWDB1.1 databases is rotated using spatial alternation network, Pan and Zoom.So that the Chinese character picture in script optional position, ratio and direction is corrected.The hand-written data collection of correction is input to convolutional neural networks afterwards classification is identified.Wherein original picture is rotated, the warp parameters increment of Pan and Zoom be by a convolutional neural networks or neutral net constantly by backpropagation Adjustable calculation and Lai.The increment and warp parameters of warp parameters are updated parameter with the mode of synthesis.Finally we have built the network frame of the handwritten Kanji recognition of the present invention on TensorFlow deep learning framework platforms.And compared with simply using convolutional neural networks Handwritten Chinese Character Recognition network frame, by the training of mass data collection, the result of test shows that handwritten Kanji recognition rate is significantly improved.

Description

Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks
Technical field
The invention belongs to Image Classfication Technology field, the Off-line Handwritten Chinese specifically based on depth convolutional neural networks is known Other method.
Background technology
Off-line Handwritten Chinese Recognition is a sub- direction in area of pattern recognition.Off line refers to handled handwriting It is the handwriting X-Y scheme collected by scanner or the first-class image-capturing apparatus of shooting, abbreviation handwritten Chinese character is known below Not.How the research paper and technical report delivered before 2011 selects feature and matching process to adapt to if mostly focusing on The change of handwritten Chinese character font.It is exactly the design in feature extraction algorithm and grader.Traditional HCCR steps include：Figure Normalization, feature extraction, dimensionality reduction, classifier training.And because Chinese character quantity is more, complicated, similar character is more and writing wind The problems such as lattice are changeable, not only step is complicated for these conventional methods, and if the feature chosen is uncomfortable in characteristic extraction step Conjunction will have a strong impact on recognition effect.Although the method based on MQDF and DLQDF has been achieved for good effect, at present Their bottleneck is reached.
In recent years because the image recognition based on deep learning achieves great breakthrough, depth convolutional neural networks (Deep Convolutional Neural Network, DCNN) has been also employed in handwritten Kanji recognition field. In 2013ICDAR international tournaments, Fujitsu team uses CNN models to be won the championship title with 94.77% discrimination.Occur afterwards Using improved model and the Handwritten Chinese Character Recognition Modal of preprocessing means.Most common improved method is to increase network The tolerance of deformation to different handwritten Chinese characters, wherein just include data enhancing technology and space pond, the two methods have with Lower shortcoming, the former, the method according to the sample new to the generation of initial data geometric deformation is using most in handwritten Kanji recognition A kind of more methods, if the learning ability of model that only sample exponentially level increases just is lifted, and individual do not allow The problem of ignorance.The latter, pond operation can damage image detail.And with the explosive growth depth convolutional Neural net of data Although network obtains very big success, also solves the Geometrical change of one species data without a kind of criterion.It is such as same Hand-written " Chinese " word, the size and the shape of stroke that different people writes out all are not quite similar.The conversion of data is various to be The key factor of influence depth convolutional neural networks recognition effect.
Improved method has：Pretreatment mode, twist distortion is carried out to Chinese character and carrys out EDS extended data set, improves convolutional Neural net The generalization ability of network；Feature extraction, with reference to traditional Gabor direction characters, HoG features etc., using different DCNN models.At present The models such as the ResLeNet of the GoogLeNet that the model through use is made up of Szegedy etc., K He, X Zhang et al. designs. These models are designed for picture classification.Still there is certain difference with handwritten Kanji recognition, although having reached good effect Still network structure is deeper for fruit, and model is complicated, and debugging is difficult.And the bad assurance of degree of twist distortion is carried out to Chinese character, no The adaptive selection angle of energy.2015, come from four Cambridge Phd of the new master AI company DeepMind under Google
Researcher devises spatial alternation network (Spatial Transformer Network), it is possible to achieve adaptive Rotation, translation, scaling.But the framework that reverse spatial alternation network and convolutional neural networks are formed directly is used for the hand-written Chinese Word identifies, also there is some problems.Such as after reverse spatial alternation network, although the direction of Chinese character is corrected, The Chinese-character stroke corrected is thicker.And sample is the sample translated by multilayer scaling, rotation in processing procedure, then directly It is input to convolutional neural networks identification.Because having cropped the marginal information of original picture by the sample of scaling, used in the hand-written Chinese In word identification font can be caused incomplete, have a strong impact on recognition effect.
Because current most of network have ignored the space distortion of actual handwritten Chinese character, because in actually writing environment, wind Lattice, direction, position, the difference of size cause sample set change various.However, CNN still lacks the space change to input sample The robustness of change.Traditional method for normalizing is only converted into sample the standardization Chinese character of prescribed level, although it appoints to classification Business plays the important and pivotal role, but method for normalizing cannot be guaranteed that HCCR tasks are optimal.And inverse composition space Converting network, according to the eigentransformation parameter of tasking learning picture, input can be schemed in the case of no mark key point Alignd on piece or the feature space of study, become so as to reduce the geometry such as the rotation due to space, translation, yardstick, distortion Change the influence to classifying and identifying.
The content of the invention
The characteristics of present invention is for the problem of existing above and Chinese character, using inverse composition spatial alternation algorithm and depth The Off-line Handwritten Chinese is identified degree convolutional neural networks scheduling algorithm.This method is to different writing styles and different books Writing environment has good robustness.Wherein we employ inverse composition spatial alternation network to solve because writing style is various The problems such as font brought is distorted, deformed, tilting.And devise corresponding depth convolutional neural networks.Depth convolution god Feature can be effectively extracted through network and is classified.The output of last inverse composition spatial alternation network is as convolutional Neural networking Input, realizes the identification for handwritten Chinese character.
Technical scheme and flow order are as follows：
(1) the TensorFlow deep learning framework platforms of depth convolutional neural networks are built；
(2) the GNT formatted datas of HWDB1.1 data sets are converted into binary system and are stored as PKL forms
(3) PKL formatted datas are read and are normalized, and are converted into training set, cross validation collection and survey Examination collection；
(4) inverse composition spatial alternation network and convolutional neural networks are realized in TensorFlow framework platforms, and will Input of the output of anti-phase blended space converting network as convolutional neural networks；
(5) setting direction blended space converting network frame model and it is added into the present inventor's self-designed convolution god Add the model structure of convolutional neural networks through in network frame, forming anti-phase blended space converting network；
(6) train and test the network frame built.
(7) using TensorBoard instruments visualization intermediate result, training and the result of test.The comparative analysis present invention The validity of network structure.
The detailed process of the step 2)：HWDB1.1 is the number of a Off-line Handwritten Chinese of CASIA-HWDB databases According to collection.It is that the Chinese Academy of Sciences collects between 2007-2010.The form of storage is exclusive binary system GNT forms.Therefore we need This data set is converted into available form first.It is current checking Model Identification because this data set writing font is various Effect is at most also one of most authoritative data set.
The step 3)：Picture is inverted and normalized, the former can make picture major part data be zero, accelerate fortune Speed is calculated, the latter can improve the convergence rate of network.
The step 4)：The detailed process of inverse composition spatial alternation network is：Training sample image and the torsion of initialization Bent parameter p is multiplied, and obtains the picture corrected for the first time.A geometry fallout predictor is entered into, geometry fallout predictor is by one Small-sized convolutional neural networks or neutral net are formed.The propagated forward prediction distortion increment Delta p of network, backpropagation BP algorithm To update the parameter of geometry fallout predictor, further iteration updates warp parameters.After often updating Δ p, updated with the mode of synthesis Parameter p.Be the picture corrected by matrix and the training sample multiplied result of P compositions afterwards, by through picture correction again this The geometry fallout predictor of input, such iteration renewal Δ p, p.Until meeting cycle-index given in advance, by last P values and original Training sample is multiplied, and feature is further extracted using obtained result as the input of the convolutional neural networks of the design of the present invention And classification.
The step 5)：Convolutional neural networks add a pond layer using the first two convolution, latter two convolutional layer, convolution It is two full articulamentums after layer.
With it is current based on the handwritten Chinese character of depth convolutional neural networks compared with, 1) present invention using inverse composition space become Switching network can by handwritten Chinese character collection by twist distortion and be converted into readily identified Chinese character picture effectively raise it is hand-written The recognition effect of the Chinese character and artificial design parameter that avoids operates to twist distortion of picture etc..2) convolutional neural networks are adopted A pond layer is added to add the network structures of two convolutional layers with two convolutional layers.Using more convolutional layer and small convolution kernel come Effective extraction Chinese-character stroke feature.
Brief description of the drawings
Fig. 1 inverse composition spatial alternation flow through a network figures
Fig. 2 CNN_4 network structures
Fig. 3 schemes one are by comparison diagram before and after reverse spatial alternation network
(a) former handwritten Chinese character
(b) handwritten Chinese character corrected by ICSTN_2
Fig. 4 schemes two are by comparison diagram before and after reverse spatial alternation network
(a) former handwritten Chinese character
(b) handwritten Chinese character corrected by ICSTN_2
Fig. 5 schemes three are by comparison diagram before and after reverse spatial alternation network
(a) former handwritten Chinese character
(b) handwritten Chinese character corrected by ICSTN_4
The cost function of Fig. 6 schemes one and test accuracy rate change curve
(a) ICSTN_1+CNN_4 loss functions are used
(b) ICSTN_1+CNN_4 test accuracy rates are used
The cost function of Fig. 7 schemes two and test accuracy rate change curve
(a) ICSTN_2+CNN_4 loss functions are used
(b) ICSTN_2+CNN_4 test accuracy rates are used
The cost function of Fig. 8 schemes three and test accuracy rate change curve
(a) ICSTN_4+CNN_4 loss functions
(b) ICSTN_4+CNN_4 test accuracy rates are used
Fig. 9 CNN_4 cost functions and test accuracy rate change curve
(a) CNN_4 loss functions are used
(b) CNN_4 test accuracy rates
Specific embodiment
Detailed description further to the present invention below in conjunction with the accompanying drawings
The specific configuration method of network given below, and the algorithm principle and its effect of the new theory of the invention used. To the implementation further instruction of the present invention.Example of the present invention is the Chinese character of HWDB1.1 200 species, each 300 Chinese characters of classification, totally 60000 samples, wherein training set, cross validation collection and test machine are respectively：48000th, 6000, 6000.The idiographic flow of the network operation designed by the present invention is：
(1) data prediction, the binary picture of acquisition is inverted, normalized, to improve the convergence of network.And Picture size is uniformly adjusted to 32*32, is then assigned as training set, cross validation collection and test set；
(2) training sample is input to next step deformation parameter, is designated as I (x), x=(x, y) represents any pixel in image Point must be sat.P=[p1 p2 … pn] represent warp parameters.The present invention is using affine transformation as p=[p1 p2 p3 p4 p5 p6], Related secondly coordinates table transformation matrix is：
(3) W (p) enters line translation to original image：
ImWarp (x)=I (x) W (p) (2)
(4) because by Skewed transformation image coordinate be it is non-it is whole cause image discontinuous, the present invention will be by conversion Image ImWarp (x) carries out bilinear interpolation：
Vi c：Represent the ith pixel point of c-th of sample；
(xi,yi)：Represent training sample ImWarp ith pixel point coordinates；
W：Represent the width of image；
H：Represent the height of image；
(5) image V (x) is input in convolutional neural networks or neutral net, this network of the invention is referred to as that geometry is pre- Survey device (why derivation that below will be detailed learns to distort increment with neutral net or convolutional neural networks), geometry fallout predictor For learning the distortion increment Delta p of image, the composition of its network structure shares three kinds of schemes.The last present invention will emulate one by one Realize the effect of each scheme.Three kinds of schemes are respectively：
A) scheme one：The structure of network is:Conv (7 × 7,4)+Conv (7 × 7,8)+P+FC (48)+FC (8), wherein with Conv represents convolutional layer, and P represents pond layer, and FC represents full articulamentum, the size and number of the interior respectively convolution kernel of bracket, Quan Lian It is neuron number to connect in layer bracket；
B) scheme two：The structure of network is:Conv(9×9,8)+FC(48)+FC(8)；
C) scheme three：The structure of network is:FC(48)+FC(8)
The propagated forward output distortion increment Delta p of geometry fallout predictor, backpropagation renewal network parameter, is further produced new Δ p.
(6) warp parameters are updated, warp parameters p update mode is synthesis.Under specific formula：
Corresponding transformation matrix is：
W(pout)=W (Δ P) W (pin) (5)
(7) if cycle-index recurN ＞ 1, step (3) is gone to until circulation terminates.Inverse composition spatial alternation network Can repeatedly it circulate to obtain alignment and the best effect of warp image.
(8) after circulation terminates, by the transformation matrix W (p of the warp parameters p compositions of final updatedout) act on original image j I.e.：
Im=I (x) W (Pout) (6)
, input an image into afterwards in convolutional neural networks；
(9) framework of convolutional neural networks such as Fig. 2, network structure are Conv (3 × 3,8)+Conv (3 × 3,16)+P+ Conv (3 × 3,32)+Conv (3 × 3,64)+P+FC (100)+FC (200), each parameter specifically represent what with it is above-mentioned It is identical.Next we carry out example explanation with one pond layer of a convolutional layer and a full articulamentum.
Convolutional layer, lamination to the effect that carry out convolution operation using trainable convolution kernel to input data, and By result, form exports in some combination, and it is substantially exactly the feature extraction to input data.Generally, in order that mould Type obtains nonlinear characteristic and export-restriction in given scope, can be by the output of convolution with a nonlinear function (Non-linear Function) is changed.This function is also referred to as activation primitive (Activation Function).This reality Example is using the faster Relu activation primitives of convergence rate.The concrete operation formula of convolutional layer is:
Wherein Im is the output of reverse spatial alternation network, is exactly the distortion Chinese character corrected by reverse spatial alternation network Training sample, W is convolution kernel, and b is convolutional layer biasing；U exports for convolutional layer, and f is Relu activation primitives, and Y is that convolution is defeated Go out the output that U passes through activation primitive, be the output of whole convolutional layer, and next layer of input.
Pond layer, main task are to carry out sampling operation, also referred to as down-sampling to the data sample of input in two-dimensional space Operation, specific calculating process are as follows：
Pond layer can make network have robustness to translation, rotation etc..Network is set to have more preferable tolerance.
Full articulamentum, the effect of full articulamentum are to be easy to the two dimensional character matrix dimensionality reduction of input to one-dimensional characteristic vector Output layer carries out classification processing.
Output layer, the effect of output layer are classified according to the one-dimensional vector of the output of full articulamentum above, Ben Liecai It is that softmax intersects entropy loss
The distortion offset Δ p of image can be asked for being because for inverse composition KL images by convolutional neural networks model Alignment algorithm：
The object function of formula (9) inverse composition image alignment, wherein I are original images, and T is target image, p：Warp parameters, Δp：The distortion increment to be asked for；(Δ p) represents the image by warp parameters effect respectively by I (p) and T.Then, by above formula (9) obtained by single order Taylor is approximate:
Its least square solution is：
Wherein+symbology generalized inverse.Next parameter p update mode is：
Wherein, it is warp parameters p synthesis computing, the solution of formula (10) is exactly a linear regression forms, can be converted For general expression：
Δ p=RI (p)+b (13)
Wherein R is the linear regression for estimating the linear relationship between image and the geometry of image, and b is biasing.Mainly It is to ask for parameter R and b.Linear regression model (LRM) we its parameter can be turned to neutral net or convolutional neural networks, network Output be Δ p.
Inverse composition spatial alternation network be not by the geometric transformation of the direct prognostic chart picture of network, it is but more before classification Secondary recycling geometry fallout predictor renewal warp parameters p, and after each circulation, warp parameters p is obtained after output image To preserve, this will to be avoided edge effect in resampling original image.Geometry prediction is sub { R, b }, is declined with supervision gradient Optimization and the object function of study, its object function are：
Δp:Increment is distorted, is the output of geometry fallout predictor
M：The sample number of image
By formula, (12) and (13) can obtain：
Using composition algorithm undated parameter, the first-order partial derivative without more new images every time to parameter, in backpropagation Ask for during gradient is, the partial derivatives of warp parameters is a closed-form mathematical expression formula, Pin, PoutThe respectively distortion ginseng of input and output Number：
I is unit matrix, and it allows gradient to propagate backward to geometry fallout predictor.
In order to intuitively illustrate the feasibility and validity of the network frame of our constructions, the present invention is entered by experiment simulation Checking is gone.And three schemes of the network of the geometry fallout predictor of our designs are demonstrated one by one.The frame that each scheme is formed Frame uses identical parameter configuration in training：Epoch=16, initial learning rate are 0.001, and each two epoch learning rates multiply With 0.8 factor.In order to contrast the validity using anti-phase blended space, the independent the Realization of Simulation only convolutional Neural of the present invention The recognition effect of network.
The present invention has been visualized by the training sample before and after reverse spatial alternation network first.Scheme one, two, Model corresponding to three is respectively：ICSTN_1+CNN_4, ICSTN_2+CNN_4 and ICSTN_4+CNN_4.After wherein ICSTN_ The numeral in face is model depth.Contrasted before and after being converted such as the respectively three different geometry fallout predictor models of Fig. 3,4,5 to sample Figure.From entirety, after by inverse composition spatial alternation network, each Chinese character size is identical, stroke is uniform, arrangement is whole Together, the stroke of some distortions is corrected, higher much than former Chinese character identification.Difference between different models is also quite notable, Overall trend is that the network of geometry fallout predictor is deeper, more clear by the Chinese-character stroke of spatial alternation, and rotates excessive feelings Condition is less, and more distortion strokes is corrected.Chinese character " towel ", " mortar ", " most " word obtain appropriate correction in such as Fig. 5 (b) Deformation.And Fig. 3 effects are worst, there is more Chinese character to generate inclination.Can to sum up draw, inverse composition spatial alternation network it is several The network of what fallout predictor is more deep more can effectively correct the Chinese character of twist distortion.
Secondly the iterativecurve figure of each network frame of analysis：Fig. 6,7,8 are respectively that three schemes form three models Loss function curve and test accuracy rate curve.The iterations of ordinate is that is, every 50 training iteration in units of 50 As a result show once, so totally 50000 iteration, 10 epoch.From ICSTN_2+CNN_4 from the point of view of (a) figure of three figures and The convergence of ICSTN_4+CNN_4 models is very fast, and model ICSTN_4+CNN_4 model stability is best because the cusp of curve compared with It is few.But the loss of ICSTN_2+CNN_4 models is closest to zero point.ICSTN_2+ can be seen to obtain from (b) figure in three figures CNN_4 discrimination highest.The effect of that model correction Chinese character why scheme three is formed is best, does not obtain but best DiscriminationReason is mainly that the parameter of ICSTN_4+CNN_4 models is more is not instructed sufficiently under identical epoch Practice.CNN models and scheme one, the model of two, three compositions is larger compared to gap, (a) (b) and (a) (b) of Fig. 6,7,8 in Fig. 9 Figure value convergent compared to loss function is maximum, and discrimination only has 82.08%.CNN_4 recognition effect is bad be because model compared with It is shallow.
It can finally be obtained from table 1, from the time of training pattern consumption, reverse spatial alternation network is deeper consumed Time is more.So the factor such as overall efficiency and discrimination, ICSTN_2+CNN_4 effect and material rate is best.From test From the point of view of accuracy rate.The present invention has obvious advantage.And CNN_4 models and ICSTN_1+CNN_4 model training time phase differences are smaller But recognition effect gap is obvious.ICSTN_1 use only one one layer of neural network.So the present invention realizes reverse sky Between converting network and shallower convolutional Neural networking be combined the effect for but having obtained a higher identification.
The recognition effect of the heterogeneous networks structure of table 1
Model classification Recognition accuracy (%) Time-consuming (h)
CNN_4 82.08 5.16
ICSTN_4+CNN_4 95.83 17.72
ICSTN_2+CNN_4 96.38 13.40
ICSTN_1+CNN_4 95.32 6.92

Claims (2)

  1. A kind of 1. method of the Off-line Handwritten Chinese Recognition based on depth convolutional neural networks, it is characterised in that including following step Suddenly：
    TensorFlow deep learning frameworks are built on Windows；
    Prepare data set, data set is converted into TensorFlow input form, and pretreatment is normalized in picture；
    Inverse composition spatial alternation network and the convolutional neural networks of depth are built under TensorFlow environment, using reverse Blended space network carries out twist distortion to handwritten Chinese character, is corrected and alignd.And output it as convolutional neural networks Input；
    Finally the network of inverse composition spatial alternation network and convolutional neural networks formation is trained using substantial amounts of data, Test.And the picture during network processes and final testing result are visualized by TensorBoard instruments, analyze its work With, and contrast recognition effect.
  2. 2. according to step described in claims 1, handwritten Chinese character is corrected using inverse composition spatial alternation, its feature exists In：
    The homogeneous matrix formed using warp parameters enters line translation alignment, exemplified by warp parameters are in a manner of affine, ginseng to original image Number p=[p1 p2 p3 p4 p5 p6], corresponding homogeneous transform matrix：
    <mrow> <mi>W</mi> <mrow> <mo>(</mo> <mi>p</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <mn>1</mn> <mo>+</mo> <msub> <mi>p</mi> <mn>1</mn> </msub> </mrow> </mtd> <mtd> <msub> <mi>p</mi> <mn>2</mn> </msub> </mtd> <mtd> <msub> <mi>p</mi> <mn>3</mn> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>p</mi> <mn>4</mn> </msub> </mtd> <mtd> <mrow> <mn>1</mn> <mo>+</mo> <msub> <mi>p</mi> <mn>1</mn> </msub> </mrow> </mtd> <mtd> <msub> <mi>p</mi> <mn>6</mn> </msub> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>1</mn> </mtd> </mtr> </mtable> </mfenced> </mrow>
    Acted on image, twist distortion is carried out to image I (x), include the operation such as the rotation of image, translation, scaling.
    ImWarp (x)=I (x) W (p)
    ImWarp (x) is the picture corrected.Parameter p is to be updated to ask for by the way of iteration, and p distortion increases Amount Δ p is that this network is referred to as into geometry fallout predictor by a convolutional neural networks or neural computing, the present invention.By Inspired in inverse composition spatial alternation principle by inverse composition Lucas-Kanade algorithms：
    <mrow> <munder> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> <mrow> <mi>&amp;Delta;</mi> <mi>p</mi> </mrow> </munder> <mo>|</mo> <mo>|</mo> <mi>I</mi> <mrow> <mo>(</mo> <mi>P</mi> <mo>)</mo> </mrow> <mo>-</mo> <mi>T</mi> <mrow> <mo>(</mo> <mi>&amp;Delta;</mi> <mi>P</mi> <mo>)</mo> </mrow> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> </mrow>
    Shift onto and convert by a series of, its solution is：
    Δ p=RI (p)+b
    Last solution is linear regression forms, it is possible to which its solution procedure parameter is turned into neutral net or convolutional Neural net Network；
    Seeking the network structure of distortion increment Delta p geometry fallout predictor has three kinds of schemes, and two schemes are convolutional neural networks structures, One neutral net for individual layer.There is detailed structure explanation in specification.We eventually verify each scheme and divided one by one Analyse its validity.
    The Δ p of geometry fallout predictor renewal synthesizes with p before, undated parameter p.The p value of renewal can be existed from new role afterwards On image, loop iteration asks for optimal p value.Defined according to synthesis, update mode is as follows：
    By the way of synthesis it is the gradient that need not remove to ask sample image every time come the benefit of undated parameter.Greatly accelerate anti- To the calculating speed of propagation.
    It will be finally input to by the picture corrected in convolutional neural networks and carry out feature extraction and classifying.Convolutional neural networks Structural order is respectively：Two convolutional layers, then a pond layer, two convolutional layers, a pond, it is finally two full connections Layer.Final output neuron number is identical with the classification of present example.
CN201710855035.7A 2017-09-20 2017-09-20 Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks Pending CN107590497A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710855035.7A CN107590497A (en) 2017-09-20 2017-09-20 Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710855035.7A CN107590497A (en) 2017-09-20 2017-09-20 Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks

Publications (1)

Publication Number Publication Date
CN107590497A true CN107590497A (en) 2018-01-16

Family

ID=61048708

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710855035.7A Pending CN107590497A (en) 2017-09-20 2017-09-20 Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks

Country Status (1)

Country Link
CN (1) CN107590497A (en)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108509881A (en) * 2018-03-22 2018-09-07 五邑大学 A kind of the Off-line Handwritten Chinese text recognition method of no cutting
CN108734168A (en) * 2018-05-18 2018-11-02 天津科技大学 A kind of recognition methods of handwritten numeral
CN108805196A (en) * 2018-06-05 2018-11-13 西安交通大学 Auto-increment learning method for image recognition
CN108932500A (en) * 2018-07-09 2018-12-04 广州智能装备研究院有限公司 A kind of dynamic gesture identification method and system based on deep neural network
CN108960301A (en) * 2018-06-20 2018-12-07 西南大学 A kind of ancient Yi nationality's text recognition methods based on convolutional neural networks
WO2019232874A1 (en) * 2018-06-04 2019-12-12 平安科技（深圳）有限公司 Chinese character model training method, chinese character recognition method, apparatus, device, and medium
CN111553423A (en) * 2020-04-29 2020-08-18 河北地质大学 Handwriting recognition method based on deep convolutional neural network image processing technology
CN111598089A (en) * 2020-05-16 2020-08-28 湖南大学 License plate correction and recognition method based on deep learning
CN112434574A (en) * 2020-11-11 2021-03-02 西安理工大学 Knuckle print identification method under non-limited state
CN112507864A (en) * 2020-12-04 2021-03-16 河北地质大学 Credit archive identification method based on convolutional neural network
WO2021091226A1 (en) * 2019-11-04 2021-05-14 Samsung Electronics Co., Ltd. Method and electronic device for correcting handwriting input
TWI823488B (en) * 2022-07-22 2023-11-21 國立中央大學 Method for implementing edge-optimized incremental learning for deep neural network and computer system

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106408039A (en) * 2016-09-14 2017-02-15 华南理工大学 Off-line handwritten Chinese character recognition method carrying out data expansion based on deformation method
CN106408038A (en) * 2016-09-09 2017-02-15 华南理工大学 Rotary Chinese character identifying method based on convolution neural network model

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106408038A (en) * 2016-09-09 2017-02-15 华南理工大学 Rotary Chinese character identifying method based on convolution neural network model
CN106408039A (en) * 2016-09-14 2017-02-15 华南理工大学 Off-line handwritten Chinese character recognition method carrying out data expansion based on deformation method

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
CHEN-HSUAN LIN ET AL.: "Inverse Compositional Spatial Transformer Networks", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 *
ZHAO ZHONG ET AL.: "Handwritten Chinese Character Recognition with Spatial Transformer and Deep Residual Networks", 《2016 23RD INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR)》 *

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108509881A (en) * 2018-03-22 2018-09-07 五邑大学 A kind of the Off-line Handwritten Chinese text recognition method of no cutting
CN108734168A (en) * 2018-05-18 2018-11-02 天津科技大学 A kind of recognition methods of handwritten numeral
WO2019232874A1 (en) * 2018-06-04 2019-12-12 平安科技（深圳）有限公司 Chinese character model training method, chinese character recognition method, apparatus, device, and medium
CN108805196A (en) * 2018-06-05 2018-11-13 西安交通大学 Auto-increment learning method for image recognition
CN108960301A (en) * 2018-06-20 2018-12-07 西南大学 A kind of ancient Yi nationality's text recognition methods based on convolutional neural networks
CN108932500A (en) * 2018-07-09 2018-12-04 广州智能装备研究院有限公司 A kind of dynamic gesture identification method and system based on deep neural network
WO2021091226A1 (en) * 2019-11-04 2021-05-14 Samsung Electronics Co., Ltd. Method and electronic device for correcting handwriting input
US11450041B2 (en) 2019-11-04 2022-09-20 Samsung Electronics Co., Ltd. Method and electronic device for correcting handwriting input
CN111553423A (en) * 2020-04-29 2020-08-18 河北地质大学 Handwriting recognition method based on deep convolutional neural network image processing technology
CN111598089A (en) * 2020-05-16 2020-08-28 湖南大学 License plate correction and recognition method based on deep learning
CN112434574A (en) * 2020-11-11 2021-03-02 西安理工大学 Knuckle print identification method under non-limited state
CN112434574B (en) * 2020-11-11 2024-04-09 西安理工大学 Knuckle pattern recognition method under unrestricted state
CN112507864A (en) * 2020-12-04 2021-03-16 河北地质大学 Credit archive identification method based on convolutional neural network
TWI823488B (en) * 2022-07-22 2023-11-21 國立中央大學 Method for implementing edge-optimized incremental learning for deep neural network and computer system

Similar Documents

Publication Publication Date Title
CN107590497A (en) Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks
US11556797B2 (en) Systems and methods for polygon object annotation and a method of training an object annotation system
CN104463209B (en) Method for recognizing digital code on PCB based on BP neural network
Saragih et al. Learning AAM fitting through simulation
Zhu et al. Bilateral weighted regression ranking model with spatial-temporal correlation filter for visual tracking
CN107016319A (en) A kind of key point localization method and device
CN111553423A (en) Handwriting recognition method based on deep convolutional neural network image processing technology
CN110674777A (en) Optical character recognition method in patent text scene
Guo et al. Creating New Chinese Fonts based on Manifold Learning and Adversarial Networks.
CN110555864B (en) Self-adaptive target tracking method based on PSPCE
CN108647726A (en) A kind of image clustering method
Suzuki et al. Adversarial transformations for semi-supervised learning
Wei et al. Non-homogeneous haze removal via artificial scene prior and bidimensional graph reasoning
Alonso-Weber et al. Combining additive input noise annealing and pattern transformations for improved handwritten character recognition
Hu et al. Face reenactment via generative landmark guidance
Xu et al. GDFace: Gated deformation for multi-view face image synthesis
CN109214269B (en) Human face posture alignment method based on manifold alignment and multi-image embedding
CN106778579A (en) A kind of head pose estimation method based on accumulative attribute
Ren et al. SAFont: Automatic Font Synthesis using Self-Attention Mechanisms.
Jiang et al. A super-resolution method for low-quality face image through RBF-PLS regression and neighbor embedding
CN111680760A (en) Clothing style identification method and device, electronic equipment and storage medium
Pavlidis et al. An on-line handwritten note recognition method using shape metamorphosis
Deokate et al. Cnn classification approach for analysis and recognition of marathi manuscript
Ahmadian et al. Model-based Persian calligraphy synthesis via learning to transfer templates to personal styles
CN117496537B (en) Handwriting writing quality evaluation method based on improved shape feature matching

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20180116

WD01 Invention patent application deemed withdrawn after publication