CN110163256A - Paper image based on joint probability matrix divides method from kinetonucleus - Google Patents
Paper image based on joint probability matrix divides method from kinetonucleus Download PDFInfo
- Publication number
- CN110163256A CN110163256A CN201910327743.2A CN201910327743A CN110163256A CN 110163256 A CN110163256 A CN 110163256A CN 201910327743 A CN201910327743 A CN 201910327743A CN 110163256 A CN110163256 A CN 110163256A
- Authority
- CN
- China
- Prior art keywords
- score
- joint probability
- result
- probability matrix
- kinetonucleus
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/14—Image acquisition
- G06V30/148—Segmentation of character regions
- G06V30/153—Segmentation of character regions using recognition of characters or words
-
- 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
- G06V30/28—Character recognition specially adapted to the type of the alphabet, e.g. Latin alphabet
- G06V30/287—Character recognition specially adapted to the type of the alphabet, e.g. Latin alphabet of Kanji, Hiragana or Katakana characters
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Present invention discloses a kind of paper images based on joint probability matrix to divide method from kinetonucleus, includes the following steps: S1, gets score identification model using training hand-written data training；S2, the rectangular area where paper score to be detected is cut using line detection algorithm；S3, using the result in S2 as the input of the score identification model, identify and export the highest N number of value of possibility；S4, building joint probability matrix simultaneously create mark search tree, calculate the value of the confidence；S5, calculated result is judged and is compared calculated result with preset threshold, final output score calculation result.The present invention calculates recognition confidence using joint probability matrix by the method for technology combination TensFlow and the CNN convolutional network of OCR, realizes the automatic checking statistics for paper total score.The present invention not only effectively improves the efficiency of core point operation, while also fully ensure that the accuracy of core point result.
Description
Technical field
Divide method from kinetonucleus based on OCR digital identification techniques the present invention relates to a kind of, and in particular to one kind is based on joint
The paper image of probability matrix divides method from kinetonucleus, i.e., is identified by the deep learning that joint probability calculation optimizes each score general
Rate achievees the purpose that paper score is effectively veritified, belongs to the field of image recognition in artificial intelligence.
Background technique
OCR (Optical Character Recognition, optical character identification) technology is a kind of by light such as scannings
It learns input mode and converts image information for the text of various bills, newpapers and periodicals, books, manuscript and other printed matters, recycle text
Word identification technology converts image information to the technology that computer can be used.
It after pre-processing image, needs to analyze the target in image, extracts and correctly represent difference
The characteristic parameter (characteristics of image) of object feature.After carrying out feature extraction to image, then of different images is carried out to target
Match.Finally object in image is identified and explained.Classifier is designed, disaggregated model is established, to object in image
It is identified and is classified.
With the rapid development of artificial intelligence, machine learning be increasingly becoming number identification main stream approach, in September, 2015,
Google has issued its second generation artificial intelligence system TensorFlow, an open source machine learning software resources storeroom.It is mesh
Preceding most popular machine learning algorithm frame is used by utilizing Google second generation artificial intelligence platform TensorFlow
The identification of CNN convolution combination Lenet5 neural network algorithm realization digital picture.
Also just because of the appearance of above-mentioned artificial intelligence system, strong technical support is provided for the realization of OCR technique.
In recent years, many insiders also start all many-sided trials in real life using OCR technique.For example,
In the treatment process at present for all kinds of papers, the verification of paper total score statistics usually require in a manner of manually-operated into
Row.It is predictable, total score is checked manually, not only inefficient operation but also is easily made a fault.Therefore, if
It can be by OCR technique applied to paper core point, then the human resources in school or organ will be liberated greatly.
In conclusion how to propose that a kind of paper image based on joint probability matrix is automatic on the basis of existing technology
Core divides method, to solve the problems, such as that current artificial nucleus cause inefficiency, accuracy not high score, also just becomes this field
The common goal in research of interior technical staff.
Summary of the invention
In view of the prior art, there are drawbacks described above, and the purpose of the present invention is to propose to a kind of papers based on joint probability matrix
Image divides method from kinetonucleus, includes the following steps:
S1, score identification model is got using training hand-written data training；
S2, the rectangular area where paper score to be detected is cut using line detection algorithm；
S3, using the result in S2 as the input of the score identification model, identify and to export possibility highest N number of
Value；
S4, building joint probability matrix simultaneously create mark search tree, calculate the value of the confidence；
S5, calculated result is judged and is compared calculated result with preset threshold, final output score calculates
As a result.
Preferably, S1 specifically comprises the following steps:
Score region on S11, foundation paper, prepares the training hand-written data collection for meeting all scores；
S12, training hand-written data collection described in S11 is trained using CNN convolutional network, obtains score identification mould
Type.
Preferably, S2 specifically comprises the following steps:
S21, the rectangular area where paper score to be detected is cut using line detection algorithm；
S22, border extended is carried out to the line detection algorithm, obtains the score area image of each major class.
Preferably, line detection algorithm described in S2, process include: according to the number of black pixel point in every n row, will be big
It is determined as lines in the pixel column of the optimal threshold, then re-optimization is carried out to lines, then adds one to every lines
Sash window with specific threshold cuts the live part of each pixel.
Preferably, S3 specifically comprises the following steps: to sequentially input each score area image obtained in S22
In the score identification model obtained in S12, by result arranged in sequence from high to low, each score region identified is provided
The N number of probable value of highest and confidence level.
Preferably, S4 specifically comprises the following steps:
S41, confidence level matrix is constructed using the N number of probable value of highest and confidence level in each score region obtained in S3；
S42, building state search tree algorithm, find from confidence level matrix obtained in S41 and meet small point of summation and be equal to
The combination of total score, and exported using all possible states and joint the value of the confidence as possible outcome, and execute S43；If not looking for
To combination of the small point of summation equal to total score is met, then S44 is directly executed；
S43, to all possible states obtained in S42 carry out confidence level sequence, and do Gauss gradient decline, differentiate obtain
Most possible assembled state, then executes S44；
S44, assembled state most possible obtained in possible state obtained in S42 or S43 is exported.
Preferably, state search tree algorithm is constructed described in S42, is specifically comprised the following steps: for excessively empty root node,
A search tree is set up, the first layer node of described search tree is first numerical value as a result, establishing under first layer node
Second layer node, the second layer node be second numerical value as a result, and so on, finally set up a mark search
Tree.
Preferably, S5 specifically comprises the following steps: to enter subsequent judgement if state exists, if there is no export state
Score calculates mistake；In the subsequent judgement, result is compared with preset threshold, the output point if result is greater than threshold value
Number calculates correctly, exports score if result is less than preset threshold and calculates mistake.
Compared with prior art, advantages of the present invention is mainly reflected in the following aspects:
The present invention is directed to the case where current artificial nucleus are to score, proposes a kind of core point method of automation.The present invention is logical
The method for crossing technology combination TensFlow and the CNN convolutional network of OCR calculates recognition confidence using joint probability matrix, real
The automatic checking statistics for paper total score is showed.The present invention not only effectively improves the efficiency of core point operation, shortens core
To the human resources in time needed for score, saving Liao Ge colleges and universities and organ, while it also fully ensure that core point process
Standardize, improve the accuracy of core point result.
Meanwhile the calculation method of score correctness is also provided for other relevant issues in same domain in the present invention
Reference, can carry out expansion extension on this basis, apply in other related art schemes in artificial intelligence field, have
There is very wide application prospect.
Just attached drawing in conjunction with the embodiments below, the embodiment of the present invention is described in further detail, so that of the invention
Technical solution is more readily understood, grasps.
Detailed description of the invention
Fig. 1 is the overall flow schematic diagram of the method for the present invention.
Fig. 2 is the local flow diagram of the method for the present invention.
Fig. 3 is the structural schematic diagram of mark search tree constructed in the present invention.
Specific embodiment
The present invention is directed to the case where artificial nucleus are to score in existing operation, and proposing one kind can be real by way of taking pictures
The now method that quickly extensive paper core divides.This method cuts picture by OpenCv, is then carried out by OCR technique
Picture pretreatment then realizes number identification by TensorFlow+CNN, calculates confidence level, substitution using joint probability matrix
Whether softmax function, analysis verification total score calculate correctly.
Specifically, present invention discloses a kind of paper images based on joint probability matrix certainly as shown in FIG. 1 to FIG. 2
Kinetonucleus divides method, includes the following steps:
S1, score identification model is got using training hand-written data training.
S1 is specifically included,
Score region on S11, foundation paper, prepares the training hand-written data collection for meeting all scores.
S12, training hand-written data collection described in S11 is trained using CNN convolutional network, obtains score identification mould
Type.
S2, the rectangular area where paper score to be detected is cut using line detection algorithm.
S2 is specifically included,
S21, the rectangular area where paper score to be detected is cut using line detection algorithm.
S22, border extended is carried out to the line detection algorithm, obtains the score area image of each major class.
The line detection algorithm, process include: that will be greater than described optimal according to the number of black pixel point in every n row
The pixel column of threshold value is determined as lines, then carries out re-optimization to lines, and then adding one to every lines has certain threshold
The sash window of value cuts the live part of each pixel.
S3, using the result in S2 as the input of the score identification model, identify and to export possibility highest N number of
Value.
S3 is specifically included, and each score area image obtained in S22 is sequentially input obtained in S12 described point
In number identification models, by result arranged in sequence from high to low, provide each score region identified the N number of probable value of highest and
Confidence level.
S4, building joint probability matrix simultaneously create mark search tree, calculate the value of the confidence.
S4 is specifically included,
S41, confidence level matrix is constructed using the N number of probable value of highest and confidence level in each score region obtained in S3.
S42, building state search tree algorithm, find from confidence level matrix obtained in S41 and meet small point of summation and be equal to
The combination of total score, and exported using all possible states and joint the value of the confidence as possible outcome, and execute S43.If not looking for
To combination of the small point of summation equal to total score is met, then S44 is directly executed.
The building state search tree algorithm, specifically includes, for excessively empty root node, it is established that a search tree, institute
The first layer node for stating search tree is first numerical value as a result, establish second layer node under first layer node, described second
Layer node be second numerical value as a result, and so on, finally set up a mark search tree.Final mark search tree graph
As shown in Figure 3.
S43, to all possible states obtained in S42 carry out confidence level sequence, and do Gauss gradient decline, differentiate obtain
Most possible assembled state, then executes S44.
S44, assembled state most possible obtained in possible state obtained in S42 or S43 is exported.
S5, calculated result is judged and is compared calculated result with preset threshold, final output score calculates
As a result.
S5 is specifically included, and enters subsequent judgement if state exists, if state calculates mistake there is no score is exported.?
In the subsequent judgement, result is compared with preset threshold, exports score if result is greater than threshold value and calculate correctly, if knot
Fruit is less than preset threshold and then exports score calculating mistake.
A concrete operations example is just combined below, and above-mentioned technical proposal is illustrated.
Training hand-written data collection described in S11 is trained using CNN convolutional network described in S12 of the present invention, is obtained
Score identification model, concrete operations process are as follows:
Step 1, building softmax function, in n classification, model has n output, i.e. y1, y2 ... yn, wherein yi table
Show a possibility that i-th kind of situation occurs size.WhereinThe output of CNN convolutional neural networks is passed through
It crosses softmax function and obtains the classification results of probability distribution.The class probability of output is distributed in model answer pair in a model
Than finding out cross entropy, obtaining loss function.
Picture is carried out single channel input by step 2, and the size of input is 28*28*1.
Step 3 carries out convolution, convolution kernel 5*5*1, number 32, step-length 1, full zero padding, at this time padding=
SAME enters length/step-length (rounding up), and the function for the convolution that tensflow is used at this time is tf.nn.conv2d, padding choosing
It is selected as SAME, specially tf.nn.conv2d (x, w, strides=[1,1,1,1], padding=' SAME '), to guarantee
Output is 5 × 5 resolution ratio, adds biasing to the output after convolution by tf.nn.bias_add (), and pass through tf.nn.relu
() completes nonlinear activation
Step 4 carries out pond, and pond size is 2*2, step-length 2, using the function in pond in Tensflow, wherein most
Great Chiization tf.nn.max_pool, average pondization tf.nn.avg_pool function, using full zero padding mode, padding
Select SAME, specially tf.nn.max_pool (x, ksize=[1,2,2,1], strides=[1,2,2,1], padding
=' SAME ').
Step 5 continues a convolution, nonlinear activation and pond.
It is specifically included in step 6, full articulamentum, in the forward propagation process, it is specified that network inputs node is 784 (generations
The number of pixels of every, table input picture), hide node layer 500, output node 100.By the ginseng of input layer to hidden layer
Number w1 shape is [784,500], is [500,100] by the parameter w2 shape of hidden layer to output layer, and parameter meets truncation normal state
Distribution, and regularization is used, the regularization loss of each parameter is added in total losses.By the biasing b1 of input layer to hidden layer
Shape is the one-dimension array that length is 500, is the one-dimension array that length is 10 by the biasing b2 shape of hidden layer to output layer, just
Beginning value is full 0.Propagated forward structure first layer is input x and parameter w1 matrix multiple plus biasing b1, using relu letter
Number obtains hidden layer output y1.The propagated forward structure second layer is that hidden layer exports y1 and parameter w2 matrix multiple plus biasing
B2 obtains output y.Reduce the instruction that over-fitting accelerates model using dropout when propagated forward constructs neural network simultaneously
Practice speed.
In back-propagation process, it is firstly introduced into tensorflow, input_data, propagated forward mnist_forward
With os module, picture number, initial learning rate, learning rate attenuation rate, the regularization coefficient, training of every wheel feeding neural network are defined
It takes turns number, model storing path and model and saves the relevant informations such as title.In backpropagation function backword, read in first
Mnist calls the propagated forward in mnist_forward file with placeholder to training data x and label y_ occupy-place
Process forword () function, and regularization is set, the prediction result y of training dataset is calculated, and take turns number meter to current calculate
Number device assignment, is set as that type can not be trained.Then, the loss function loss comprising all parameter regularizations loss is called, and
Indexing decaying learning rate learning_rate.Then, loss function is reduced to model optimization using gradient decay algorithm,
And the sliding average of defined parameters.Finally, realizing all parameter initializations in structure, feeding batch_size every time
Group (i.e. 200 groups) training data and corresponding label, loop iteration steps wheel, and a loss function is printed every 1000 wheels
Value information, and current sessions are loaded into specified path.Finally, loading the training under specified path by principal function main ()
Data set, and call defined backward () function training pattern.
Step 7 identifies the picture of each small topic and total score that obtain after cutting, deletes in lenet5 network
Softmax step records the highest N number of output valve of each small examination paper product value and confidence level, constructs confidence level matrix.In following reality
In example, by identifying the column fraction of certain paper, highest 3 output valves of every examination paper product value are obtained, tectonic syntaxis probability matrix is such as
Shown in table 1:
Table 1
Step 8: carrying out result using multiway tree and correctly determine, first of all for our traversal is facilitated, add a sky
Root node, first layer be followed successively by the first topic as a result, then the child node of each result is respectively the score of next topic, according to
It is secondary to analogize, it finally sets up a tree about score as shown in figure, then by the preorder traversal of number, successively traverses then
The results added of traversal is compared with total score, the correctness of judging result is recorded as long as there is correct result
Come, the joint probability matrix traversed in example is as shown in table 2:
Table 2
| Group number | First topic | Second topic | Third topic | Total score |
| 1 | (20,0.81) | (16,0.77) | (25,0.56) | (61,0.15) |
| 2 | (20,0.81) | (16,0.77) | (28,0.49) | (64,0.69) |
| 3 | (20,0.81) | (18,0.23) | (26,0.02) | (64,0.69) |
| 4 | (20,0.81) | (15,0.05) | (26,0.02) | (61,0.15) |
| 5 | (28,0.14) | (18,0.23) | (28,0.49) | (74,0.12) |
| 6 | (30,0.03) | (16,0.77) | (28,0.49) | (74,0.12) |
| 7 | (30,0.03) | (18,0.23) | (26,0.02) | (74,0.12) |
Step 9: joint probability being calculated for joint probability matrix table 2, is standardized, is i.e. calculating joint probability, is chosen
The value of the confidence is compared with selected threshold value, is only greater than after threshold value and just returns to knot by highest joint probability, that is, the value of the confidence
Fruit is correct.
The calculation method of joint probability are as follows: by after the probability multiplication of each small topic and total score, calculate their correspondence time
Evolution, obtained result maximum value are joint probability, that is, the value of the confidence.
Calculation formula are as follows:
Confidence level are as follows: α=Max (Pj)。
Each group joint probability matrix such as table 3 in example, obtained confidence level are 0.677649666, recognition result are as follows: topic
One: 20, topic 2 16, topic 3 28, total score 64.
Table 3
| Group number | Joint probability | Joint probability after standardization |
| 1 | 0.0523908 | 0.478424881 |
| 2 | 0.21087297 | 0.677649666 |
| 3 | 0.00257094 | 0.225176459 |
| 4 | 0.0001215 | 0.104989065 |
| 5 | 0.00189336 | 0.208597116 |
| 6 | 0.00135828 | 0.191976172 |
| 7 | 0.00001656 | 0.063791833 |
In conclusion the present invention utilizes joint by the method for technology combination TensFlow and the CNN convolutional network of OCR
Probability matrix calculates recognition confidence, realizes the automatic checking statistics for paper total score.The present invention not only effectively improves
The efficiency of core point operation, shorten verification score needed for the time, save human resources in Liao Ge colleges and universities and organ, simultaneously
Also the accuracy that fully ensure that the standardization of core point process, improve core point result.
Meanwhile the calculation method of score correctness is also provided for other relevant issues in same domain in the present invention
Reference, can carry out expansion extension on this basis, apply in other related art schemes in artificial intelligence field, have
There is very wide application prospect.
It is obvious to a person skilled in the art that invention is not limited to the details of the above exemplary embodiments, Er Qie
In the case where without departing substantially from spirit and essential characteristics of the invention, the present invention can be realized in other specific forms.Therefore, no matter
From the point of view of which point, the present embodiments are to be considered as illustrative and not restrictive, and the scope of the present invention is by appended power
Benefit requires rather than above description limits, it is intended that all by what is fallen within the meaning and scope of the equivalent elements of the claims
Variation is included within the present invention, and any reference signs in the claims should not be construed as limiting the involved claims.
In addition, it should be understood that although this specification is described in terms of embodiments, but not each embodiment is only wrapped
Containing an independent technical solution, this description of the specification is merely for the sake of clarity, and those skilled in the art should
It considers the specification as a whole, the technical solutions in the various embodiments may also be suitably combined, forms those skilled in the art
The other embodiments being understood that.
Claims (8)
1. a kind of paper image based on joint probability matrix divides method from kinetonucleus, which comprises the steps of:
S1, score identification model is got using training hand-written data training；
S2, the rectangular area where paper score to be detected is cut using line detection algorithm；
S3, using the result in S2 as the input of the score identification model, identify and export the highest N number of value of possibility；
S4, building joint probability matrix simultaneously create mark search tree, calculate the value of the confidence；
S5, calculated result is judged and is compared calculated result with preset threshold, final output score calculates knot
Fruit.
2. the paper image according to claim 1 based on joint probability matrix divides method from kinetonucleus, which is characterized in that S1
Specifically comprise the following steps:
Score region on S11, foundation paper, prepares the training hand-written data collection for meeting all scores；
S12, training hand-written data collection described in S11 is trained using CNN convolutional network, obtains score identification model.
3. the paper image according to claim 2 based on joint probability matrix divides method from kinetonucleus, which is characterized in that S2
Specifically comprise the following steps:
S21, the rectangular area where paper score to be detected is cut using line detection algorithm；
S22, border extended is carried out to the line detection algorithm, obtains the score area image of each major class.
4. the paper image according to claim 3 based on joint probability matrix divides method from kinetonucleus, which is characterized in that S2
Described in line detection algorithm, process includes: that will be greater than the optimal threshold according to the number of black pixel point in every n row
Pixel column is determined as lines, then carries out re-optimization to lines, and then adding one to every lines has the upper of specific threshold
Lower sliding window, the live part of each pixel is cut.
5. the paper image according to claim 3 based on joint probability matrix divides method from kinetonucleus, which is characterized in that S3
Specifically comprise the following steps: each score area image obtained in S22 sequentially inputting the score obtained in S12
In identification model, by result arranged in sequence from high to low, provides the N number of probable value of highest in each score region identified and set
Reliability.
6. the paper image according to claim 5 based on joint probability matrix divides method from kinetonucleus, which is characterized in that S4
Specifically comprise the following steps:
S41, confidence level matrix is constructed using the N number of probable value of highest and confidence level in each score region obtained in S3；
S42, building state search tree algorithm, find from confidence level matrix obtained in S41 and meet small point of summation equal to total score
Combination, and all possible states and joint the value of the confidence are exported as possible outcome, and execute S43；If not finding full
Small point of summation of foot is equal to the combination of total score, then directly executes S44；
S43, confidence level sequence is carried out to all possible states obtained in S42, and does the decline of Gauss gradient, differentiates most had
Possible assembled state, then executes S44；
S44, assembled state most possible obtained in possible state obtained in S42 or S43 is exported.
7. the paper image according to claim 6 based on joint probability matrix divides method from kinetonucleus, which is characterized in that
State search tree algorithm is constructed described in S42, is specifically comprised the following steps: for excessively empty root node, it is established that a search
Tree, the first layer node of described search tree is first numerical value as a result, establish second layer node under first layer node, described
Second layer node be second numerical value as a result, and so on, finally set up a mark search tree.
8. the paper image according to claim 6 based on joint probability matrix divides method from kinetonucleus, which is characterized in that S5
Specifically comprise the following steps: to enter subsequent judgement if state exists, if state calculates mistake there is no score is exported；Institute
It states in subsequent judgement, result is compared with preset threshold, export score if result is greater than threshold value and calculate correctly, if result
Score is then exported less than preset threshold calculates mistake.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910327743.2A CN110163256B (en) | 2019-04-23 | 2019-04-23 | Automatic examination paper image checking method based on joint probability matrix |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910327743.2A CN110163256B (en) | 2019-04-23 | 2019-04-23 | Automatic examination paper image checking method based on joint probability matrix |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110163256A true CN110163256A (en) | 2019-08-23 |
| CN110163256B CN110163256B (en) | 2022-11-15 |
Family
ID=67639749
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910327743.2A Active CN110163256B (en) | 2019-04-23 | 2019-04-23 | Automatic examination paper image checking method based on joint probability matrix |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110163256B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111242131A (en) * | 2020-01-06 | 2020-06-05 | 北京十六进制科技有限公司 | Method, storage medium and device for image recognition in intelligent marking |
| CN117671849A (en) * | 2023-12-14 | 2024-03-08 | 浙江南星科技有限公司 | Vertical image scanning banknote counter adopting banknote sliding structure |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180025256A1 (en) * | 2015-10-20 | 2018-01-25 | Tencent Technology (Shenzhen) Company Limited | Method and apparatus for recognizing character string in image |
| CN108805222A (en) * | 2018-05-08 | 2018-11-13 | 南京邮电大学 | A kind of deep learning digital handwriting body recognition methods based on ARM platforms |
| CN109389173A (en) * | 2018-10-13 | 2019-02-26 | 河南理工大学 | A kind of paper score automatic statistical analysis method and apparatus based on M-CNN |
-
2019
- 2019-04-23 CN CN201910327743.2A patent/CN110163256B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180025256A1 (en) * | 2015-10-20 | 2018-01-25 | Tencent Technology (Shenzhen) Company Limited | Method and apparatus for recognizing character string in image |
| CN108805222A (en) * | 2018-05-08 | 2018-11-13 | 南京邮电大学 | A kind of deep learning digital handwriting body recognition methods based on ARM platforms |
| CN109389173A (en) * | 2018-10-13 | 2019-02-26 | 河南理工大学 | A kind of paper score automatic statistical analysis method and apparatus based on M-CNN |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111242131A (en) * | 2020-01-06 | 2020-06-05 | 北京十六进制科技有限公司 | Method, storage medium and device for image recognition in intelligent marking |
| CN111242131B (en) * | 2020-01-06 | 2024-05-10 | 北京十六进制科技有限公司 | Method, storage medium and device for identifying images in intelligent paper reading |
| CN117671849A (en) * | 2023-12-14 | 2024-03-08 | 浙江南星科技有限公司 | Vertical image scanning banknote counter adopting banknote sliding structure |
| CN117671849B (en) * | 2023-12-14 | 2024-05-14 | 浙江南星科技有限公司 | Vertical image scanning banknote counter adopting banknote sliding structure |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110163256B (en) | 2022-11-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111489358B (en) | Three-dimensional point cloud semantic segmentation method based on deep learning | |
| CN110929610B (en) | Plant disease identification method and system based on CNN model and transfer learning | |
| CN110555399B (en) | Finger vein identification method and device, computer equipment and readable storage medium | |
| CN109597891A (en) | Text emotion analysis method based on two-way length Memory Neural Networks in short-term | |
| CN109543031A (en) | A kind of file classification method based on multitask confrontation study | |
| CN111046661B (en) | Reading understanding method based on graph convolution network | |
| CN111652202B (en) | Method and system for solving video question-answer problem by improving video-language representation learning through self-adaptive space-time diagram model | |
| CN110222163A (en) | A kind of intelligent answer method and system merging CNN and two-way LSTM | |
| CN109977199A (en) | A kind of reading understanding method based on attention pond mechanism | |
| CN111881262A (en) | Text emotion analysis method based on multi-channel neural network | |
| CN115858847B (en) | Combined query image retrieval method based on cross-modal attention reservation | |
| CN110096991A (en) | A kind of sign Language Recognition Method based on convolutional neural networks | |
| CN115563327A (en) | Zero sample cross-modal retrieval method based on Transformer network selective distillation | |
| CN111159345A (en) | Chinese knowledge base answer obtaining method and device | |
| CN110163256A (en) | Paper image based on joint probability matrix divides method from kinetonucleus | |
| CN108985442A (en) | Handwriting model training method, hand-written character recognizing method, device, equipment and medium | |
| CN115937852A (en) | Text-driven efficient weak supervision semantic segmentation method and device | |
| CN112766134B (en) | Expression recognition method for strengthening distinction between classes | |
| CN111783688B (en) | Remote sensing image scene classification method based on convolutional neural network | |
| CN116386148B (en) | Knowledge graph guide-based small sample action recognition method and system | |
| CN113239678A (en) | Multi-angle attention feature matching method and system for answer selection | |
| CN116258147A (en) | Multimode comment emotion analysis method and system based on heterogram convolution | |
| CN115375984A (en) | Chart question-answering method based on graph neural network | |
| Chen et al. | Towards Pedestrian Target Detection with Optimized Mask R‐CNN | |
| CN115168678A (en) | Time sequence perception heterogeneous graph nerve rumor detection model |
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 |