CN109886105A - Price tickets recognition methods, system and storage medium based on multi-task learning - Google Patents
Price tickets recognition methods, system and storage medium based on multi-task learning Download PDFInfo
- Publication number
- CN109886105A CN109886105A CN201910033930.XA CN201910033930A CN109886105A CN 109886105 A CN109886105 A CN 109886105A CN 201910033930 A CN201910033930 A CN 201910033930A CN 109886105 A CN109886105 A CN 109886105A
- Authority
- CN
- China
- Prior art keywords
- price
- price tickets
- tickets
- obtains
- processing
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses price tickets recognition methods, system and storage medium based on multi-task learning, method includes identification model training step and price tickets identification step.The present invention by price tickets integer part and fractional part be trained and identify respectively, not only reduce the difficulty of neural metwork training, and the position of decimal point can be distinguished, accuracy of identification is improved, can be widely applied to depth learning technology field.
Description
Technical field
The present invention relates to depth learning technology fields, are based especially on price tickets recognition methods, the system of multi-task learning
And storage medium.
Background technique
Recently as the rapid development of depth learning technology, it is also more and more extensive for applying, and is led from traditional security protection
Domain, the wisdom retail domain to rising in recent years have its figure.It is the important link in being newly sold that disappears fastly that channel, which is verified,
Traditional operation mode mainly includes that business agent's on-site verification and third party verify, and it is respective scarce that both modes suffer from its
Point, such as human error is big, the verification period is long, data error can not trace.Image recognition technology based on deep learning has
The features such as precision is high, scene can restore, be very suitable to the business scenario of channel verification.Channel based on image verifies main packet
Containing two big identification contents, SKU identification and price tickets identification.Important component of the price as sales data, based on image
The required precision of price tickets identification is very high, and price tickets identification is easy to be influenced by price tickets design pattern, quality of taking pictures etc., such as
The influence of fuzzy, very oblique etc. factors.
Existing price tickets identification technology be all largely by the various pieces of price tickets, as integer part, fractional part with
And decimal point is as a whole, then unifies to carry out indiscriminate identification, but due to many kinds of, the Yi Jirong of price tickets
Vulnerable to illumination, the influence of photo angle, feature in the picture is not that clearly, identification difficulty is high, even with
Recognition sequence algorithm with context relation, it is also difficult to position the position of decimal point, therefore accuracy is not high.
Summary of the invention
In order to solve the above technical problems, it is an object of the invention to: provide that a kind of difficulty is low and accuracy is high, based on more
Price tickets recognition methods, system and the storage medium of tasking learning.
The technical solution that one aspect of the present invention is taken are as follows:
Price tickets recognition methods based on multi-task learning, including identification model training step and price tickets identification step,
Wherein,
The identification model training step the following steps are included:
According to the image data of price tickets, the pricing information on price tickets is detected；
According to the pricing information on price tickets, based on preset data format to the integer part and fractional part of pricing information
Divide and be labeled, obtains price data；
Enhancing processing is carried out to the price data marked；
Enhancing treated price data is inputted preset neural network model to be trained, obtains identification model；
The price tickets identification step the following steps are included:
The price tickets position in images to be recognized is detected by target detection model, obtains price tickets picture number
According to；
Price tickets image data is pre-processed；
By identification model, pretreated price tickets image data is identified, obtains the integer portion of pricing information
Divide recognition result and fractional part recognition result.
Further, the identification model training step, further comprising the steps of:
Acquire shelf image；
Collected shelf image is detected, confirms the position in price tickets region in shelf image；
According to the position in price tickets region, the image data of price tickets is intercepted.
Further, the described pair of price data marked carried out in the step for enhancing processing, and the enhancing processing includes
Brightness processed, rotation processing, scaling processing, translation processing, increases noise processed, skimulated motion Fuzzy Processing at contrast processing
With simulation ambiguity of space angle processing.
Further, described enhancing treated price data is inputted into preset neural network model to be trained, it obtains
The step for identification model, comprising the following steps:
By convolutional neural networks and LSTM network, to enhancing, treated that price data calculates, and obtains the whole of price
Number part and fractional part；
The feature vector of integer part and fractional part is calculated separately by LSTM network；
Processing is optimized by the feature vector that Attention mechanism obtains LSTM network query function；
According to the feature vector after optimization processing, the loss function of integer part and fractional part is calculated separately；
According to the loss function being calculated, identification model is obtained.
It is further, described that by convolutional neural networks and LSTM network, to enhancing, treated that price data calculates,
The step for obtaining the integer part and fractional part of price, comprising the following steps:
To enhancing, treated that price data carries out normalization processing, obtains picture to be trained；
The characteristic pattern of picture to be trained is extracted by convolutional neural networks；
Processing is reconstructed to characteristic pattern；
By LSTM network, to reconstruct, treated that characteristic pattern calculates, and obtains the integer part and fractional part of price
Point.
Further, the step for the feature vector that integer part and fractional part are calculated separately by LSTM network,
The following steps are included:
The characteristic pattern of the characteristic pattern of integer part and fractional part is inputted into several LSTM networks respectively；
Pass through the timing information between LSTM network query function feature；
According to the timing information being calculated, the characteristic pattern that can express timing information is generated.
Further, the feature vector obtained by Attention mechanism to LSTM network query function optimizes processing
The step for, comprising the following steps:
Notice that force parameter is normalized by context of the softmax to Attention mechanism, obtains weight ginseng
Number；
Be weighted processing by the feature vector that weight parameter obtains LSTM network query function, obtain new feature to
Amount.
Another aspect of the present invention is adopted the technical scheme that:
Price tickets identifying system based on multi-task learning, comprising: training module and identification module, wherein
The training module includes:
First detection unit detects the pricing information on price tickets for the image data according to price tickets；
Unit is marked, for according to the pricing information on price tickets, based on preset data format to the whole of pricing information
Number part and fractional part are labeled, and obtain price data；
Enhancement unit, for carrying out enhancing processing to the price data marked；
Training unit inputs preset neural network model for the price data that will enhance that treated and is trained, obtains
To identification model；
The identification module includes:
Second detection unit, for being detected by target detection model to the price tickets position in images to be recognized,
Obtain price tickets image data；
Pretreatment unit, for being pre-processed to price tickets image data；
Recognition unit, for being identified to pretreated price tickets image data, obtaining price by identification model
The integer part recognition result and fractional part recognition result of information.
Another aspect of the present invention is adopted the technical scheme that:
Price tickets identifying system based on multi-task learning, comprising:
At least one processor；
At least one processor, for storing at least one program；
When at least one described program is executed by least one described processor, so that at least one described processor is realized
The price tickets recognition methods based on multi-task learning.
Another aspect of the present invention is adopted the technical scheme that:
A kind of storage medium, wherein be stored with the executable instruction of processor, the executable instruction of the processor by
For executing the price tickets recognition methods based on multi-task learning when processor executes.
The beneficial effects of the present invention are: the present invention by price tickets integer part and fractional part be trained respectively with
And identification, the difficulty of neural metwork training is not only reduced, and the position of decimal point can be distinguished, improves identification
Precision.
Detailed description of the invention
Fig. 1 is the step flow chart of the embodiment of the present invention；
Fig. 2 is the schematic network structure of the embodiment of the present invention.
Specific embodiment
The present invention is further explained and is illustrated with specific embodiment with reference to the accompanying drawings of the specification.For of the invention real
The step number in example is applied, is arranged only for the purposes of illustrating explanation, any restriction is not done to the sequence between step, is implemented
The execution sequence of each step in example can be adaptively adjusted according to the understanding of those skilled in the art.
The price tickets recognition methods based on multi-task learning that the embodiment of the invention provides a kind of, including identification model training
Step and price tickets identification step, wherein
The identification model training step the following steps are included:
According to the image data of price tickets, the pricing information on price tickets is detected；
According to the pricing information on price tickets, based on preset data format to the integer part and fractional part of pricing information
Divide and be labeled, obtains price data；
Enhancing processing is carried out to the price data marked；
Enhancing treated price data is inputted preset neural network model to be trained, obtains identification model；
The price tickets identification step the following steps are included:
The price tickets position in images to be recognized is detected by target detection model, obtains price tickets picture number
According to；
Price tickets image data is pre-processed；
By identification model, pretreated price tickets image data is identified, obtains the integer portion of pricing information
Divide recognition result and fractional part recognition result.
It is further used as preferred embodiment, the identification model training step is further comprising the steps of:
Acquire shelf image；
Collected shelf image is detected, confirms the position in price tickets region in shelf image；
According to the position in price tickets region, the image data of price tickets is intercepted.
It is further used as preferred embodiment, the described pair of price data marked carries out the step for enhancing is handled
In, the enhancing processing includes brightness processed, contrast processing, rotation processing, scaling processing, translation processing, increases at noise
Reason, skimulated motion Fuzzy Processing and simulation ambiguity of space angle processing.
Be further used as preferred embodiment, it is described will enhancing treated that price data inputs preset neural network
The step for model is trained, and obtains identification model, comprising the following steps:
By convolutional neural networks and LSTM network, to enhancing, treated that price data calculates, and obtains the whole of price
Number part and fractional part；
The feature vector of integer part and fractional part is calculated separately by LSTM network；
Processing is optimized by the feature vector that Attention mechanism obtains LSTM network query function；
According to the feature vector after optimization processing, the loss function of integer part and fractional part is calculated separately；
According to the loss function being calculated, identification model is obtained.
It is further used as preferred embodiment, described treated to enhancing by convolutional neural networks and LSTM network
The step for price data is calculated, and the integer part and fractional part of price are obtained, comprising the following steps:
To enhancing, treated that price data carries out normalization processing, obtains picture to be trained；
The characteristic pattern of picture to be trained is extracted by convolutional neural networks；
Processing is reconstructed to characteristic pattern；
By LSTM network, to reconstruct, treated that characteristic pattern calculates, and obtains the integer part and fractional part of price
Point.
It is further used as preferred embodiment, it is described that integer part and fractional part are calculated separately by LSTM network
The step for feature vector, comprising the following steps:
The characteristic pattern of the characteristic pattern of integer part and fractional part is inputted into several LSTM networks respectively；
Pass through the timing information between LSTM network query function feature；
According to the timing information being calculated, the characteristic pattern that can express timing information is generated.
It is further used as preferred embodiment, the spy obtained by Attention mechanism to LSTM network query function
The step for sign vector optimizes processing, comprising the following steps:
Notice that force parameter is normalized by context of the softmax to Attention mechanism, obtains weight ginseng
Number；
Be weighted processing by the feature vector that weight parameter obtains LSTM network query function, obtain new feature to
Amount.
It is corresponding with method, the price tickets identifying system based on multi-task learning that the embodiment of the invention also provides a kind of,
It include: training module and identification module, wherein
The training module includes:
First detection unit detects the pricing information on price tickets for the image data according to price tickets；
Unit is marked, for according to the pricing information on price tickets, based on preset data format to the whole of pricing information
Number part and fractional part are labeled, and obtain price data；
Enhancement unit, for carrying out enhancing processing to the price data marked；
Training unit inputs preset neural network model for the price data that will enhance that treated and is trained, obtains
To identification model；
The identification module includes:
Second detection unit, for being detected by target detection model to the price tickets position in images to be recognized,
Obtain price tickets image data；
Pretreatment unit, for being pre-processed to price tickets image data；
Recognition unit, for being identified to pretreated price tickets image data, obtaining price by identification model
The integer part recognition result and fractional part recognition result of information.
It is corresponding with method, the price tickets identifying system based on multi-task learning that the embodiment of the invention also provides a kind of,
Include:
At least one processor；
At least one processor, for storing at least one program；
When at least one described program is executed by least one described processor, so that at least one described processor is realized
The price tickets recognition methods based on multi-task learning.
Corresponding with method, the embodiment of the invention also provides a kind of storage mediums, wherein it is executable to be stored with processor
Instruction, the executable instruction of the processor is when executed by the processor for executing the valence based on multi-task learning
Lattice board recognition methods.
With reference to the accompanying drawings of the specification 1, the tool of price tickets recognition methods of the present invention is described in detail based on multi-task learning
Body implementation steps:
(1), price tickets data are acquired, the price tickets data in the present invention are mainly derived from true shelf image, data
It is rich and varied, cover various design patterns, angle change, illumination variation；
(2), shelf image collected is detected, to find out the position in price tickets region；
(3), training data marks, according to the pricing information on the price tickets detected, the artificial integer for marking price tickets
And fractional part supplies number, such as price 100.05 according to 9 digits in mark, and in mark, integer and fractional part
It is respectively labeled as 100AAAAAA, 05AAAAAAA；
(4), processing of the present embodiment Jing Guo above step, price tickets data bulk about 4w generated, in order to
The diversity of data is further increased, the present invention has carried out enhancing processing to above-mentioned truthful data, mainly included, brightness, comparison
Degree, scaling, translation, increases several aspects such as noise, fuzzy, the simulation ambiguity of space angle of skimulated motion at rotation；
(5), data are sent into pre-designed neural network model to be trained, training pattern is until model is restrained；
(6), it in the actual use stage, first by target detection model, detects the position of price tickets and intercepts picture, in advance
Trained model, the integer and fractional part of export price board, then removal filling respectively in input step (5) after processing
Character finally combines the pricing information on export price board.
Referring to network structure shown in Fig. 2, specifically, the step (5) the following steps are included:
S1, the picture that training data is normalized to 96 × 200 first, utilize convolutional neural networks to extract the spy of picture
Sign obtains the characteristic pattern of picture, and the specification of characteristic pattern is 12 × 25 × 64, then by characteristic pattern reshape to 25 × 768, and with
This is characterized input LSTM network, calculates integer and fractional part in price tickets picture；
S2,25 LSTM nets are sequentially input by the characteristic pattern of the extraction of upper level network for the integer part in picture
Network, the input of each network are the feature vector of 1 × 768 dimension；By 25 feature vectors by LSTM network respectively from a left side to
It right connection and connects from right to left, calculates the timing information between feature, finally obtain 25 × 768 of expressed sequence information after energy
Characteristic pattern；
S3, the feature vector obtained by LSTM network query function, the problem is that: no matter list entries length all can be by
The vector for being encoded into a regular length indicates, and decodes the vector expression for being then limited to the regular length, limits model
Performance；For more accurate expression characteristic vector, invention introduces Attention mechanism, fundamental design idea is
By reservation LSTM encoder to the intermediate output of list entries as a result, training a model then to select these inputs
The study of selecting property and output sequence is associated therewith in model output, Attention mechanism proposed by the present invention
It is implemented as follows:
Firstly, input feature value c:c={ c1,c2,......,cL, L=25, c thereiniIt indicates to pass through LSTM network
Some the spatial position feature being calculated；The length of L expression input feature vector sequence.
Then, the context for Attention mechanism being arranged pays attention to force parameter e:
ei=fATT(h,ci)
Wherein, fATT() representation remaps function；The function f of the present embodimentATTIt is realized using multitiered network.ciRepresent
The feature vector in i channel；H represents the hidden state parameter of multitiered network.
Then, it is normalized using softmax, obtains weight parameter a:
The present embodiment uses the feature c obtained after Attention mechanismtIt can indicate are as follows:
Finally export the feature vector c obtained after Attention mechanismt。
S4,1 × 768 feature vector c has been obtained after learning by Attention mechanismt, to each LSTM network list
The output of member is weighted, and calculates separately the feature vector of 9 bit digital of integer part, is calculated and is classified further according to softmax loss
Loss；
Wherein, the network losses function L of integer partintAre as follows:
zi,j=ωi,jx+bi,j
Wherein, M=9 indicates the loss of 9 bit digitals of integer part output respectively；N indicates to participate in the number of samples of training；
yi,jIndicate target category；si,jIndicate the softmax output of j-th of integer part, i-th of training sample position；It indicates
zI, jIndex mapping；Indicate zi,kIndex mapping；zi,jIndicate the response magnitude of neuron；K=11 presentation class
Classification number；ωi,jAnd bi,jIndicate network parameter；X is the feature vector of network.
S5, the algorithmic procedure of fractional part are identical as integer part, loss function LdecCalculation method also with integer part
It is identical, it may be assumed that
The present invention calculates the loss function of integer part and fractional part in network training, is then calculated
Total loss function L are as follows:
L=Lint+Ldec；
Finally, the present invention carries out dynamic update by reverse propagated error, to network parameter, until completing when network convergence
Training, obtains final identification model.
In conclusion the problem of present invention is difficult to differentiate between with regard to scaling position in existing price tickets identification technology, proposing will
Integer part and fractional part in price tickets separately identify that main process is divided into following several stages to solve the problems, such as this, the
One stage was to carry out CNN feature extraction to the image of input；Second stage be using LSTM to the feature in a upper stage into
The expression of row sequence；Three phases are the interference that background is eliminated using attention mechanism；The last stage is by integer part
Multi-task learning modeling is carried out with fractional part.Data in price tickets are divided into two parts and not identified by present invention proposition, i.e., whole
Several and decimal separately identifies, and is learnt using multitask Training strategy end to end, not only reduces identification in this way
Difficulty can also distinguish the position of decimal point.
It is to be illustrated to preferable implementation of the invention, but the present invention is not limited to the embodiment above, it is ripe
Various equivalent deformation or replacement can also be made on the premise of without prejudice to spirit of the invention by knowing those skilled in the art, this
Equivalent deformation or replacement are all included in the scope defined by the claims of the present application a bit.
Claims (10)
1. the price tickets recognition methods based on multi-task learning, it is characterised in that: including identification model training step and price tickets
Identification step, wherein
The identification model training step the following steps are included:
According to the image data of price tickets, the pricing information on price tickets is detected；
According to the pricing information on price tickets, based on preset data format to the integer part of pricing information and fractional part into
Rower note, obtains price data；
Enhancing processing is carried out to the price data marked；
Enhancing treated price data is inputted preset neural network model to be trained, obtains identification model；
The price tickets identification step the following steps are included:
The price tickets position in images to be recognized is detected by target detection model, obtains price tickets image data；
Price tickets image data is pre-processed；
By identification model, pretreated price tickets image data is identified, the integer part for obtaining pricing information is known
Other result and fractional part recognition result.
2. the price tickets recognition methods according to claim 1 based on multi-task learning, it is characterised in that: the identification mould
Type training step, further comprising the steps of:
Acquire shelf image；
Collected shelf image is detected, confirms the position in price tickets region in shelf image；
According to the position in price tickets region, the image data of price tickets is intercepted.
3. the price tickets recognition methods according to claim 1 based on multi-task learning, it is characterised in that: described pair of mark
Good price data carried out in the step for enhancing processing, and the enhancing is handled including at brightness processed, contrast processing, rotation
Reason, translation processing, increases noise processed, skimulated motion Fuzzy Processing and simulation ambiguity of space angle processing at scaling processing.
4. the price tickets recognition methods according to claim 1 based on multi-task learning, it is characterised in that: described to enhance
Treated, and price data inputs the step for preset neural network model is trained, obtains identification model, including following
Step:
By convolutional neural networks and LSTM network, to enhancing, treated that price data calculates, and obtains the integer portion of price
Point and fractional part；
The feature vector of integer part and fractional part is calculated separately by LSTM network；
Processing is optimized by the feature vector that Attention mechanism obtains LSTM network query function；
According to the feature vector after optimization processing, the loss function of integer part and fractional part is calculated separately；
According to the loss function being calculated, identification model is obtained.
5. the price tickets recognition methods according to claim 4 based on multi-task learning, it is characterised in that: described to pass through volume
Treated that price data calculates to enhancing for product neural network and LSTM network, obtains the integer part and fractional part of price
The step for dividing, comprising the following steps:
To enhancing, treated that price data carries out normalization processing, obtains picture to be trained；
The characteristic pattern of picture to be trained is extracted by convolutional neural networks；
Processing is reconstructed to characteristic pattern；
By LSTM network, to reconstruct, treated that characteristic pattern calculates, and obtains the integer part and fractional part of price.
6. the price tickets recognition methods according to claim 4 based on multi-task learning, it is characterised in that: described to pass through
LSTM network calculates separately the step for feature vector of integer part and fractional part, comprising the following steps:
The characteristic pattern of the characteristic pattern of integer part and fractional part is inputted into several LSTM networks respectively；
Pass through the timing information between LSTM network query function feature；
According to the timing information being calculated, the characteristic pattern that can express timing information is generated.
7. the price tickets recognition methods according to claim 4 based on multi-task learning, it is characterised in that: described to pass through
The step for feature vector that Attention mechanism obtains LSTM network query function optimizes processing, comprising the following steps:
Notice that force parameter is normalized by context of the softmax to Attention mechanism, obtains weight parameter；
It is weighted processing by the feature vector that weight parameter obtains LSTM network query function, obtains new feature vector.
8. the price tickets identifying system based on multi-task learning, it is characterised in that: include: training module and identification module, wherein
The training module includes:
First detection unit detects the pricing information on price tickets for the image data according to price tickets；
Unit is marked, for according to the pricing information on price tickets, based on preset data format to the integer portion of pricing information
Divide and fractional part is labeled, obtains price data；
Enhancement unit, for carrying out enhancing processing to the price data marked；
Training unit inputs preset neural network model for the price data that will enhance that treated and is trained, known
Other model；
The identification module includes:
Second detection unit is obtained for being detected by target detection model to the price tickets position in images to be recognized
Price tickets image data；
Pretreatment unit, for being pre-processed to price tickets image data；
Recognition unit, for being identified to pretreated price tickets image data, obtaining pricing information by identification model
Integer part recognition result and fractional part recognition result.
9. the price tickets identifying system based on multi-task learning, it is characterised in that: include:
At least one processor；
At least one processor, for storing at least one program；
When at least one described program is executed by least one described processor, so that at least one described processor is realized as weighed
Benefit requires the price tickets recognition methods described in any one of 1-7 based on multi-task learning.
10. a kind of storage medium, wherein being stored with the executable instruction of processor, it is characterised in that: the processor is executable
Instruction when executed by the processor for executes such as the price of any of claims 1-7 based on multi-task learning
Board recognition methods.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910033930.XA CN109886105B (en) | 2019-01-15 | 2019-01-15 | Price tag identification method, system and storage medium based on multi-task learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910033930.XA CN109886105B (en) | 2019-01-15 | 2019-01-15 | Price tag identification method, system and storage medium based on multi-task learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109886105A true CN109886105A (en) | 2019-06-14 |
| CN109886105B CN109886105B (en) | 2021-12-14 |
Family
ID=66925927
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910033930.XA Active CN109886105B (en) | 2019-01-15 | 2019-01-15 | Price tag identification method, system and storage medium based on multi-task learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109886105B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110766050A (en) * | 2019-09-19 | 2020-02-07 | 北京捷通华声科技股份有限公司 | Model generation method, text recognition method, device, equipment and storage medium |
| CN113536989A (en) * | 2021-06-29 | 2021-10-22 | 广州博通信息技术有限公司 | Refrigerator frosting monitoring method and system based on camera video frame-by-frame analysis |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH0944601A (en) * | 1995-07-28 | 1997-02-14 | Chubu Electric Power Co Inc | Image recognizing device |
| CN107045641A (en) * | 2017-04-26 | 2017-08-15 | 广州图匠数据科技有限公司 | A kind of identification of pallets method based on image recognition technology |
| CN108399419A (en) * | 2018-01-25 | 2018-08-14 | 华南理工大学 | Chinese text recognition methods in natural scene image based on two-dimentional Recursive Networks |
| CN108647553A (en) * | 2018-05-10 | 2018-10-12 | 上海扩博智能技术有限公司 | Rapid expansion method, system, equipment and the storage medium of model training image |
| CN108875836A (en) * | 2018-06-27 | 2018-11-23 | 浙江大学 | A kind of simple-complicated activity collaboration recognition methods based on depth multi-task learning |
-
2019
- 2019-01-15 CN CN201910033930.XA patent/CN109886105B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH0944601A (en) * | 1995-07-28 | 1997-02-14 | Chubu Electric Power Co Inc | Image recognizing device |
| CN107045641A (en) * | 2017-04-26 | 2017-08-15 | 广州图匠数据科技有限公司 | A kind of identification of pallets method based on image recognition technology |
| CN108399419A (en) * | 2018-01-25 | 2018-08-14 | 华南理工大学 | Chinese text recognition methods in natural scene image based on two-dimentional Recursive Networks |
| CN108647553A (en) * | 2018-05-10 | 2018-10-12 | 上海扩博智能技术有限公司 | Rapid expansion method, system, equipment and the storage medium of model training image |
| CN108875836A (en) * | 2018-06-27 | 2018-11-23 | 浙江大学 | A kind of simple-complicated activity collaboration recognition methods based on depth multi-task learning |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110766050A (en) * | 2019-09-19 | 2020-02-07 | 北京捷通华声科技股份有限公司 | Model generation method, text recognition method, device, equipment and storage medium |
| CN113536989A (en) * | 2021-06-29 | 2021-10-22 | 广州博通信息技术有限公司 | Refrigerator frosting monitoring method and system based on camera video frame-by-frame analysis |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109886105B (en) | 2021-12-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107945204B (en) | Pixel-level image matting method based on generation countermeasure network | |
| CN110008338B (en) | E-commerce evaluation emotion analysis method integrating GAN and transfer learning | |
| CN111444343B (en) | Cross-border national culture text classification method based on knowledge representation | |
| CN110533024B (en) | Double-quadratic pooling fine-grained image classification method based on multi-scale ROI (region of interest) features | |
| CN106960206A (en) | Character identifying method and character recognition system | |
| CN107680579A (en) | Text regularization model training method and device, text regularization method and device | |
| CN105740790B (en) | Color face recognition method based on multicore dictionary learning | |
| CN110175248B (en) | Face image retrieval method and device based on deep learning and Hash coding | |
| CN109102024A (en) | A kind of Layer semantics incorporation model finely identified for object and its implementation | |
| CN110349229A (en) | A kind of Image Description Methods and device | |
| CN107491729B (en) | Handwritten digit recognition method based on cosine similarity activated convolutional neural network | |
| CN112651940B (en) | Collaborative visual saliency detection method based on dual-encoder generation type countermeasure network | |
| CN108229287A (en) | Image-recognizing method and device, electronic equipment and computer storage media | |
| CN113361251A (en) | Text image generation method and system based on multi-stage generation countermeasure network | |
| CN111126464A (en) | Image classification method based on unsupervised domain confrontation field adaptation | |
| Yu et al. | Unbiased multi-modality guidance for image inpainting | |
| CN113378949A (en) | Dual-generation confrontation learning method based on capsule network and mixed attention | |
| CN115222998A (en) | Image classification method | |
| CN109886105A (en) | Price tickets recognition methods, system and storage medium based on multi-task learning | |
| CN116258990A (en) | Cross-modal affinity-based small sample reference video target segmentation method | |
| CN111522979A (en) | Picture sorting recommendation method and device, electronic equipment and storage medium | |
| CN108108769A (en) | A kind of sorting technique of data, device and storage medium | |
| CN113420289B (en) | Hidden poisoning attack defense method and device for deep learning model | |
| Yao et al. | [Retracted] Facial Expression Recognition Based on Convolutional Neural Network Fusion SIFT Features of Mobile Virtual Reality | |
| Roy et al. | Tips: Text-induced pose synthesis |
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 |