CN109034281A - The Chinese handwritten body based on convolutional neural networks is accelerated to know method for distinguishing - Google Patents
The Chinese handwritten body based on convolutional neural networks is accelerated to know method for distinguishing Download PDFInfo
- Publication number
- CN109034281A CN109034281A CN201810789695.4A CN201810789695A CN109034281A CN 109034281 A CN109034281 A CN 109034281A CN 201810789695 A CN201810789695 A CN 201810789695A CN 109034281 A CN109034281 A CN 109034281A
- Authority
- CN
- China
- Prior art keywords
- classifier
- network
- layer
- parameter
- neural networks
- 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
- G06N3/04—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
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)
Abstract
A kind of accelerated method of the Chinese handwritten body identification based on convolutional neural networks, includes the following steps: that hand-written character image data pre-processes, the classification of convolutional neural networks is facilitated to realize;Convolutional neural networks structure is built, the first classifier is added in shallow-layer, the second classifier is added in middle layer, and third classifier is added in top layer;Initialization network parameter, the parameter of fixed shallow-layer network and the first classifier instruct network training using the loss function of the second classifier, update the parameter of mid-level network and the second classifier;The parameter of fixed shallow-layer network, mid-level network, the first classifier and the second classifier instructs network training using the loss function of third classifier, updates the parameter of overlay network and third classifier;The loss function of three classifiers is weighted, the parameter of loss function fine tuning whole network and three classifiers after weighting is utilized;Three classifiers of network are tested on test set, assess the performance of classifier, and calculation amount needed for three classifiers of calculating.
Description
Technical field
The present invention relates to depth learning technology, especially application of the convolutional neural networks in handwriting recongnition is specifically related to
And be it is a kind of based on convolutional neural networks Chinese handwritten body identification accelerated method.
Background technique
Due to defeated in document of taking pictures, post envelope, check, manuscript letter Optical Character Recognition system and handwriting
Enter the wide application prospect in equipment, the identification of Chinese handwritten body is always an important research direction of area of pattern recognition, is obtained
The extensive research and concern of academia and industry are arrived.Traditional Chinese handwritten body identifying system is mainly manually set using some
The feature of meter classifies to character using classifiers such as second judgement functions such as Gabor characteristic and Gradient feature.Closely
Nian Lai, with the rise of deep learning, convolutional neural networks bring extremely effective solution, base to Chinese handwriting recongnition
In result significantly leading traditional method that the method for convolutional neural networks obtains.
However, due to the particularity that convolutional neural networks calculate, the huge multiply-add operation of network needs, such as classical
VGG16 network needs about 15,300,000,000 multiply-add operations, even existing some lightweight networks, it is still desirable to millions of time to multiply
Add operation, such huge operand limits the application of convolutional neural networks.The present invention is based on convolutional neural networks existing
The identification of Chinese handwritten body on, take the scheme for increasing middle layer output, simple sample and difficult sample treated with a certain discrimination, with shallow
The network of layer identifies most of simple sample, and difficult sample then continues to extract the spy for more having distinction in a network
Sign realizes the classification to most of simple sample with a small amount of operation cost, and difficult sample is real with relatively large number of calculating cost
Now classify, the Chinese handwritten body identification accelerated based on convolutional neural networks is reached with this.
Summary of the invention
The accelerated method for the Chinese handwritten body identification that the object of the present invention is to provide a kind of based on convolutional neural networks is
By increasing the scheme of middle layer output, simple sample and difficult sample are treated with a certain discrimination, with the network of shallow-layer to most of letter
Single sample is identified, and difficult sample then continues to extract the feature for more having distinction in a network, with a small amount of operation cost
Realize the classification to most of simple sample, difficult sample is realized with relatively large number of calculating cost classifies, and accelerates to be based on this
The Chinese handwritten body identifying schemes of convolutional neural networks.
The present invention provides a kind of accelerated method of Chinese handwritten body identification based on convolutional neural networks, including walks as follows
It is rapid:
Step 1: the pretreatment of hand-written character image data facilitates the classification of convolutional neural networks to realize;
Step 2: building convolutional neural networks structure, the first classifier C is added in shallow-layer1, the second classification of middle layer addition
Device C2, top layer addition third classifier C3;
Step 3: initialization network parameter, using dividing first kind device C1Loss function instruct network training, update shallow-layer
Network and the first classifier C1Parameter;Fixed shallow-layer network and the first classifier C1Parameter, utilize the second classifier C2Damage
It loses function and instructs network training, update mid-level network and the second classifier C2Parameter;Fixed shallow-layer network, middle layer net
Network, the first classifier C1With the second classifier C2Parameter, utilize third classifier C3Loss function instruct network training, more
New overlay network and third classifier C3Parameter;
Step 4: the loss function of three classifiers being weighted, finely tunes whole network using the loss function after weighting
With the parameter of three classifiers;
Step 5: three classifiers of network being tested on test set, assess the performance of classifier, and calculate
Calculation amount needed for three classifiers.
It can be seen from the above technical proposal that a kind of Chinese handwritten body identification of the acceleration of the present invention based on convolutional neural networks
Method at least have the advantages that one of them:
1, most of character picture can correctly classify in shallow-layer network, and only about 5.7% data need operation to top
Layer network could correctly classify;
2, mid-level network and the classification accuracy of overlay network are closer to, using the voting results of two classifiers,
The discrimination of difficult sample is obviously improved.
Detailed description of the invention
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with specific embodiment, and reference
Attached drawing is described in further detail as rear technology contents of the invention, in which:
Fig. 1 is flow chart of the method for the present invention;
Fig. 2 is the flow chart of step 2 in Fig. 1;
Fig. 3 is the flow chart of step 3 in Fig. 1;
Fig. 4 is the flow chart of step 5 in Fig. 1;
Fig. 5 is the convolutional neural networks structural schematic diagram that the present invention uses;
Fig. 6 is that the present invention accelerates the Chinese handwritten body recognition principle schematic diagram based on convolutional neural networks.
Specific embodiment
It please refers to shown in Fig. 1 to Fig. 6, the present invention provides a kind of adding for Chinese handwritten body identification based on convolutional neural networks
Fast method, includes the following steps:
Step 1: hand-written character image data pre-processes (Preprocess), and the classification of convolutional neural networks is facilitated to realize,
In the step 1, classifying hand-written characters picture is pre-processed, and character picture is used bilinear interpolation (Bilinear
Interpolation) mode zooms to 64 × 64 sizes, and picture is converted into grayscale image, and to the foreground color of picture
(Foreground) it is adjusted with background (Background) color, so that prospect has biggish gray value, close to 255,
Background has lesser gray value, close to 0;
Step 2: building convolutional neural networks structure, the step 2 includes:
Step 21: the basic convolutional neural networks structure of use shares seven convolutional layers (Convolutional Layer),
Three pond layers (Pooling Layer);
Step 22: all convolutional layer convolution kernel sizes (Kernel Size) of use are 3 × 3, sliding step
It (Stride) is 1, convolutional layer output channel number (Channels) is respectively 64,128,128,256,256,512 and 512, respectively
It is referred to as conv1, conv2, conv3, conv4, conv5, conv6 and conv7;
Step 23: it is activation primitive that all convolutional layers, which are all made of prelu function, and formula is as follows:
Wherein, x is the input value of node, and prelu (x) is the input of node by the output valve after activation primitive, and a is
One parameter that can learn;
Step 24: batch normalization (BatchNornalization Layer) is all added between all convolution sum activation,
Formula is as follows:
Wherein x is obtained after convolution as a result, μ and σ are the mean value and variance of data respectively, and γ and β, which are two, to be learnt
Parameter controls the zooming and panning of data respectively;Batch normalization can effectively accelerate network convergence, can also play the effect of regularization
Fruit improves the generalization ability of network;
Step 25: all pond layers are all made of maximum pond (Max Pooling), and window size is 2 × 2, sliding step
A length of 2, after pond layer is located at conv1, conv3 and conv5, respectively refer on behalf of pool1, pool2 and pool3;
Step 26, shallow-layer network includes conv1, pool1, conv2 and conv3 module, and middle layer includes pool2, conv4
With conv5 module, overlay network includes pool3, conv6 and conv7 module;
Step 27: carrying out global uniformly pond (Global Average before being sent into classifier to the output of each layer network
Pooling), i.e., averaged on the various channels is exported to network, to reduce full articulamentum number of parameters, reduces network and deposit
Storage scale;
Step 28: classifier uses full articulamentum, and carries out probabilistic forecasting to each character using softmax function, false
If the node number of output layer is n, respectively (y1, y2…yn), the value y of each output nodei, after softmax function
Become a probability value, n probability value forms a probability distribution (Py1, Py2…Pyn) for i-th of node of output layer, it
Probability value is
Step 3: substep three classifiers of training, initialization network parameter utilize the first classifier C1Loss function refer to
Network training is led, shallow-layer network and the first classifier C are updated1Parameter;Fixed shallow-layer network and the first classifier C1Parameter,
Utilize the second classifier C2Loss function instruct network training, update mid-level network and the second classifier C2Parameter;It is fixed
Shallow-layer network, mid-level network, the first classifier C1With the second classifier C2Parameter, utilize third classifier C3Loss letter
Number instructs network training, updates overlay network and third classifier C3Parameter, include: in the step 3
Step 31: classifier is all made of cross entropy as loss function, and formula is as follows:
H (y, p)=- ∑ yilogpyi
Wherein yiFor the true tag of i-th of sample, pyiFor probability value of i-th of sample on true tag;
Step 32: training process uses the stochastic gradient descent method with momentum term, and formula is as follows:
Wt+1=Wt+Vt+1
Wherein, wt is current network parameter value,For current network gradient value, indicating momentum term, (history gradient is tired
Accumulated amount), it is learning rate, is momentum term weight, initial learning rate is set as 0.1, and when loss stops decline, learning rate is reduced
10 times, momentum term weight is set as 0.9, batch_size and is set as 256;
Step 33: in training process before being sent into classifier, neuron node carries out random drop (Dropout), with
The probability that machine abandons is respectively as follows: the first classifier C1: 0.1, the second classifier C2: 0.2 and third classifier C3: 0.5, dropout
As a kind of effective regularization means, it can be avoided network over-fitting, improve the generalization ability of network;
Step 34: training uses Institute of Automation Chinese monocase data set CASIA-HWDB1.0-1.2 offline, is 2,600,000
Picture, totally 3755 class first class word-base of the national standard;
Step 4: the loss function of three classifiers being weighted, finely tunes whole network using the loss function after weighting
With the parameter of three classifiers, in step 4, the first classifier C1, the second classifier C2With third classifier C3Loss weighted value
Respectively 0.5,0.5 and 1.0;
Step 5: network being tested on test set, assesses the performance of classifier, and calculate three classifier institutes
The calculation amount needed includes: in the step 5
Step 51: test data set is data set used in the offline Chinese handwritten body identification of ICDAR-2013 is competed, and is 260,000
Open character picture;
Step 52: classifier performance takes top-1 accuracy rate as standard, i.e., classifier output most probable value is corresponding
Label as prediction label compared with true tag;
Step 53: test data concentration character picture is first used into the first classifier C1Classify, if the first classifier C1It is defeated
The big Mr. Yu's threshold value t of most probable value out1, then the classifier is exported as final prediction result, otherwise character picture is with the
Two classifier C2Classification;If the second classifier C2Export the big Mr. Yu's threshold value t of most probable value2, then by the second classifier C2Output is made
For final prediction result, otherwise by the second classifier C2With third classifier C3Output is weighted as final prediction result;
Step 54, by the class probability value for the 53 policy class output that takes steps compared with true tag, detection network point
Class accuracy rate, and with directly adopt third classifier C3The accuracy rate of classification compares;
Step 55: be directed to entire data set, that average computational load needed for classification of assessment, and with directly adopt third classifier C3
Calculation amount needed for classification compares, wherein calculation amount is multiply-add number (MAC, Multiply- needed for assorting process
accumulate operation)。
The independent test result of 1 classifier of table
Accuracy rate/% | MAC/million | |
First classifier C1 | 91.73 | 229 |
Second classifier C2 | 95.64 | 457 |
Third classifier C3 | 96.43 | 685 |
The independent test result of classifier is as shown in table 1, wherein the first classifier C1Reached with about 2.29 hundred million operands
91.73% accuracy rate, the second classifier C2Reach 95.64% accuracy rate, third classifier C with about 4.57 hundred million operands3
Reach 96.43% accuracy rate with about 6.85 hundred million operands;
Take E3 strategy as follows to picture progress classification results, wherein about 83.7% sample is in the first classifier C1It completes
Classification, about 10.6% sample is in the second classifier C2Classification is completed, remaining 5.7% sample is in third classifier C3It completes to divide
Class, accuracy rate 97.25%, operand needed for averagely completing a subseries are 2.79 hundred million times;Compared to directlying adopt third
Classifier C3Classify to picture, accuracy rate improves nearly 0.82%, and that average computational load reduces about 60%.
So far, a kind of Chinese handwritten body of the acceleration based on convolutional neural networks of the present invention is known method for distinguishing introduction and is finished.According to
According to above description, those skilled in the art should have clear understanding to the present invention.
It should be noted that in attached drawing or specification text, the convolutional neural networks base unit not described is institute
Belong to form known to a person of ordinary skill in the art in technical field, is not described in detail.In addition, above-mentioned to each element and method
Definition is not limited in various specific structures, shape or the mode mentioned in embodiment, and those of ordinary skill in the art can be to it
It is simply changed or is replaced.
It should also be noted that, the present invention can provide the demonstrations of the parameter comprising particular value, but these parameters are without definite
Equal to corresponding value, but analog value can be similar in acceptable error margin or design constraint.In addition, except non-specifically retouching
The step of stating or must sequentially occurring, there is no restriction for the sequences of above-mentioned steps in listed above, and can be become according to required design
Change or rearranges.And above-described embodiment can based on the considerations of design and reliability, be mixed with each other collocation use or and other
Embodiment mix and match uses, i.e., the technical characteristic in different embodiments can freely form more embodiments.
Particular embodiments described above has carried out further in detail the purpose of the present invention, technical scheme and beneficial effects
Describe in detail bright, it should be understood that the above is only a specific embodiment of the present invention, is not intended to restrict the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution, improvement and etc. done should be included in protection of the invention
Within the scope of.
Claims (6)
1. a kind of accelerated method of the Chinese handwritten body identification based on convolutional neural networks, includes the following steps:
Step 1: the pretreatment of hand-written character image data facilitates the classification of convolutional neural networks to realize;
Step 2: building convolutional neural networks structure, the first classifier is added in shallow-layer, the second classifier, top layer is added in middle layer
Third classifier is added;
Step 3: initialization network parameter instructs network training using the loss function of the first classifier, update shallow-layer network and
The parameter of first classifier;The parameter of fixed shallow-layer network and the first classifier is instructed using the loss function of the second classifier
Network training updates the parameter of mid-level network and the second classifier;Fixed shallow-layer network, mid-level network, the first classifier
With the parameter of the second classifier, network training is instructed using the loss function of third classifier, updates overlay network and third point
The parameter of class device;
Step 4: the loss function of three classifiers being weighted, finely tunes whole network and three using the loss function after weighting
The parameter of a classifier;
Step 5: three classifiers of network being tested on test set, assess the performance of classifier, and calculate three
Calculation amount needed for classifier.
2. the accelerated method of the Chinese handwritten body identification according to claim 1 based on convolutional neural networks, wherein described
In step 1, classifying hand-written characters picture is pre-processed, and picture is zoomed to 64 × 64 sizes using bilinear interpolation mode, is schemed
Piece is converted into grayscale image, and is adjusted to the foreground color and background color of picture, so that prospect has biggish gray scale
Value, close to 255, background has lesser gray value, close to 0.
3. the accelerated method of the Chinese handwritten body identification according to claim 1 based on convolutional neural networks, wherein described
Step 2 includes:
Step 21: the basic convolutional neural networks structure of use shares seven convolutional layers, three pond layers;
Step 22: all convolutional layer convolution kernel sizes of use are 3 × 3, sliding step 1, convolutional layer output channel number point
Not Wei 64,128,128,256,256,512 and 512, respectively refer on behalf of conv1, conv2, conv3, conv4, conv5, conv6
And conv7;
Step 23: it is activation primitive that all convolutional layers, which are all made of prelu function, and formula is as follows:
Wherein, x is the input value of node, and prelu (x) is the input of node by the output valve after activation primitive, and a is one
The parameter that can learn;
Step 24: batch is all added between all convolution sum activation and normalizes, formula is as follows:
Wherein x is to obtain after convolution as a result, μ and σ are the mean value and variance of data respectively, γ and β be two can learning parameter,
The zooming and panning of data are controlled respectively;Batch normalization can effectively accelerate network convergence, can also play the effect of regularization, mention
The generalization ability of high network;
Step 25: all pond layers are all made of maximum pond, and window size is 2 × 2, and sliding step 2, pond layer is located at
After conv1, conv3 and conv5, respectively refer on behalf of pool1, pool2 and pool3;
Step 26, shallow-layer network includes conv1, pool1, conv2 and conv3 module, middle layer include pool2, conv4 and
Conv5 module, overlay network include pool3, conv6 and conv7 module;
Step 27: carrying out global uniformly pond before being sent into classifier to the output of each layer network, i.e., network is exported each
Averaged on channel reduces network storage scale to reduce full articulamentum number of parameters;
Step 28: classifier uses full articulamentum, and carries out probabilistic forecasting to each character using softmax function, it is assumed that defeated
The node number of layer is n, respectively (y out1, y2…yn), the value y of each output nodei, become after softmax function
One probability value, n probability value form a probability distribution (Py1, Py2…Pyn) for i-th of node of output layer, its probability
Value is
4. the accelerated method of the Chinese handwritten body identification according to claim 1 based on convolutional neural networks, wherein described
Include: in step 3
Step 31: classifier is all made of cross entropy as loss function, and formula is as follows:
Wherein yiFor the true tag of i-th of sample, pyiFor probability value of i-th of sample on true tag;
Step 32: training process uses the stochastic gradient descent method with momentum term, and initial learning rate is set as 0.1, works as loss
When stopping decline, learning rate is reduced by 10 times, momentum term is set as 0.9, and batch size is set as 256;
Step 33: in training process before being sent into classifier, neuron node carries out random drop, the probability point of random drop
Not are as follows: the first classifier: 0.1, the second classifier: 0.2 and third classifier: 0.5, dropout as a kind of effective canonical
Change means can be avoided network over-fitting, improve the generalization ability of network;
Step 34: training uses Institute of Automation Chinese monocase data set CASIA-HWDB1.0-1.2 offline, for 2,600,000 figures
Piece, totally 3755 class first class word-base of the national standard.
5. the accelerated method of the Chinese handwritten body identification according to claim 1 based on convolutional neural networks, wherein step 4
In, the loss weighted value of the first classifier, the second classifier and third classifier is respectively 0.5,0.5 and 1.0.
6. the accelerated method of the Chinese handwritten body identification according to claim 5 based on convolutional neural networks, wherein described
Include: in step 5
Step 51: test data set is data set used in the offline Chinese handwritten body identification of ICDAR-2013 is competed, and is 260,000 words
Accord with picture;
Step 52: classifier performance takes top-1 accuracy rate as standard, i.e., by the corresponding mark of classifier output most probable value
Label are as prediction label compared with true tag;
Step 53: test data concentration character picture first being classified with the first classifier, if the output of the first classifier is maximum
The big Mr. Yu's threshold value t of probability value1, then the classifier is exported as final prediction result, otherwise the second classification of character picture
Device classification;If the second classifier exports the big Mr. Yu's threshold value t of most probable value2, then by the output of the second classifier as final prediction
As a result, otherwise the second classifier and the output of third classifier are weighted as final prediction result;
Step 54: by the class probability value for the 53 policy class output that takes steps compared with true tag, detection network class is quasi-
True rate, and compared with the accuracy rate for directlying adopt the classification of third classifier;
Step 55: be directed to entire data set, that average computational load needed for classification of assessment, and with directly adopt third classifier classification institute
The calculation amount needed compares, wherein calculation amount is multiply-add number needed for assorting process.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810789695.4A CN109034281A (en) | 2018-07-18 | 2018-07-18 | The Chinese handwritten body based on convolutional neural networks is accelerated to know method for distinguishing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810789695.4A CN109034281A (en) | 2018-07-18 | 2018-07-18 | The Chinese handwritten body based on convolutional neural networks is accelerated to know method for distinguishing |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109034281A true CN109034281A (en) | 2018-12-18 |
Family
ID=64643252
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810789695.4A Pending CN109034281A (en) | 2018-07-18 | 2018-07-18 | The Chinese handwritten body based on convolutional neural networks is accelerated to know method for distinguishing |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109034281A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109918499A (en) * | 2019-01-14 | 2019-06-21 | 平安科技(深圳)有限公司 | A kind of file classification method, device, computer equipment and storage medium |
CN109948742A (en) * | 2019-03-25 | 2019-06-28 | 西安电子科技大学 | Handwritten form picture classification method based on quantum nerve network |
CN109993162A (en) * | 2019-03-01 | 2019-07-09 | 昆明理工大学 | Laotian block letter text optical character recognition methods based on convolutional neural networks |
CN110567967A (en) * | 2019-08-20 | 2019-12-13 | 武汉精立电子技术有限公司 | Display panel detection method, system, terminal device and computer readable medium |
CN112308058A (en) * | 2020-10-25 | 2021-02-02 | 北京信息科技大学 | Method for recognizing handwritten characters |
CN112616043A (en) * | 2020-12-22 | 2021-04-06 | 杭州电子科技大学 | PYNQ-based neural network identification video monitoring alarm system and method |
CN112632319A (en) * | 2020-12-22 | 2021-04-09 | 天津大学 | Method for improving overall classification accuracy of long-tail distributed speech based on transfer learning |
CN114549906A (en) * | 2022-02-28 | 2022-05-27 | 长沙理工大学 | Improved image classification algorithm for step-by-step training of Top-k loss function |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108595643A (en) * | 2018-04-26 | 2018-09-28 | 重庆邮电大学 | Text character extraction and sorting technique based on more class node convolution loop networks |
-
2018
- 2018-07-18 CN CN201810789695.4A patent/CN109034281A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108595643A (en) * | 2018-04-26 | 2018-09-28 | 重庆邮电大学 | Text character extraction and sorting technique based on more class node convolution loop networks |
Non-Patent Citations (3)
Title |
---|
HONGWEI QIN 等: "Joint Training of Cascaded CNN for Face Detection", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
ZHIYUAN LI 等: "Building Efficient CNN Architecture for Offline Handwritten Chinese Character Recognition", 《ARXIV》 * |
何西麟: "基于深度学习的手写体字符识别研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109918499A (en) * | 2019-01-14 | 2019-06-21 | 平安科技(深圳)有限公司 | A kind of file classification method, device, computer equipment and storage medium |
CN109993162A (en) * | 2019-03-01 | 2019-07-09 | 昆明理工大学 | Laotian block letter text optical character recognition methods based on convolutional neural networks |
CN109948742A (en) * | 2019-03-25 | 2019-06-28 | 西安电子科技大学 | Handwritten form picture classification method based on quantum nerve network |
CN109948742B (en) * | 2019-03-25 | 2022-12-02 | 西安电子科技大学 | Handwritten picture classification method based on quantum neural network |
CN110567967A (en) * | 2019-08-20 | 2019-12-13 | 武汉精立电子技术有限公司 | Display panel detection method, system, terminal device and computer readable medium |
CN112308058A (en) * | 2020-10-25 | 2021-02-02 | 北京信息科技大学 | Method for recognizing handwritten characters |
CN112308058B (en) * | 2020-10-25 | 2023-10-24 | 北京信息科技大学 | Method for recognizing handwritten characters |
CN112616043A (en) * | 2020-12-22 | 2021-04-06 | 杭州电子科技大学 | PYNQ-based neural network identification video monitoring alarm system and method |
CN112632319A (en) * | 2020-12-22 | 2021-04-09 | 天津大学 | Method for improving overall classification accuracy of long-tail distributed speech based on transfer learning |
CN112632319B (en) * | 2020-12-22 | 2023-04-11 | 天津大学 | Method for improving overall classification accuracy of long-tail distributed speech based on transfer learning |
CN114549906A (en) * | 2022-02-28 | 2022-05-27 | 长沙理工大学 | Improved image classification algorithm for step-by-step training of Top-k loss function |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109034281A (en) | The Chinese handwritten body based on convolutional neural networks is accelerated to know method for distinguishing | |
CN103605972B (en) | Non-restricted environment face verification method based on block depth neural network | |
Thai et al. | Image classification using support vector machine and artificial neural network | |
Gando et al. | Fine-tuning deep convolutional neural networks for distinguishing illustrations from photographs | |
Owusu et al. | An SVM-AdaBoost facial expression recognition system | |
CN105184298B (en) | A kind of image classification method of quick local restriction low-rank coding | |
CN110866530A (en) | Character image recognition method and device and electronic equipment | |
Haghighi et al. | Stacking ensemble model of deep learning and its application to Persian/Arabic handwritten digits recognition | |
CN104834941A (en) | Offline handwriting recognition method of sparse autoencoder based on computer input | |
Yang et al. | An ensemble classification algorithm for convolutional neural network based on AdaBoost | |
CN108416270A (en) | A kind of traffic sign recognition method based on more attribute union features | |
CN105023006A (en) | Face recognition method based on enhanced nonparametric margin maximization criteria | |
CN109472733A (en) | Image latent writing analysis method based on convolutional neural networks | |
Almodfer et al. | Enhancing AlexNet for arabic handwritten words recognition using incremental dropout | |
CN110348280A (en) | Water book character recognition method based on CNN artificial neural | |
CN108154156A (en) | Image Ensemble classifier method and device based on neural topic model | |
Silanon | Thai Finger‐Spelling Recognition Using a Cascaded Classifier Based on Histogram of Orientation Gradient Features | |
CN110569971A (en) | convolutional neural network single-target identification method based on LeakyRelu activation function | |
Feng et al. | Robust shared feature learning for script and handwritten/machine-printed identification | |
Mahto et al. | Deep learning based models for offline Gurmukhi handwritten character and numeral recognition | |
Borra et al. | Face recognition based on convolutional neural network | |
Zhang et al. | A novel deep LeNet-5 convolutional neural network model for image recognition | |
ul Haq et al. | Neural network based adaboosting approach for hyperspectral data classification | |
CN110287973B (en) | Image feature extraction method based on low-rank robust linear discriminant analysis | |
Boveiri | Persian printed numerals classification using extended moment invariants |
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: 20181218 |
|
WD01 | Invention patent application deemed withdrawn after publication |