CN109583328A - A kind of depth convolutional neural networks character identifying method being embedded in partially connected - Google Patents
A kind of depth convolutional neural networks character identifying method being embedded in partially connected Download PDFInfo
- Publication number
- CN109583328A CN109583328A CN201811345088.5A CN201811345088A CN109583328A CN 109583328 A CN109583328 A CN 109583328A CN 201811345088 A CN201811345088 A CN 201811345088A CN 109583328 A CN109583328 A CN 109583328A
- Authority
- CN
- China
- Prior art keywords
- convolutional neural
- neural networks
- text
- depth convolutional
- partially connected
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
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
- G06V30/40—Document-oriented image-based pattern recognition
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/24—Aligning, centring, orientation detection or correction of the image
- G06V10/243—Aligning, centring, orientation detection or correction of the image by compensating for image skew or non-uniform image deformations
-
- 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
- G06V30/40—Document-oriented image-based pattern recognition
- G06V30/41—Analysis of document content
- G06V30/413—Classification of content, e.g. text, photographs or tables
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Character Discrimination (AREA)
Abstract
The invention discloses a kind of depth convolutional neural networks character identifying methods for being embedded in partially connected, belong to Information and Communication Engineering field.Character identifying method provided by the invention, it is based on depth convolutional neural networks, unlike existing method, the present invention is the improvement to depth convolutional neural networks structure, aiming at the problem that memory overflows when the training present in the legacy network, it is embedded in partially connected layer in original depth convolutional neural networks, increases the width of network, reduce parameter, reduces the dimension of parameter space.To save the required hardware spending in network training and test, reduce occupied memory headroom when trained network, required duration when shortening trained.
Description
Technical field
The present invention relates to Information and Communication Engineering technical field, especially a kind of depth convolutional Neural for being embedded in partially connected
Network character identifying method.
Background technique
Information technology is developed rapidly with the universal of computer, image as the important transmitting carrier in information technology,
Very important status is accounted in information propagation, and is incorporated in our daily life.The artificial intelligence of computer is general
In the life for being applied to us, all trades and professions are assisted in using the efficiently automatic processing capacity of computer and have become mitigation people
Employee makees, and accelerates a kind of trend of working efficiency.However computer can not pass through vision as the mankind directly acquires image
Included in information, thus image information be identified as this process of text information for computer understanding image have it is important
Meaning.Currently, image recognition has been applied to various scenes, is such as used for the Vehicular real time monitoring of road, obtains license plate number information,
The identification of the certificate photo of various government bodies and data input etc., assistance application greatly improves working efficiency, simultaneously also
Reduce error rate, largely release human resources, to reduce cost of labor.
Character recognition is intended to that text information is detected and identified from picture, that is, is scanned to text information, then to image
File is analyzed and processed, and obtains the process of text and layout information.It is mainly used in document identification and certificate identification.Document
Identification can digitize document printing quick and precisely to extract effective information, and certificate identification is then by papers-scanning or copy
Digitlization reduces working strength to improve working efficiency.For the identification process of character in image, compared to artificial, meter
Embodiment is had the advantage that undoubtedly in the simple and duplicate work of calculation machine processing.It is to simulate the intelligence of human vision to target image
Judged and is identified.By obtaining the image comprising text object, and the image comprising text information is subjected to digitlization knowledge
Other process.
Conventional method requires to train detection model and identification model by manual extraction feature greatly, due to low-level image feature
The distinctive semantic gap between high-level semantic is single when coping with the variation of multiclass font and complex background interferes
Feature selecting is just no longer applicable in.Existing character identifying method, the character recognition engine Tesseract for having Google to increase income are (a kind of
Character recognition technologies), although it is in identification English alphabet and fruitful in Arabic numerals, its identification on Chinese character
There is no so ideal, subsequent improvement work are also cumbersome and complicated.Due to the fast development of deep learning, it is based on deep learning
Character recognition technologies emerge in multitude.Artificial neural network serves as feature extractor and classifier, using character picture as input,
Recognition result can directly be exported.This is also achieved that identification technology end to end.But character recognition is carried out with deep learning
Also there is its drawback, the training of neural network needs a large amount of training data, and with the intensification of network structure, required hardware is opened
Pin also has to take into account.
Although being carried out using the method for deep learning, character recognition is more efficient, and the training of neural network needs a large amount of
Training data.Nowadays, in order to keep recognition result more accurate, precision is higher, is usually mentioned using depth convolutional neural networks
Take feature, identification classification.But as network is deeper and deeper, gradient extinction tests are more obvious, required hardware spending also compared with
It is considerable.
Summary of the invention
A kind of insertion partially connected is provided the technical problem to be solved by the present invention is to overcome the deficiencies in the prior art
Depth convolutional neural networks character identifying method improves original depth convolutional neural networks.Retain its close connection
Thought, to guarantee the mutual transmitting of gradient information.It is embedded in partially connected layer in the module, to save in network training and survey
Required hardware spending when examination.
The present invention uses following technical scheme to solve above-mentioned technical problem:
A kind of depth convolutional neural networks character identifying method of the insertion partially connected proposed according to the present invention, including with
Lower step:
Training set is sent into the depth convolutional neural networks detection mould for being used for the insertion partially connected of character recognition by step 1
In type, the depth convolutional neural networks detection model of insertion partially connected is trained, after the completion of training, what preservation was trained
It is embedded in the model parameter of partially connected, is used for character recognition;
The structure of depth convolutional neural networks detection model are as follows: binary channels first is carried out simultaneously to the picture in the training set of input
Row processing, wherein a channel processing mode are as follows: it first passes through convolutional layer and extracts feature, form characteristic pattern, then by pond layer, it is right
Extracted characteristic pattern is compressed;The processing mode in another channel are as follows: using convolution kernel to the picture in the training set of input
Carry out dimensionality reduction;The processing result in the two channels is connected by cascading layers;Then, then by sequentially connected first closely connect
Connection module, First Transition module, the second close link block, the second transitional module and the close link block of third extract feature,
After feature extraction finishes, the output of the close link block of third is passed through to normalization layer, active coating, displacement layer and smoothing layer again
Subsequent processing is carried out, the result of output layer is finally obtained;
Step 2 first uses the existing convolutional neural networks model for text detection under natural scene to test picture
In text information positioned and extracted, to the text information extracted be put into it is trained insertion partially connected depth roll up
In product neural network detection model, tested.
It is further as a kind of depth convolutional neural networks character identifying method for being embedded in partially connected of the present invention
Prioritization scheme, the detailed process being trained to the depth convolutional neural networks detection model of insertion partially connected in step 1 is such as
Under:
Picture in data set is uniformly truncated into unified size by step 1.1, includes fixed number in every picture
Character;Meanwhile one dictionary comprising commonly used word of production;Each word corresponds to a unique serial number, each in every picture
Sequence composed by the corresponding serial number of word is exactly the label of this picture;Every picture is tagged, and be proportionally divided into
Training set and test set;
Step 1.2, be loaded into the insertion partially connected depth convolutional neural networks detection model pre-training Model Weight,
Start to train the depth convolutional neural networks detection model of the insertion partially connected;
Step 1.3, in the training process sets the threshold value stopped ahead of time as n, and n is the integer greater than 2 and less than 7, if
The best test set error of n iteration is not identical, then return step 1.2, if the best test set error of n iteration no longer mentions
Precision is risen, then deconditioning, stores the model parameter of trained insertion partially connected;And record training used in memory with
Duration.
It is further as a kind of depth convolutional neural networks character identifying method for being embedded in partially connected of the present invention
Prioritization scheme, is tested that specific step is as follows in step 2:
Step 2.1, using the existing convolutional neural networks model for text detection under natural scene to test picture
In text information positioned and extracted;Essential characteristic first is extracted using depth convolutional neural networks and forms characteristic pattern, in institute
Sliding window is done with k × k window on obtained characteristic pattern, it is k × k × C feature vector that each window, which obtains a length, and C is logical
Road number, k are the size of window in sliding window operation, these feature vectors are for determining that target waits for favored area;V={ vc,vhIt is prediction
The coordinate of text box,For actual text frame coordinate, vcTo predict the longitudinally central position of text box, vhFor prediction
The height of text box,The longitudinally central position of actual text frame,For the height of actual text frame;
Wherein,Indicate the centre coordinate of text box vertical direction, haFor the height of text box, cyIndicate predicted text
The centre coordinate of this frame vertical direction, h are the height for the text box predicted,Indicate the center of actual text frame vertical direction
Coordinate, h*Indicate the height of actual text frame;
Obtaining the coordinate v={ v of prediction text boxc,vhAfter, it is filtered out with the non-maxima suppression algorithm of a standard
Extra prediction block;And the text chunk of prediction is merged into line of text;It completes to position the text information in test picture;
Step 2.2, the text information for extracting positioning are sent into the depth convolution of the trained insertion partially connected of step 1
In neural network detection model, the test of character recognition is carried out.
It is further as a kind of depth convolutional neural networks character identifying method for being embedded in partially connected of the present invention
Prioritization scheme, the first close link block are used to extract initial characteristics, and First Transition module is used to reduce the first close connection mould
The port number of block output carries out dimensionality reduction to the output of the first close link block, and the second close link block is used to further mention
Feature is taken, the second transitional module is used to reduce the port number of the second close link block output, to the second close link block
Output carries out dimensionality reduction, and the close link block of third is used to extract feature again, increases the ability to express of feature.
It is further as a kind of depth convolutional neural networks character identifying method for being embedded in partially connected of the present invention
Prioritization scheme, k 3.
The invention adopts the above technical scheme compared with prior art, has following technical effect that
The present invention improves original depth convolutional neural networks, retains its close-connected thought, to guarantee ladder
Spend the mutual transmitting of information;It is embedded in partially connected layer in the module, increases the width of network, reduces parameter, improve simultaneously
Its detection accuracy;Reduce the dimension of parameter space;To save the required hardware spending in network training and test, subtract
Occupied memory headroom when trained network is lacked, required duration when shortening trained.
Detailed description of the invention
Fig. 1 is the depth convolutional neural networks character recognizing process figure for being embedded in partially connected.
Fig. 2 is accuracy curve on training set and test set.
Fig. 3 is error curve on training set and test set.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with the accompanying drawings and the specific embodiments
The present invention will be described in detail.
Fig. 1 is the depth convolutional neural networks character recognizing process figure for being embedded in partially connected;It is specific as follows:
Step 1: improving original depth convolutional neural networks
The thought that partially connected is combined in original depth convolutional neural networks structure, reduces parameter, and it is empty to reduce parameter
Between dimension, to achieve the purpose that save the required hardware spending in network training and test.Improved nerve net
Network structure is as shown in table 1.Table 2 is explained the term in table 1.
Table 1 is improved deep neural network structure
Table 2 is that the term in table 1 is explained
The structure of depth convolutional neural networks detection model are as follows: binary channels first is carried out simultaneously to the picture in the training set of input
Row processing, wherein a channel processing mode are as follows: it first passes through convolutional layer and extracts feature, form characteristic pattern, then by pond layer, it is right
Extracted characteristic pattern is compressed;The processing mode in another channel are as follows: utilize the figure in the training set of small convolution verification input
Piece carries out dimensionality reduction;The processing result in the two channels is connected by cascading layers;Then, then it is close by sequentially connected first
Link block, First Transition module, the second close link block, the second transitional module and the close link block of third extract feature
And then subsequent processing is carried out by normalization layer, active coating, displacement layer and smoothing layer, finally obtain the result of output layer;
First close link block is mainly used to extract initial characteristics, and First Transition module is placed on the first close connection mould
After block, for reducing the port number of the first close link block output, dimensionality reduction is carried out to the output of the first close link block,
To reach reduction parameter, the purpose of parameter space dimension is reduced.Second close link block is mainly used to further extract feature,
Second transitional module is placed on after the second close link block, for reducing the port number of the second close link block output,
Dimensionality reduction is carried out to the output of the second close link block, to reach reduction parameter, reduces the purpose of parameter space dimension.Third is tight
Close link block is mainly used to extract feature again, increases the ability to express of feature.
The structure of depth convolutional neural networks detection model are as follows: binary channels first is carried out simultaneously to the picture in the training set of input
Row processing, wherein a channel processing mode are as follows: it first passes through convolutional layer and extracts feature, form characteristic pattern, then by pond layer, it is right
Extracted characteristic pattern is compressed;The processing mode in another channel are as follows: using convolution kernel to the picture in the training set of input
Carry out dimensionality reduction;The processing result in the two channels is connected by cascading layers;Then, then by sequentially connected first closely connect
Connection module, First Transition module, the second close link block, the second transitional module and the close link block of third extract feature.
First close link block includes n (value of n is between 6-48) two kinds of convolution kernel A and B of different sizes, first to cascading layers
Output be normalized, then make to normalize result by active coating, be then utilized respectively convolution kernel A of different sizes and B pairs
The output of active coating carries out convolution operation, and the result of convolution kernel A and B processing is spliced by cascading layers.At so repeating
N times are managed, are sent in First Transition module.In First Transition module, closely connect first with the convolution kernel that size is C to first
The output of connection module carries out convolution operation, then carries out pondization operation to the result after process of convolution.The is sent into after pondization processing
Two close link blocks.Second close link block includes n (value of n is between 6-48) two kinds of convolution kernels of different sizes
A and B (value of n, A, B are identical as the first close link block), is first normalized the output of First Transition module, then make
Result is normalized by active coating, convolution kernel A and B of different sizes is then utilized respectively and convolution behaviour is carried out to the output of active coating
Make, the result of convolution kernel A and B processing is spliced by cascading layers.So reprocessing n times, are sent to the second stage die
In block.In the second transitional module, first with the convolution kernel that size is C (value of C is identical with First Transition module) to the
The output of two close link blocks carries out convolution operation, then carries out pondization operation to the result after process of convolution.Pondization processing
It is sent into the close link block of third afterwards.The close link block of third also include n (value of n is between 6-48) two kinds of sizes not
Same convolution kernel A and B (value of n, A, B are identical as the first close link block), is first normalized the output of cascading layers,
Make to normalize result again by active coating, is then utilized respectively convolution kernel A and B of different sizes and the output of active coating is rolled up
Product operation splices the result of convolution kernel A and B processing by cascading layers.So reprocessing n times, so far, feature extraction
It finishes.Later, after the output of the close link block of third to be passed through to normalization layer, active coating, displacement layer and smoothing layer progress again
Continuous processing finally obtains the result of output layer.
In deep learning network, with the intensification of network, gradient disappearance problem can be further obvious.Dense block is (tight
Close link block) layer there are problems that efficiently solve gradient disappearance.Each dense block layers by normalizing
Layer, active coating, convolutional layer and cascading layers are constituted.Its premise of maximum information transmission between layers in guaranteeing network
Under, all layers are connected, that is, output of each layer of the input from all layers of front.xlRepresent l layers of output, HlIndicate one
Nonlinear transformation.[x0,x1,...,xl-1] represent 0 Dao l layers output.Then,
xl=Hl([x0,x1,...,xl-1])
Block layers of each dense convolution operation comprising 81 × 1 and 3 × 3.1 × 1 convolution operation is mainly used for
Dimensionality reduction, it can greatly reduce number of parameters and calculation amount.
Each transition layer (transition zone/transitional module) is abandoned by normalization layer, active coating, convolutional layer
Layer, pond layer composition.Block layers of institute's output ground channel number of Dense are more, therefore transition layer is placed on
Two block layers of dense intermediate, by 1 × 1 convolution kernel come dimensionality reduction.Since each layer is all mutual in dense block
Connection, in order to avoid over-fitting, dropout layer (discarding layers) are added in transition layer to reduce at random point
Branch.
Step 2: production data set.
Picture in training set is uniformly truncated into 280 × 32 sizes, includes ten characters in every picture.From corpus
Choose include Chinese character, English alphabet, number and punctuate totally 5990 characters as label.By the tagged conduct of every picture
Training sample.Made training sample shares 3644007, goes out training set and test set according to the ratio cut partition of 99:1,
Middle training set is 3607567, and test set is 36440.
Step 3: being loaded into pre-training model, initialize parameters, start to train network.For trained GPU (at figure
Reason device) it is configured to Tesla P100, parameter initialization value is as shown in table 3.
3 parameter initialization value of table
Parameter | img_h | img_w | batch_size | maxlabellength | epoch | nclass |
Meaning | Picture height | Picture width | Block size | Maximum tag length | Wheel | Classification |
Value | 32 | 280 | 64 | 10 | 30 | 5990 |
In the training process, the threshold value of early stopping is set as 5, if the preferably verifying of 5 iteration (epoch)
It is not identical to collect error, then return step 1.3, if 5 iteration (epoch) verifying collection errors will not promote precision again, stops
Training, storage model.And record memory used in training and duration.Trained model is drawn again in training set and test set
On curve, as shown in Figure 2.Error change curve of the model trained on training set and test set is as shown in Figure 3.It improves
Preceding network can not be trained normally under the conditions of the hardware memory of 32G, it may appear that the problem of memory overflows.And improved net
Network can complete training under the hardware condition, and shorten trained duration to a certain extent.
Step 4: utilizing the existing convolutional neural networks model for text detection under natural scene in test picture
Text information positioned and extracted.The network is designed for text detection, first (is used for large-scale image using VGG16
The depth convolutional neural networks of identification) essential characteristic formation characteristic pattern is extracted, it is made on obtained characteristic pattern of 3 × 3 windows
Sliding window, each window, which obtains the feature vector that a length is 3 × 3 × C (C is port number), to be used to predict and set between frame
Offset distance.Detail is as follows, v={ vc,vhIt is opposite prediction coordinate,For actual text frame coordinate.Table
Show the centre coordinate of text box vertical direction, haFor the height of text box, these can directly be calculated from input picture.
cyIndicating the centre coordinate of predicted text box vertical direction, h is the height for the text box predicted,Indicate actual text
The centre coordinate of frame vertical direction, h*Indicate the height of actual text frame.
After obtaining the text box of dense prediction, extra prediction is filtered out with the non-maxima suppression algorithm of a standard
Frame.The text chunk predicted is finally merged into line of text, prepares output.
Step 5: the text information extracted being sent into trained model, the test of character recognition is carried out.Test institute
Business license and text segment are sent into deep by 1050 Ti of GPU (graphics processor) model GeForce GTX used respectively
Degree convolutional neural networks are tested.
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by those familiar with the art, all answers
It is included within the scope of protection of the present invention.
Claims (5)
1. a kind of depth convolutional neural networks character identifying method for being embedded in partially connected, which comprises the following steps:
Training set is sent into the depth convolutional neural networks detection model for the insertion partially connected for being used for character recognition by step 1,
The depth convolutional neural networks detection model of insertion partially connected is trained, after the completion of training, saves trained insertion
The model parameter of partially connected is used for character recognition;
The structure of depth convolutional neural networks detection model are as follows: binary channels first is carried out to the picture in the training set of input and is located parallel
Reason, wherein a channel processing mode are as follows: first pass through convolutional layer and extract feature, form characteristic pattern, then by pond layer, to being mentioned
The characteristic pattern taken is compressed;The processing mode in another channel are as follows: the picture in the training set of input is carried out using convolution kernel
Dimensionality reduction;The processing result in the two channels is connected by cascading layers;Then, then by sequentially connected first mould is closely connected
Block, First Transition module, the second close link block, the second transitional module and the close link block of third extract feature, feature
After extraction finishes, the output of the close link block of third is passed through to normalization layer, active coating, displacement layer and smoothing layer again and is carried out
Subsequent processing finally obtains the result of output layer;
Step 2 first uses the existing convolutional neural networks model for text detection under natural scene in test picture
Text information is positioned and is extracted, and the depth convolution mind of trained insertion partially connected is put into the text information extracted
Through being tested in network detection model.
2. a kind of depth convolutional neural networks character identifying method for being embedded in partially connected according to claim 1, special
Sign is, is trained that detailed process is as follows in step 1 to the depth convolutional neural networks detection model of insertion partially connected:
Picture in data set is uniformly truncated into unified size by step 1.1, includes the word of fixed number in every picture
Symbol;Meanwhile one dictionary comprising commonly used word of production;Each word corresponds to a unique serial number, each word in every picture
Sequence composed by corresponding serial number is exactly the label of this picture;Every picture is tagged, and proportionally it is divided into instruction
Practice collection and test set;
Step 1.2, be loaded into the insertion partially connected depth convolutional neural networks detection model pre-training Model Weight, start
The depth convolutional neural networks detection model of the training insertion partially connected;
Step 1.3, in the training process sets the threshold value stopped ahead of time as n, and n is the integer greater than 2 and less than 7, if n is a
The best test set error of iteration is not identical, then return step 1.2, if the best test set error of n iteration is no longer promoted
Precision, then deconditioning, stores the model parameter of trained insertion partially connected;And record training used in memory and when
It is long.
3. a kind of depth convolutional neural networks character identifying method for being embedded in partially connected according to claim 1, special
Sign is, is tested that specific step is as follows in step 2:
Step 2.1, using the existing convolutional neural networks model for text detection under natural scene in test picture
Text information is positioned and is extracted;Essential characteristic first is extracted using depth convolutional neural networks and forms characteristic pattern, acquired
Characteristic pattern on k × k window do sliding window, it is k × k × C feature vector that each window, which obtains a length, and C is port number,
K is the size of window in sliding window operation, these feature vectors are for determining that target waits for favored area;V={ vc,vhIt is prediction text
The coordinate of frame,For actual text frame coordinate, vcTo predict the longitudinally central position of text box, vhTo predict text
The height of frame,The longitudinally central position of actual text frame,For the height of actual text frame;
vh=log (h/ha)
Wherein,Indicate the centre coordinate of text box vertical direction, haFor the height of text box, cyIndicate predicted text box
The centre coordinate of vertical direction, h are the height for the text box predicted,Indicate that the center of actual text frame vertical direction is sat
Mark, h*Indicate the height of actual text frame;
Obtaining the coordinate v={ v of prediction text boxc,vhAfter, it is extra to be filtered out with the non-maxima suppression algorithm of a standard
Prediction block;And the text chunk of prediction is merged into line of text;It completes to position the text information in test picture;
Step 2.2, the text information for extracting positioning are sent into the depth convolutional Neural of the trained insertion partially connected of step 1
In network detection model, the test of character recognition is carried out.
4. a kind of depth convolutional neural networks character identifying method for being embedded in partially connected according to claim 1, special
Sign is that the first close link block is used to extract initial characteristics, and First Transition module is used to reduce by the first close link block
The port number of output carries out dimensionality reduction to the output of the first close link block, and the second close link block is used to further extract
Feature, the second transitional module is used to reduce the port number of the second close link block output, to the defeated of the second close link block
Dimensionality reduction is carried out out, and the close link block of third is used to extract feature again, increases the ability to express of feature.
5. a kind of depth convolutional neural networks character identifying method for being embedded in partially connected according to claim 3, special
Sign is, k 3.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811345088.5A CN109583328B (en) | 2018-11-13 | 2018-11-13 | Sparse connection embedded deep convolutional neural network character recognition method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811345088.5A CN109583328B (en) | 2018-11-13 | 2018-11-13 | Sparse connection embedded deep convolutional neural network character recognition method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109583328A true CN109583328A (en) | 2019-04-05 |
CN109583328B CN109583328B (en) | 2021-09-03 |
Family
ID=65922152
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811345088.5A Active CN109583328B (en) | 2018-11-13 | 2018-11-13 | Sparse connection embedded deep convolutional neural network character recognition method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109583328B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111832437A (en) * | 2020-06-24 | 2020-10-27 | 万翼科技有限公司 | Building drawing identification method, electronic equipment and related product |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3282397A1 (en) * | 2016-08-11 | 2018-02-14 | Vivante Corporation | Zero coefficient skipping convolution neural network engine |
CN107977620A (en) * | 2017-11-29 | 2018-05-01 | 华中科技大学 | A kind of multi-direction scene text single detection method based on full convolutional network |
CN108229479A (en) * | 2017-08-01 | 2018-06-29 | 北京市商汤科技开发有限公司 | The training method and device of semantic segmentation model, electronic equipment, storage medium |
CN108734284A (en) * | 2017-04-24 | 2018-11-02 | 英特尔公司 | Rely on the deep learning of real-time context |
-
2018
- 2018-11-13 CN CN201811345088.5A patent/CN109583328B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3282397A1 (en) * | 2016-08-11 | 2018-02-14 | Vivante Corporation | Zero coefficient skipping convolution neural network engine |
CN108734284A (en) * | 2017-04-24 | 2018-11-02 | 英特尔公司 | Rely on the deep learning of real-time context |
CN108229479A (en) * | 2017-08-01 | 2018-06-29 | 北京市商汤科技开发有限公司 | The training method and device of semantic segmentation model, electronic equipment, storage medium |
CN107977620A (en) * | 2017-11-29 | 2018-05-01 | 华中科技大学 | A kind of multi-direction scene text single detection method based on full convolutional network |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111832437A (en) * | 2020-06-24 | 2020-10-27 | 万翼科技有限公司 | Building drawing identification method, electronic equipment and related product |
CN111832437B (en) * | 2020-06-24 | 2024-03-01 | 万翼科技有限公司 | Building drawing identification method, electronic equipment and related products |
Also Published As
Publication number | Publication date |
---|---|
CN109583328B (en) | 2021-09-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10817741B2 (en) | Word segmentation system, method and device | |
CN107563385A (en) | License plate character recognition method based on depth convolution production confrontation network | |
CN111652332B (en) | Deep learning handwritten Chinese character recognition method and system based on two classifications | |
CN113011357B (en) | Depth fake face video positioning method based on space-time fusion | |
CN106156766A (en) | The generation method and device of line of text grader | |
CN105469047A (en) | Chinese detection method based on unsupervised learning and deep learning network and system thereof | |
CN103544504B (en) | Scene character recognition method based on multi-scale map matching core | |
CN105224951A (en) | A kind of vehicle type classification method and sorter | |
CN108776774A (en) | A kind of human facial expression recognition method based on complexity categorization of perception algorithm | |
CN105760891A (en) | Chinese character verification code recognition method | |
CN108664975A (en) | A kind of hand-written Letter Identification Method of Uighur, system and electronic equipment | |
CN111815609B (en) | Pathological image classification method and system based on context awareness and multi-model fusion | |
Chaabouni et al. | Static and dynamic features for writer identification based on multi-fractals. | |
CN116012653A (en) | Method and system for classifying hyperspectral images of attention residual unit neural network | |
CN106650696A (en) | Handwritten electrical element identification method based on singular value decomposition | |
CN105469099A (en) | Sparse-representation-classification-based pavement crack detection and identification method | |
CN111242114B (en) | Character recognition method and device | |
CN109583328A (en) | A kind of depth convolutional neural networks character identifying method being embedded in partially connected | |
Xue | Optical character recognition | |
Gandhi et al. | An attempt to recognize handwritten Tamil character using Kohonen SOM | |
US20240013368A1 (en) | Pavement nondestructive detection and identification method based on small samples | |
Rashid et al. | Discriminative learning for script recognition | |
CN110598650A (en) | License plate false alarm filtering method and device, electronic equipment and storage medium | |
CN114581780A (en) | Tunnel surface crack detection method for improving U-Net network structure | |
Álvaro et al. | Page segmentation of structured documents using 2d stochastic context-free grammars |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |