CN106408038A - Rotary Chinese character identifying method based on convolution neural network model - Google Patents
Rotary Chinese character identifying method based on convolution neural network model Download PDFInfo
- Publication number
- CN106408038A CN106408038A CN201610813866.3A CN201610813866A CN106408038A CN 106408038 A CN106408038 A CN 106408038A CN 201610813866 A CN201610813866 A CN 201610813866A CN 106408038 A CN106408038 A CN 106408038A
- Authority
- CN
- China
- Prior art keywords
- chinese character
- rotation
- training
- sample
- data set
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/24—Character recognition characterised by the processing or recognition method
- G06V30/242—Division of the character sequences into groups prior to recognition; Selection of dictionaries
- G06V30/244—Division of the character sequences into groups prior to recognition; Selection of dictionaries using graphical properties, e.g. alphabet type or font
- G06V30/2455—Discrimination between machine-print, hand-print and cursive writing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Character Discrimination (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a hand-written Chinese character rotation-independent identifying method based on a convolution neural network. The method includes the steps of building a Caffe deep learning framework platform containing a plurality of convolution neural network models on a graphics processor, preparing a training data set and a test data set with labels, training the convolution neural network models to identify primary-level hand-written Chinese characters on the graphics processor by using the data sets, and inputting original images of hand-written Chinese characters in HCL2000 database and images rotated randomly in various directions into the convolution neural network models to train the network, and finally inputting unknown rotated Chinese characters for testing to obtain the identification result of the Chinese character images. The method has the advantages of high smart level, simple method, accurate classification and rapid detection speed. The method has good identification performance for low-magnitude databases and has excellent identification performance for hand-written Chinese characters rotated by any angle.
Description
Technical field
The invention belongs to Image Classfication Technology field, particularly one kind are based on convolutional neural networks（CNN）The rotation of model
Chinese characters recognition method.
Background technology
Off-line Handwritten Chinese Recognition is always one of difficult point of area of pattern recognition.How to strengthen the Off-line Handwritten Chinese with
Machine rotation individual character recognition capability has very strong realistic meaning.In daily life, the reason such as the characteristic of sensor and limitation is normal
Lead to the input of a PRS not ideal enough, the input to off line Chinese character recognition system often natural rotation, this
The recognition capability that identifying system will be led to declines；For the off line Chinese character of big angle rotary, almost it is difficult to.Offline handwriting
The Rotation of Chinese character, does not have good solution so far.The present invention is directed to GB2312 80 one-level character set, to selection
The similar handwritten Chinese character image of all off lines carry out random angles rotation processing after be identified again.Rotation the Off-line Handwritten Chinese
The main difficulty of identification is that stroke order is unknown, angle is difficult to determine, the presence of particularly a large amount of similar Chinese character and not advising
Then write deformation.After rotation, off line Chinese character is more difficult to identify.Therefore, the recognition performance improving rotation Chinese character has very
Strong realistic meaning.
In recent years, for solving the problems, such as off line rotation character recognition, there has been proposed many effectively methods：U.Pal et al.
In order to solve the problems, such as the identification of the artistic english character of multi-faceted multiple dimensioned printing it is proposed that a kind of returning based on character boundary point
Quadric discriminant function (MQDF) recognition methods of the correction of one change information, can effectively identify to the English character of rotation.C.
Singh et al. proposes and proposes to change Ze Nike square（The magnitude of Zernike moments）Single real component
With the new method of imaginary component, the real component after change and imaginary component be subsequently used as constant image descriptor, and this method is to rotation
Turning character recognition has good recognition performance.But above method has high demands to data set sample number magnitude, need substantial amounts of number
According to training network, for low order of magnitude sample not effect well.And method is complicated, different pieces of information collection effect is had
Certain difference, does not have universality well.Therefore, for the classification of the Chinese character of Random-Rotation, there is presently no more pervasive has
The method of effect.
Content of the invention
In order to solve the above problems, it is an object of the invention to provide a kind of rotation Chinese based on convolutional neural networks model
Word recognition methods, the fast sorting technique based on convolutional neural networks of classification speed.
The present invention adopts the following technical scheme that realization.
A kind of rotation Chinese characters recognition method based on convolutional neural networks model, it includes the following step carrying out in order
Suddenly：
1）The caffe deep learning framework platform of convolutional neural networks is built on Linux system；
2）Prepare data set（First-level Chinese characters under HCL2000 data set）：Training dataset and the test data set with label.
Training set is original handwritten Chinese character and the sample obtaining original Chinese character Arbitrary Rotation adds in training set jointly, test
Collection is the handwritten Chinese character of the Arbitrary Rotation with label.
3）Using above-mentioned data set on caffe platform training convolutional neural networks model, obtain test result, it is right to realize
The identification of Arbitrary Rotation Chinese character；
Further, in step 2）In, training dataset is gained after HCL2000 one-level handwritten Chinese character Data Centralized Processing,
Processing method is to concentrate Chinese character to rotate data, thus in the training process, it is special that model may learn more rotations
Levy；Test set is to concentrate the arbitrarily angled Random-Rotation of Chinese character to data, and carries label.
Compared with prior art, what the present invention provided is had based on the rotation Chinese character sorting technique of convolutional neural networks model
Advantage and good effect be：(1) adopt the outstanding multilayer convolutional neural networks model of current classifying quality, classification is accurately
Rate is high；(2) by Arbitrary Rotation is carried out to Chinese character image, sample set is extended, the model obtaining can learn
To the feature of rotated sample, to improve the robustness to rotation Character Font Recognition for the convolutional neural networks model；(3) experiment is based on GPU
Parallel computation, training and test speed be significantly larger than CPU arithmetic speed.
Brief description
Fig. 1 is convolutional neural networks topological structure in example.
Fig. 2 is rotation transformation geometrical relationship figure in example.
Fig. 3 is Random-Rotation angle in exampleUnder rotation image.
Fig. 4 is for non-rotated sample in example and the common training pattern of rotated sample to rotated sample recognition capability curve.
Specific embodiment
Below in conjunction with accompanying drawing and example to the present invention be embodied as be described further, but the enforcement of the present invention and protection
Not limited to this, if in place of having special detailed description in detail below, are all that those skilled in the art can refer to prior art realization.
As Fig. 1, convolutional neural networks are made up of input layer, hidden layer and output layer, hidden layer then mainly include convolutional layer,
Maximize pond sample level and full articulamentum composition.
（1）Convolutional layer.Convolutional layer is used for extracting basic visual signature, also referred to as Feature Mapping in vision acceptance region, behaviour
It is also referred to as neuron as unit.
（2）Maximum sample level.Because image has the attribute of a kind of " nature static ", this also implies that in an image district
Domain useful feature is very likely equally applicable in another region.Therefore, in order to describe big image, can be from diverse location
Feature carry out aggregate statistics, statistical nature not only can reduce dimension, also can improve result it is not easy to over-fitting simultaneously.Adopt
With maximum pond, take the input as next layer for the maximum in pond region.
（3）Full articulamentum.After convolutional layer and pond layer, the feature extracting is combined again, finally
The unique feature having to each word.
In the initialized selection of weight parameter, using Xavier strategy, this is to be proposed in 2010 by Xavier et al.
A kind of normalization initialization strategy, this strategy makes network can keep activating difference and dorsad gradient side in the training process
Difference, makes network convergence obtain very fast.This strategy is by formulaBe given：
Here U refers to be uniformly distributed,It is respectively the quantity of the neuron of current layer and next layer.
Additionally, in network training parameter adjustment, for improving the robustness of network and accelerating network convergence, setting with off line
Network parameter：
（1）Learning rate.Increase with iterations is gradually reduced by learning rate.Using following more New Policy：
Wherein,Based on learning rate,、For the parameter setting, iter is iterations, takes in experiment,,.
（2）Neutral net activation primitive.Using RELUNonlinear activation function replaces conventional sigmoid function.
（3）Error function increased momentum term and regularization term.Momentum term is based on the Newton's law in physics, when by mistake
After difference curved surface enters " flat region ", network can quickly be restrained.Regular terms is then to introduce in order to avoid network over-fitting
A regularization coefficient, also referred to as weight attenuation coefficient.
Last layer of network, i.e. full articulamentum, its design is to be associated with the classification task of network.The god of output layer
It is set as the class number of required classification through first quantity.In this example, employ softmax grader, cross entropy error by
FormulaBe given.
ForThe classification task of individual classification, orderForThe cross entropy error value of individual sample, and discriminant classification rule is by formulaBe given, whereinIt is a constant independent of classification.ForIf having, thenFor minimum of a value,
SoIt is divided intoClass.
After having built network, next rotation transformation is carried out to sample.Rotation transformation is the operation side of affine transformation
One of formula, by carrying out rotation transformation to original sample, and is added among training set, so that model learning is to rotation
Sample characteristics.Therefore, rotation processing is carried out to original sample, and rotated sample is added to training pattern in training set.
Geometrically, an affine transformation between two vector spaces can be by a linear transformation and a translation group
Become.For two-dimensional space, affine transformation can be in the form of with matrixRepresent：
Wherein size isMatrixAnd column vectorIt is all the coefficient of affine transformation,WithRepresent former two respectively
The base vector in dimensional linear space and the base vector in the two-dimensional linear space after affine transformation.Sample is obtained newly as affine transformation
Sample can pass through formulaEach pixel that conversion coefficient is applied on image to be realized.This interactively by
FormulaBe given.
WhereinAfter representing conversion, image coordinate isPixel gray value,Represent original image
Coordinate isPixel gray value (due toWithValue be likely to beyond given image and be non-integer, because
This suppose beyond given image size pixel gray value be 255 and in an experiment application bilinear interpolation method).
Common affine transformation has translation, scaling, rotation, Shear Transform (in horizontal direction and on vertical direction) etc..This
Example employs the method that sample is made to rotate to extend sample set.
Common affine transformation has translation, scaling, rotation, Shear Transform (in horizontal direction and on vertical direction) etc., and
Rotation is one of important affine transformation operation.As shown in Fig. 2 with pointFor axle center rotate counterclockwiseIt is assumed that pointFormer two
Coordinate in dimension space is, then can be obtained a little by geometrical relationshipCoordinate in new two-dimensional space's
Value is respectively：
Therefore, rotation transformation coefficient is respectively：
,
Each of image pixel is all passed throughWithTo calculate new coordinate value, then according to formula
Calculate the new gray value of each pixel, you can obtain rotating image.ChangeIn parameterCan change with image
The heart is the angle of axle center image rotate counterclockwise.Several Random-Rotation angles are illustrated in Fig. 3Under rotation image（Top is
Original image, lower section is corresponding random angles rotation image）.
Finally inventive method is tested, this example employs the handwritten Chinese character sample database of HCL2000,
HCL2000 is the large-scale Off-line Handwritten Chinese Character Recognition Sample Storehouse that Beijing University of Post ＆ Telecommunication issues.Comprise in database
3755 conventional simplified Chinese characters, respectively by 1000 different person writings.For the recognition capability of the Off-line Handwritten Chinese, pass through
N candidate's diversity method of most possible property, has had higher recognition capability to the Off-line Handwritten Chinese.But for similar character,
Because stroke is close, font architecture is similar, is still a major challenge in Chinese Character Recognition field at present.Therefore, ten groups of phases are randomly selected
Like Chinese character, every group of Chinese character includes ten similar fonts, each font totally 300 original samples.For each font, choose wherein
, as training set, 75 as test set for 225 samples.In order to reduce dimension and reduce training burden, using arest neighbors interpolation
The sample image size of 64*64 is adjusted to 28*28 by algorithm.In addition, in order to avoid leading to because font is in edge
During network extraction feature may lost part feature, image surrounding after being sized adds 2 blank pixel（Pixel ash
Angle value is 255）, the Hanzi specimen image size finally giving is 32*32.
By the convolutional neural networks training pattern built, adjust network parameter, finally give the result shown in table 1.
Table 1 original sample recognition capability experimental result
By experiment test, the recognition accuracy mean value of ten groups of similar Chinese characters has reached 92.26%, and every group of Chinese character
Recognition accuracy all very high it was demonstrated that the reasonability of network and validity.But the recognition accuracy that there is part sample is not high
Situation, by checking error sample, find that the accuracy rate of Chinese Character Recognition is affected by the similitude of Chinese character itself.For example
" scholar " and " native ", " big " " too " and " fiery " etc., due to no constraining the impact writing deformation in handwritten Chinese character, it will lead to these
Similar character is difficult to correctly identify.After Chinese character spinning, mistake may still exist.In addition " scholar " and " doing " this kind of Chinese character may
Due to the rotation of certain wordAfterwards, become another font, and lead to identify mistake.
Using above-mentioned original sample recognition capability Experiment Training model out to any rotation（）Ten groups of similar Chinese characters carry out class test respectively, test result such as table 2.
The original sample training of table 2 draws the result to the direct test of rotated sample for the model
| Group | 01 | 02 | 03 | 04 | 05 | 06 | 07 | 08 | 09 | 10 | Mean value |
| Accuracy rate | 0.190 | 0.159 | 0.242 | 0.163 | 0.150 | 0.171 | 0.193 | 0.196 | 0.217 | 0.167 | 0.1848 |
The average recognition capability of the similar Chinese character that can rotate that the test accuracy rate that this is tested is averaged only has 18.48%, completely not
Enough to apply.Therefore, using original sample training, model out cannot accurately identify rotation Chinese character.
First pass through and original sample is fixed with angle rotation, with the geometric center of font as pivot, oftenProduce a rotated sample.For single sample, in the anglec of rotationIn the range of can obtain
To 36 rotated samples.Then the new rotated sample producing is added in training set, training network model.Finally to random angle
Degree rotated sample is tested.Then, oftenProduce a rotated sample, identical with foregoing description, after training
Network model is tested to Random-Rotation sample.Ten similar characters taking 01 group are tested, and obtain result shown in table 3.
Table 3 often rotatesThe sample obtaining adds the recognition capability to Random-Rotation sample for the training set
Experiment obtains rotated sample by carrying out Random-Rotation to original sample, postrotational sample is added training set, and changes
The ratio of change rotated sample and original sample is as comparison.Will after rotation the sample of different proportion and original sample collectively as
Training set carrys out training network.Ten similar characters taking 01 group are tested, and obtain as shown in table 4.
The non-rotated sample of table 4 and the common training pattern of rotated sample are to rotated sample recognition capability
As Fig. 4, test result indicate that, with the increase of added rotated sample, rotate font recognition capability and lifted, simultaneously when
After rotated sample ratio reaches 30 times of normal samples, accuracy rate will stabilise near 93%, close to not adding rotation sample
The degree of accuracy that the network that the training of this when obtains is obtained（94.2%）.Random-Rotation sample and original sample in training set
Ratio is 50：When 1, reach best effects（0.939）.Below, to the experiment carrying out same ratio in remaining nine groups, obtain as follows
Similar result：
Add in table 5 training set during 50 times of Random-Rotation samples to Random-Rotation Character Font Recognition result
| Group | 01 | 02 | 03 | 04 | 05 | 06 | 07 | 08 | 09 | 10 | Mean value |
| Accuracy rate | 0.939 | 0.942 | 0.858 | 0.953 | 0.864 | 0.949 | 0.920 | 0.913 | 0.901 | 0.947 | 0.9186 |
The results show, is rotated by fixed angle and produces sample, and random angles produce great amount of samples and carry out training pattern
The recognition accuracy to rotation Chinese character all can be improved.Proof is proposed by the present invention to produce some by rotating original sample
Rotated sample carrys out training pattern and can efficiently identify unknown rotated sample.
Claims (2)
1. a kind of rotation Chinese characters recognition method based on convolutional neural networks model is it is characterised in that comprise the steps：
The caffe deep learning framework platform of convolutional neural networks is built on Linux system；
Training dataset and the test data set with label；Training dataset is original handwritten Chinese character and appoints original Chinese character
The sample that meaning angle rotation obtains adds in training set jointly, and test data set is the hand-written of the Arbitrary Rotation with label
Chinese character；
The training convolutional neural networks model on caffe deep learning framework platform using training dataset and test data set,
Obtain test result, realize the identification to Arbitrary Rotation Chinese character.
2. a kind of rotation Chinese characters recognition method based on convolutional neural networks model according to claim 1, its feature exists
In：In step 2）In, training dataset is gained after HCL2000 one-level handwritten Chinese character Data Centralized Processing, and processing method is
Chinese character is concentrated to rotate data, thus in the training process, described convolutional neural networks model can learn to revolve to more
Turn feature；Test data set is the Chinese character of the arbitrarily angled Random-Rotation that HCL2000 data set is randomly drawed, and with mark
Sign.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610813866.3A CN106408038A (en) | 2016-09-09 | 2016-09-09 | Rotary Chinese character identifying method based on convolution neural network model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610813866.3A CN106408038A (en) | 2016-09-09 | 2016-09-09 | Rotary Chinese character identifying method based on convolution neural network model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106408038A true CN106408038A (en) | 2017-02-15 |
Family
ID=57999511
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610813866.3A Pending CN106408038A (en) | 2016-09-09 | 2016-09-09 | Rotary Chinese character identifying method based on convolution neural network model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106408038A (en) |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107220640A (en) * | 2017-05-23 | 2017-09-29 | 广州绿怡信息科技有限公司 | Character identifying method, device, computer equipment and computer-readable recording medium |
| CN107346629A (en) * | 2017-08-22 | 2017-11-14 | 贵州大学 | A kind of intelligent blind reading method and intelligent blind reader system |
| CN107451559A (en) * | 2017-07-31 | 2017-12-08 | 邱宇轩 | Parkinson's people's handwriting automatic identifying method based on machine learning |
| CN107590497A (en) * | 2017-09-20 | 2018-01-16 | 重庆邮电大学 | Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks |
| CN107844740A (en) * | 2017-09-05 | 2018-03-27 | 中国地质调查局西安地质调查中心 | A kind of offline handwriting, printing Chinese character recognition methods and system |
| CN107958285A (en) * | 2017-11-21 | 2018-04-24 | 深圳普思英察科技有限公司 | The mapping method and device of the neutral net of embedded system |
| CN108229470A (en) * | 2017-12-22 | 2018-06-29 | 北京市商汤科技开发有限公司 | character image processing method, device, equipment and storage medium |
| CN109102037A (en) * | 2018-06-04 | 2018-12-28 | 平安科技（深圳）有限公司 | Chinese model training, Chinese image-recognizing method, device, equipment and medium |
| CN109271989A (en) * | 2018-09-03 | 2019-01-25 | 广东电网有限责任公司东莞供电局 | A kind of hand-written test data automatic identifying method based on CNN and RNN model |
| CN109409363A (en) * | 2018-10-13 | 2019-03-01 | 长沙芯希电子科技有限公司 | The reverse judgement of text image based on content and bearing calibration |
| CN109447081A (en) * | 2018-11-14 | 2019-03-08 | 广东小天才科技有限公司 | A kind of acquisition methods and system of optical character identification data set |
| CN109858488A (en) * | 2018-12-28 | 2019-06-07 | 众安信息技术服务有限公司 | A kind of handwriting samples recognition methods and system based on sample enhancing |
| CN109871904A (en) * | 2019-03-11 | 2019-06-11 | 广东工业大学 | Inscriptions on bones or tortoise shells word identification model and training method, system, equipment, computer media |
| CN110084239A (en) * | 2019-04-10 | 2019-08-02 | 中国科学技术大学 | The method of network training over-fitting when reducing offline hand-written mathematical formulae identification |
| CN110223277A (en) * | 2019-05-28 | 2019-09-10 | 深圳新视智科技术有限公司 | Method, apparatus, terminal device and the storage medium that image generates |
| CN110866499A (en) * | 2019-11-15 | 2020-03-06 | 爱驰汽车有限公司 | Handwritten text recognition method, system, device and medium |
| CN111126408A (en) * | 2019-12-24 | 2020-05-08 | 华中师范大学 | Handwritten chemical organic ring symbol recognition method based on rotation invariance |
| CN112418321A (en) * | 2020-11-24 | 2021-02-26 | 北京沃东天骏信息技术有限公司 | Identification method and device of sign image |
| CN112784675A (en) * | 2020-11-25 | 2021-05-11 | 上海芯翌智能科技有限公司 | Target detection method and device, storage medium and terminal |
| US11302108B2 (en) | 2019-09-10 | 2022-04-12 | Sap Se | Rotation and scaling for optical character recognition using end-to-end deep learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104899571A (en) * | 2015-06-12 | 2015-09-09 | 成都数联铭品科技有限公司 | Random sample generation method for recognition of complex character |
| CN105631482A (en) * | 2016-03-03 | 2016-06-01 | 中国民航大学 | Convolutional neural network model-based dangerous object image classification method |
| CN105893968A (en) * | 2016-03-31 | 2016-08-24 | 华南理工大学 | Text-independent end-to-end handwriting recognition method based on deep learning |
-
2016
- 2016-09-09 CN CN201610813866.3A patent/CN106408038A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104899571A (en) * | 2015-06-12 | 2015-09-09 | 成都数联铭品科技有限公司 | Random sample generation method for recognition of complex character |
| CN105631482A (en) * | 2016-03-03 | 2016-06-01 | 中国民航大学 | Convolutional neural network model-based dangerous object image classification method |
| CN105893968A (en) * | 2016-03-31 | 2016-08-24 | 华南理工大学 | Text-independent end-to-end handwriting recognition method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 高学 等: "基于CNN和随机弹性形变的相似手写汉字识别", 《华南理工大学学报（自然科学版）》 * |
Cited By (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107220640A (en) * | 2017-05-23 | 2017-09-29 | 广州绿怡信息科技有限公司 | Character identifying method, device, computer equipment and computer-readable recording medium |
| CN107220640B (en) * | 2017-05-23 | 2020-07-17 | 广州绿怡信息科技有限公司 | Character recognition method, character recognition device, computer equipment and computer-readable storage medium |
| CN107451559A (en) * | 2017-07-31 | 2017-12-08 | 邱宇轩 | Parkinson's people's handwriting automatic identifying method based on machine learning |
| CN107346629A (en) * | 2017-08-22 | 2017-11-14 | 贵州大学 | A kind of intelligent blind reading method and intelligent blind reader system |
| CN107844740A (en) * | 2017-09-05 | 2018-03-27 | 中国地质调查局西安地质调查中心 | A kind of offline handwriting, printing Chinese character recognition methods and system |
| CN107590497A (en) * | 2017-09-20 | 2018-01-16 | 重庆邮电大学 | Off-line Handwritten Chinese Recognition method based on depth convolutional neural networks |
| CN107958285A (en) * | 2017-11-21 | 2018-04-24 | 深圳普思英察科技有限公司 | The mapping method and device of the neutral net of embedded system |
| CN108229470A (en) * | 2017-12-22 | 2018-06-29 | 北京市商汤科技开发有限公司 | character image processing method, device, equipment and storage medium |
| US11275961B2 (en) * | 2017-12-22 | 2022-03-15 | Beijing Sensetime Technology Development Co., Ltd. | Character image processing method and apparatus, device, and storage medium |
| CN109102037B (en) * | 2018-06-04 | 2024-03-05 | 平安科技（深圳）有限公司 | Chinese model training and Chinese image recognition method, device, equipment and medium |
| CN109102037A (en) * | 2018-06-04 | 2018-12-28 | 平安科技（深圳）有限公司 | Chinese model training, Chinese image-recognizing method, device, equipment and medium |
| CN109271989A (en) * | 2018-09-03 | 2019-01-25 | 广东电网有限责任公司东莞供电局 | A kind of hand-written test data automatic identifying method based on CNN and RNN model |
| CN109409363B (en) * | 2018-10-13 | 2021-11-12 | 长沙芯希电子科技有限公司 | Content-based text image inversion judgment and correction method |
| CN109409363A (en) * | 2018-10-13 | 2019-03-01 | 长沙芯希电子科技有限公司 | The reverse judgement of text image based on content and bearing calibration |
| CN109447081B (en) * | 2018-11-14 | 2021-07-23 | 广东小天才科技有限公司 | Method and system for acquiring optical character recognition data set |
| CN109447081A (en) * | 2018-11-14 | 2019-03-08 | 广东小天才科技有限公司 | A kind of acquisition methods and system of optical character identification data set |
| CN109858488A (en) * | 2018-12-28 | 2019-06-07 | 众安信息技术服务有限公司 | A kind of handwriting samples recognition methods and system based on sample enhancing |
| CN109858488B (en) * | 2018-12-28 | 2021-09-17 | 众安信息技术服务有限公司 | Handwritten sample recognition method and system based on sample enhancement |
| CN109871904A (en) * | 2019-03-11 | 2019-06-11 | 广东工业大学 | Inscriptions on bones or tortoise shells word identification model and training method, system, equipment, computer media |
| CN110084239A (en) * | 2019-04-10 | 2019-08-02 | 中国科学技术大学 | The method of network training over-fitting when reducing offline hand-written mathematical formulae identification |
| CN110223277A (en) * | 2019-05-28 | 2019-09-10 | 深圳新视智科技术有限公司 | Method, apparatus, terminal device and the storage medium that image generates |
| US11302108B2 (en) | 2019-09-10 | 2022-04-12 | Sap Se | Rotation and scaling for optical character recognition using end-to-end deep learning |
| CN110866499A (en) * | 2019-11-15 | 2020-03-06 | 爱驰汽车有限公司 | Handwritten text recognition method, system, device and medium |
| CN111126408A (en) * | 2019-12-24 | 2020-05-08 | 华中师范大学 | Handwritten chemical organic ring symbol recognition method based on rotation invariance |
| CN112418321A (en) * | 2020-11-24 | 2021-02-26 | 北京沃东天骏信息技术有限公司 | Identification method and device of sign image |
| CN112784675A (en) * | 2020-11-25 | 2021-05-11 | 上海芯翌智能科技有限公司 | Target detection method and device, storage medium and terminal |
| CN112784675B (en) * | 2020-11-25 | 2023-06-30 | 上海芯翌智能科技有限公司 | Target detection method and device, storage medium and terminal |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106408038A (en) | Rotary Chinese character identifying method based on convolution neural network model | |
| CN106408039A (en) | Off-line handwritten Chinese character recognition method carrying out data expansion based on deformation method | |
| CN109214282B (en) | A kind of three-dimension gesture critical point detection method and system neural network based | |
| CN108427924B (en) | Text regression detection method based on rotation sensitive characteristics | |
| Chen et al. | Beyond human recognition: A CNN-based framework for handwritten character recognition | |
| Tao et al. | Principal component 2-D long short-term memory for font recognition on single Chinese characters | |
| CN111753828B (en) | Natural scene horizontal character detection method based on deep convolutional neural network | |
| CN105426919B (en) | The image classification method of non-supervisory feature learning is instructed based on conspicuousness | |
| CN108009222B (en) | Three-dimensional model retrieval method based on better view and deep convolutional neural network | |
| CN108509881A (en) | A kind of the Off-line Handwritten Chinese text recognition method of no cutting | |
| WO2021139175A1 (en) | Electric power operation ticket character recognition method based on convolutional neural network | |
| CN105224937B (en) | Fine granularity semanteme color pedestrian recognition methods again based on human part position constraint | |
| Mahmoud | Recognition of writer-independent off-line handwritten Arabic (Indian) numerals using hidden Markov models | |
| CN107729865A (en) | A kind of handwritten form mathematical formulae identified off-line method and system | |
| CN107844740A (en) | A kind of offline handwriting, printing Chinese character recognition methods and system | |
| CN104809731B (en) | A kind of rotation Scale invariant scene matching method based on gradient binaryzation | |
| CN105893968A (en) | Text-independent end-to-end handwriting recognition method based on deep learning | |
| Bhattacharya et al. | A sigma-lognormal model-based approach to generating large synthetic online handwriting sample databases | |
| CN104239872A (en) | Abnormal Chinese character identification method | |
| Lin et al. | Font generation based on least squares conditional generative adversarial nets | |
| Akhand et al. | Bangla handwritten numeral recognition using convolutional neural network | |
| CN110197255A (en) | A kind of deformable convolutional network based on deep learning | |
| He et al. | Context-aware mathematical expression recognition: An end-to-end framework and a benchmark | |
| Qi et al. | Exploring illumination robust descriptors for human epithelial type 2 cell classification | |
| Tripathy | Reconstruction of oriya alphabets using Zernike moments |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170215 |
|
| RJ01 | Rejection of invention patent application after publication |