CN110059694A - The intelligent identification Method of lteral data under power industry complex scene - Google Patents
The intelligent identification Method of lteral data under power industry complex scene Download PDFInfo
- Publication number
- CN110059694A CN110059694A CN201910318881.4A CN201910318881A CN110059694A CN 110059694 A CN110059694 A CN 110059694A CN 201910318881 A CN201910318881 A CN 201910318881A CN 110059694 A CN110059694 A CN 110059694A
- Authority
- CN
- China
- Prior art keywords
- text
- picture
- data
- feature
- model
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the intelligent identification Methods of the lteral data under power industry complex scene.This method uses series of preprocessing before text detection, enhance the lteral data information taken in frame picture, irrelevant information is weakened, the CBOW model for carrying out post-processing and using natural language processing field after recognition is post-processed, and the effect of detection and identification is promoted with this.Text data verification and measurement ratio is 96%, and recognition accuracy is 85% or so, accuracy rate with higher, while having faster operation speed on local server, can satisfy the requirement of real-time.This method is applicable not only to distribution room environmental simultaneously, applies also for the various complex scenes even natural scene of electric utility, is widely used.
Description
Technical field
The present invention relates to the intelligent identification Methods of the lteral data under power industry complex scene, and it is artificial to belong to power industry
The technical field of intelligent recognition.
Background technique
Electric power resource is basis and the lifeblood of national economy, and distribution is directly connected with user in electric system and to user
The link of electric energy is distributed, whether the state of controller switching equipment is normally directly related to power quality and safety, therefore guarantees that distribution is set
Standby normal operation is most important.
Traditional controller switching equipment monitoring method mainly uses manual inspection mode, i.e., patrol officer is by periodically setting distribution
Standby, instrument and other equipment carry out artificial data reading, record and interpretation.Need to take a significant amount of time, human and material resources and financial resources into
Row manual inspection and maintenance, to realize the monitoring of power status and the elimination of security risk.The mode cannot achieve real-time monitoring,
Also it cannot achieve the early warning to potential risk.
Multi-source fusion data acquisition technology provides data branch to monitor based on the controller switching equipment of machine vision and artificial intelligence
Support.But due to controller switching equipment it is many kinds of, come in every shape, scene is complicated, this causes, and monitoring data redundancy is big, coupling
Height, data processing and intellectual analysis are the technical bottlenecks for limiting its application at present.
Current text detection localization method both domestic and external mainly has the positioning side of the localization method of connected region, edge detection
Method, the method based on textural characteristics, the method based on machine learning.Character recognition method mainly has the identification based on statistical nature
Method, structure character identifying method, statistics identify the method combined with structure recognition, the identification side based on artificial neural network
Method.Although at home and abroad all being studied lteral data recognition methods, in electric utility practical application, exist automatic
Change degree is lower, and real-time is poor, can not adapt to complex environment and identification accuracy it is not high the problems such as, it is complicated in electric utility
The automation and intelligent and real-time that text data identifies under scene increasingly show its urgency and importance.
The identification of power industry lteral data mainly has text information to identify and the digital reading of digital displaying meter identifies, this hair
The bright intelligent recognition for being intended to the lteral data that power industry is realized under complex environment, to replace artificial text information to read and realize
The real-time and intelligence of identification.
A kind of billing information recognizer, equipment and storage medium based on CRNN of CN109214382A.This method comprises:
Specific region, such as purchaser Name area are positioned and are syncopated as, from bill picture in order to preferably identify；Using CRNN
Neural metwork training generates identification model, to identify to the information of the specific region on bill.This method is preferably suitable
Recognizer for invoice information.But the recognizer is applied to bank slip recognition for specific type identification, is suitable only for
Identify bill, the lteral data that can not be suitable under Complex Power scene identifies demand.
A kind of text detection of CN109034155A and the method and system of identification, applied to the text letter in identification picture
Breath, according to the character area position in data set, using deep neural network training word area detection model, then according to text
Block domain picture be can detecte out and be schemed by word area detection model with corresponding text information training Text region model
Character area in piece can identify the text information in character area picture, two models couplings by Text region model
The text information that getting up may be implemented in picture identifies.Although general recognition methods, general data, which is concentrated, is free of electric power
Industry technical term and peculiar data type (such as seven segment digital tubes), effect is poor when for power industry identification, is suitable only for one
As scene, and can not solve the problems, such as that text such as lacks, blocks at the power industries complex scene.
In summary, the technical difficulty for Text region being carried out in existing power industry mainly covers following aspect:
(1) text information and digital reading test problems under complex environment, in power industry, distribution environmental background is multiple
Miscellaneous, text information target is small, and there are text occlusion issues.Lteral data is a complete text lines, and Yao Zuowei mono- whole
Body detected.The detection difficulty of text information identification and digital reading increases and real-time is poor, lteral data and background
Difference is huge, the layout of font size and multiplicity complexity.
(2) illumination, weather condition are complicated under natural scene, can detect to lteral data and text identification generates to a certain degree
Interference, such as reversible-light shooting, insufficient light or over-exposed, greasy weather will affect lteral data detection and identification.
(3) there is deformation problem in text data, and the image text region in power industry is also possible to be deformed, revolve
Turn, distortion etc. a variety of styles.
(4) the complicated multiplicity of the text data content of power industry, there are multilingual character etc. and seven segment digital tubes etc.
A variety of text data types.
(5) lteral data of power industry there are problems that missing, it is excessively fuzzy or be blocked and Text region exist it is wrong
How misrecognition problem restores lteral data to greatest extent and is a technical problem to be solved urgently.
Summary of the invention
In view of the problems of the existing technology, the invention discloses the intelligence of the lteral data under power industry complex scene
Recognition methods.
Explanation of technical terms:
Multi-source fusion data acquisition technology has merged visual light imaging, infrared imaging, laser point cloud imaging, inertial navigation
The multiple technologies such as IMU technology, global location GPS technology, digital photogrammetry and image processing techniques breach traditional single
Channel information cognitive method.Multi-source fusion video monitoring system described herein can be with the texture of quick obtaining scene and target
Data, point cloud data and temperature field data have high efficiency, high-precision advantage.Multi-source fusion video monitoring system, Utilities Electric Co.
System can synchronize high quality acquisition to the multiple dimensioned signal of multidimensional in complex scene.
Technical scheme is as follows:
The intelligent identification Method of lteral data under power industry complex scene, which comprises the following steps:
S1: the video data saved from the acquisition of Utilities Electric Co.'s multi-source fusion video monitoring system takes frame, the figure that will acquire
Piece carries out starting picture mark work after image enhancement, standardization: text data position and text in mark picture
Notebook data content；
S2: using the picture of step S1 acquisition mark as data set and Reading Chinese Text in the Wild
(RCTW-17) data set is in conjunction with training text detection model, Text region model；
S3: it after above-mentioned two model training, is carried out in real time firstly, capturing multi-source fusion video monitoring system video flowing
Frame is taken, then, carries out series of preprocessing to picture is obtained using digital image processing techniques: so as to subsequent detection identification；
S4: step S3 treated picture is inputted in trained text detection model, to the text in detection image
Notebook data region and the location information for saving text block cut text data region to be identified；
S5: the text data region picture after step S4 cutting is input to trained Text region model and carries out text
Data content identification；
S6: post-processing the content of text of identification, filters out undesirable character, and utilize NLP nature language
It says that Processing Algorithm carries out missing information prediction, is lacked with this completion text data region, covered lteral data information.
Preferred according to the present invention, the step S1 is specifically included:
S11: a large amount of video data is saved from the acquisition of Utilities Electric Co.'s multi-source fusion video monitoring system and takes frame；
S12: the picture saved to step S11 first does deduplication processing, then unifies dimension of picture, then to carry out picture bright
Degree adjustment and picture contrast enhancing processing are later successively being handled picture progress direction of rotation correction and perspective transform；
S13: text detection model is carried out to the picture of step S12 respectively and text identification model marks twice:
Text detection model mark: mark the rectangle frame position of the entire text data in all pictures coordinate (x1,
Y1), (x2, y2), (x3, y3), (x4, y4) and save；
Text identification model mark: it on the basis of text detection model mark, is cut according to the position coordinates of mark
Text data is cut, and marks the content of text data.
Preferred according to the present invention, the step S2 is specifically included:
S21: training text detection model, text detection mode input is picture, and output is lteral data in picture
In position:
Using training method end to end, text is obtained using preceding 5 convolution stages of convolutional neural networks VGG16 model
The Feature Mapping W*H*C of picture, wherein C is the number in Feature Mapping or channel, and W × H is space layout, in the 5th convolutional layer
Feature Mapping each position on take 3*3*C window feature, the feature will be used to predict that the k anchor point in the position to be corresponding
Classification information and location information；The feature W*3*3*C of the corresponding 3*3*C of all windows of every a line is input to circulation nerve
In network, the output of W*256 is obtained, the W*256 of Recognition with Recurrent Neural Network is input to the full articulamentum of 512 dimensions, by full articulamentum
Feature is input to three classification or returns in layer；Finally with simple line of text construction algorithm, the text that classification is obtained
Zonule is merged into text filed；
S22: three kinds of loss functions are introduced when training text detection model:Text/non-textual loss function using
Soft max function,The loss function of text vertical coordinate using L1 norm regularization,Text edges refinement
Loss function using L1 norm regularization, specific formula is as follows:
In formula, each described anchor point is a training sample, and i is an anchor point in a small lot data
Index, siIt is text/non-textual prediction probability of anchor point i,It is true value；vjWithBe anchor point j prediction and it is true hang down
Straight coordinate value；K is the index of edge anchor point, is defined as in the left or right side horizontal distance of actual text row bound frame
One group of anchor point, okWithIt is the prediction and practical true excursions amount of the associated x-axis of k-th of anchor point；Parameter lambda1=1, parameter lambda2=
2, parameter Ns,Nv, and NoIt is normalizing parameter, indicatesThe anchor point sum used respectively；
S23: training Text region model, the network architecture of Text region model include:
1. convolutional layer extracts characteristic sequence from input picture；
2. circulation layer predicts the label distribution of each frame；
3. transcribing layer, the label that each frame is predicted is become into final sequence label；
The training first step is that the picture that will be trained is sent to convolutional layer progress convolution extraction feature, obtains Feature Mapping
Then feature map is inputted by each column of Feature Mapping feature map or per several column as a time series
Feature, be sent into circulation layer, when obtain length memory network LSTM after soft max function obtains character probabilities score matrix,
Output result is obtained through transcription；During carrying out Text region model training, need to do picture altitude once normalizing,
The step S13 picture marked is switched into LMDB database file, wherein save two kinds of data, one is image data, one
Kind is label data；Timing class classification (CTC) neural network based is the calculation method of Text region model loss function, is used
Timing class classification (CTC) neural network based replaces Soft max loss function.
Preferred according to the present invention, the step S3 is specifically included:
S31: gaussian filtering process is carried out to picture to be identified；
S32: interfering for the light under natural environment, carries out limitation Contrast-limited adaptive histogram equalization to picture and calculates
Method changes picture contrast and brightness to redistribute brightness；
The limitation Contrast-limited adaptive histogram equalization algorithm comprises the following steps that
1) image block first calculates histogram in blocks, then trims histogram, last equilibrium set clipped value as
Clip Limit, ask in histogram higher than the value part and total Excess, it is assumed that total Excess is equal
All gray levels are given, the height L=total Excess/N that histogram rises overally are found out, with upper=Clip Limit-
L is that histogram is handled as follows in boundary:
1. being directly set to Clip Limit if amplitude is higher than Clip Limit；
2. being padded to Clip Limit if amplitude is between Upper and Clip Limit；
3. directly filling up L pixel if amplitude is lower than Upper；
2) linear interpolation between, the value that each pixel is pointed out carry out bilinearity by the mapping function value of 4 sub-blocks around it and insert
Value obtains, and directly does transformation for the pixel of boundary with the mapping function of a sub-block or two sub-blocks do mapping function
Do linear interpolation；
S33: the processing of figure layer colour filter hybrid algorithm function is carried out to image:
F (a, b)=1- (1-a) * (1-b).
Preferred according to the present invention, the step S4 is specifically included:
S41: using step S3 treated picture as inputting in trained text detection model to come in detection image
Text data region；
S42: the coordinate (x1, y1) of testing result text box, (x2, y2), (x3, y3), (x4, y4) and according to text are saved
Frame coordinate cuts text block, and the text block after cutting needs corresponding in saving location information of the text block in original image.
Preferred according to the present invention, the step S5 is specifically included:
S51: multiple text blocks that S42 is obtained carry out perspective transform, to correct text；
S52: the text block after perspective transform is denoised, light correction process；
S53: by treated, text block is used as input to enter Text region model；
S54: extracting image convolution feature by convolutional neural networks first, then by length memory network LSTM into
Sequence signature in onestep extraction image convolution feature is finally introducing timing class classification CTC neural network based and solves training
When character the problem of can not being aligned, save output recognition result.
Preferred according to the present invention, the step S6 is specifically included:
Word content after S61: read step S5 identification；
S62: filtering undesirable character, then handles text information and lacks problem:
The text of missing is predicted using the continuous bag of words CBOW of word2vec, firstly, utilizing open Chinese corpus
And electrical system correlation corpus trains CBOW model, makes the missing text of this model prediction electric utility；Mode input
For filtered content of text, the content of text of the missing for prediction is exported, finally by the missing content of text of prediction and existing
Content of text is integrated, and word content is restored；
The objective function of the CBOW model are as follows:
Training process optimizes it using stochastic gradient rise method；
S63: the recognition result of step S42 location information and step S62 is mapped, and a text filed coordinate pair is answered
Corresponding text identification content, i.e. completion identification mission.
Excellent technique effect of the invention is as follows:
Lteral data of the method for the invention practical application on the controller switching equipment of the switchgear house of power industry is intelligently known
When other, good result is obtained.This method uses series of preprocessing before text detection, enhances and takes in frame picture
Lteral data information weakens irrelevant information, carries out post-processing after recognition and uses natural language processing field
CBOW model is post-processed, and the effect of detection and identification is promoted with this.Text data verification and measurement ratio is 96%, and identification is accurate
Rate is 85% or so, accuracy rate with higher, while having faster operation speed on local server, can satisfy reality
The requirement of when property.This method is applicable not only to distribution room environmental simultaneously, applies also for the various complex scenes of electric utility even
Natural scene is widely used.
Detailed description of the invention
Fig. 1 is the procedure chart of present invention training text detection and Text region model；
Fig. 2 is applicating flow chart of the present invention；
Fig. 3 is algorithm flow chart of the present invention in training text detection model；
The present invention predicts the algorithm flow example diagram of missing text when Fig. 4, wherein the multiple existing word Context (w) of input,
It is predicted, the conduct of maximum probability is chosen to export prediction word w；
Fig. 5 is Baidu API recognition effect figure；
Fig. 6 is recognition effect figure of the present invention.
Specific embodiment
The present invention is described in detail below with reference to embodiment and Figure of description, but not limited to this.
Embodiment,
As shown in Figs 1-4.
The intelligent identification Method of lteral data under power industry complex scene, the following steps are included:
S1: the video data saved from the acquisition of Utilities Electric Co.'s multi-source fusion video monitoring system takes frame, the figure that will acquire
Piece carries out starting picture mark work after image enhancement, standardization: text data position and text in mark picture
Notebook data content；
S2: using the picture of step S1 acquisition mark as data set and Reading Chinese Text in the Wild
(RCTW-17) data set is in conjunction with training text detection model, Text region model；
S3: it after above-mentioned two model training, is carried out in real time firstly, capturing multi-source fusion video monitoring system video flowing
Frame is taken, then, carries out series of preprocessing to picture is obtained using digital image processing techniques: including gaussian filtering, limitation pair
The pretreatment such as mix than the self-adapting histogram equilibrium algorithm of degree, figure layer colour filter, so as to subsequent detection identification；
S4: step S3 treated picture is inputted in trained text detection model, to the text in detection image
Notebook data region and the location information for saving text block cut text data region to be identified；
S5: the text data region picture after step S4 cutting is input to trained Text region model and carries out text
Data content identification；
S6: post-processing the content of text of identification, filters out undesirable character, and utilize NLP nature language
It says that Processing Algorithm carries out missing information prediction, is lacked with this completion text data region, covered lteral data information.
The step S1 is specifically included:
S11: a large amount of video data is saved from the acquisition of Utilities Electric Co.'s multi-source fusion video monitoring system and takes frame；Every ten frames
It takes primary and saves picture；
S12: the picture saved to step S11 first does deduplication processing, deletes repeated and redundant picture, then unifies picture
Size (height is fixed, width adequate ratio), then picture luminance adjustment and picture contrast enhancing processing are carried out, later successively
Direction of rotation correction and perspective transform processing are carried out to picture；
S13: text detection model is carried out to the picture of step S12 respectively and text identification model marks twice:
Text detection model mark: mark the rectangle frame position of the entire text data in all pictures coordinate (x1,
Y1), (x2, y2), (x3, y3), (x4, y4) and save；
Text identification model mark: it on the basis of text detection model mark, is cut according to the position coordinates of mark
Text data is cut, and marks the content of text data.
The step S2 is specifically included:
S21: training text detection model, text detection mode input is picture, and output is lteral data in picture
In position:
Using training method end to end, text is obtained using preceding 5 convolution stages of convolutional neural networks VGG16 model
The Feature Mapping W*H*C of picture, wherein C is the number in Feature Mapping or channel, and W × H is space layout, in the 5th convolutional layer
Feature Mapping each position on take 3*3*C window feature, the feature will be used to predict that the k anchor point in the position to be corresponding
Classification information and location information；The feature W*3*3*C of the corresponding 3*3*C of all windows of every a line is input to circulation nerve
In network (two-way length memory network), the output of W*256 is obtained, the W*256 of Recognition with Recurrent Neural Network is input to the complete of 512 dimensions
Full articulamentum feature is input to three classification or returned in layer by articulamentum；Finally with simple line of text construction algorithm,
The zonule for obtained text of classifying is merged into text filed；Its training parameter is provided that
1) it for each Zhang Xunlian picture, needs picture to be marked a series of text for being converted into that fixed widths are 16 pixels
Frame, a figure extract 128 samples in total, and 64 positive 64 is negative, if positive sample not enough just with negative sample polishing；Detection zone and reality
As positive sample, detection zone also defines anchor point of the border region degree of overlapping greater than 0.7 with the maximum anchor point of actual area degree of overlapping
For positive sample, detection zone and anchor point of the actual area degree of overlapping less than 0.5 are defined as negative sample；
2) training picture is all by short side scaling to 600 pixels, and keeps the pantograph ratio of original image；
3) Recognition with Recurrent Neural Network layer and output layer use Stochastic Mean-Value for 0, and the parameter that variance is 1 is initialized；
4) in training, two layer parameters are fixed before convolutional neural networks；
S22: three kinds of loss functions are introduced when training text detection model:Text/non-textual loss function using
Soft max function,The loss function of text vertical coordinate using L1 norm regularization,Text edges refinement
Loss function using L1 norm regularization, specific formula is as follows:
In formula, each described anchor point is a training sample, and i is an anchor point in a small lot data
Index, siIt is text/non-textual prediction probability of anchor point i,It is true value；vjWithBe anchor point j prediction and it is true hang down
Straight coordinate value；K is the index of edge anchor point, is defined as in the left or right side horizontal distance of actual text row bound frame
One group of anchor point, okWithIt is the prediction and practical true excursions amount of the associated x-axis of k-th of anchor point；Parameter lambda1=1, parameter lambda2=
2, parameter Ns,Nv, and NoIt is normalizing parameter, indicatesThe anchor point sum used respectively；It is sharp in training process
With the backpropagation of overall goal L function, reversed derivation, stochastic gradient descent makes error minimum to determine gradient vector, passes through ladder
Vector is spent to adjust each weight, so that overall goal L function is tended to 0 or convergent trend, repetitive process is until setting number
Or the average value that damage error is lost no longer declines, i.e. minimum point；
S23: training Text region model, the network architecture of Text region model include:
4. convolutional layer extracts characteristic sequence from input picture；
5. circulation layer predicts the label distribution of each frame；
6. transcribing layer, the label that each frame is predicted is become into final sequence label；
The training first step is that the picture that will be trained is sent to convolutional layer progress convolution extraction feature, obtains Feature Mapping
Then feature map is inputted by each column of Feature Mapping feature map or per several column as a time series
Feature is sent into circulation layer (two-way length memory network), is obtained when obtaining length memory network LSTM by soft max function
After character probabilities score matrix, output result is obtained through transcription；During carrying out Text region model training, need to figure
Image height degree, which is done, to be normalized, and the step S13 picture marked is switched to LMDB database file, wherein saving two kinds of numbers
According to one is image datas, and one is label datas, they respectively have its key value；Timing class classification neural network based
(CTC) be Text region model loss function calculation method, with timing class neural network based classification (CTC) replace Soft
Max loss function, training sample introduce null character, solve the problems, such as that some positions do not have character, can pass through without alignment
Recursion quickly calculates gradient.
The step S3 is specifically included:
S31: gaussian filtering process is carried out to picture to be identified, gaussian filtering is a kind of linear smoothing filtering, suitable for disappearing
Except Gaussian noise, gaussian filtering is exactly the process being weighted and averaged to entire image, the value of each pixel, all by it
Other pixel values in body and neighborhood obtain after being weighted averagely；
S32: interfering for the light under natural environment, carries out limitation Contrast-limited adaptive histogram equalization to picture and calculates
Method changes picture contrast and brightness to redistribute brightness；Contrast self-adapting histogram equilibrium CLAHE algorithm is limited, with
Common self-adapting histogram equilibrium different places are contrast clipping, i.e., following histogram trims process, after trimming
When histogram equalization image, picture contrast can be more natural.In addition CLAHE algorithm is fine to Misty Image treatment effect, can
It realizes dehazing function, realizes foggy-dog intelligent identification；
The limitation Contrast-limited adaptive histogram equalization algorithm comprises the following steps that
1) image block first calculates histogram in blocks, then trims histogram, last equilibrium set clipped value as
Clip Limit, ask in histogram higher than the value part and total Excess, it is assumed that total Excess is equal
All gray levels are given, the height L=total Excess/N that histogram rises overally are found out, with upper=Clip Limit-
L is that histogram is handled as follows in boundary:
4. being directly set to Clip Limit if amplitude is higher than Clip Limit；
5. being padded to Clip Limit if amplitude is between Upper and Clip Limit；
6. directly filling up L pixel if amplitude is lower than Upper；
2) linear interpolation between needs exist for traversing, operates each image block, deal with it is more complex, need using insert
It is worth operation, that is, the value that each pixel is pointed out is obtained by the mapping function value progress bilinear interpolation of 4 sub-blocks around it, it is right
Transformation is directly done with the mapping function of a sub-block in the pixel of boundary or two sub-blocks do mapping function and do linear insert
Value；
S33: some colour casts of image after limitation Contrast-limited adaptive histogram equalization algorithm process is complete, so having handled
Figure layer colour filter hybrid manipulation is carried out afterwards, be can be very good to reduce colour cast, is looked at enhanced picture more natural, image is carried out
The processing of figure layer colour filter hybrid algorithm function:
F (a, b)=1- (1-a) * (1-b).
The step S4 is specifically included:
S41: using step S3 treated picture as inputting in trained text detection model to come in detection image
Text data region；
S42: the coordinate (x1, y1) of testing result text box, (x2, y2), (x3, y3), (x4, y4) and according to text are saved
Frame coordinate cuts text block, and the text block after cutting needs corresponding in saving location information of the text block in original image.
The step S5 is specifically included:
S51: multiple text blocks that S42 is obtained carry out perspective transform, to correct text；
S52: the text block after perspective transform is denoised, light correction process；
S53: by treated, text block is used as input to enter Text region model；
S54: extracting image convolution feature by convolutional neural networks first, then by length memory network LSTM into
Sequence signature in onestep extraction image convolution feature is finally introducing timing class classification CTC neural network based and solves training
When character the problem of can not being aligned, save output recognition result.
The step S6 is specifically included:
Word content after S61: read step S5 identification, because incomplete or excessively fuzzy word can be identified as some surprises
Strange word and symbol, so needing to judge whether each character meets the requirements, if accord with for rarely used word (seldom in power industry
Using) or ambiguity character etc.；
S62: filtering undesirable character, then handles text information and lacks problem:
The text of missing is predicted using the continuous bag of words CBOW of word2vec, firstly, utilizing open Chinese corpus
And electrical system correlation corpus trains CBOW model, makes the missing text of this model prediction electric utility；Mode input
For filtered content of text, the content of text of the missing for prediction is exported, finally by the missing content of text of prediction and existing
Content of text is integrated, with this come solve the problems, such as text information lack, restore word content；Such as input is ' straight * feeder line
Screen ', model prediction output is ' stream ', is ' DC feeder screen ' after integration；
The objective function of the CBOW model are as follows:
Training process optimizes it using stochastic gradient rise method；
S63: the recognition result of step S42 location information and step S62 is mapped, and a text filed coordinate pair is answered
Corresponding text identification content, i.e. completion identification mission.
Application examples:
It takes frame one to open photo and carries out identification comparison with the API recognition methods of the method for the present invention and existing Baidu respectively:
Fig. 5 is Baidu API recognition effect figure, and Fig. 6 is recognition effect figure of the present invention, is compared it is found that using of the present invention
When method identifies the text in same power industry picture, verification and measurement ratio and recognition accuracy all have clear superiority.
Claims (7)
1. the intelligent identification Method of the lteral data under power industry complex scene, which comprises the following steps:
S1: from Utilities Electric Co.'s multi-source fusion video monitoring system acquisition save video data take frame, the picture that will acquire into
Start picture after row image enhancement, standardization and mark work: text data position and textual data in mark picture
According to content；
S2: using the picture of step S1 acquisition mark as data set and Reading Chinese Text in the Wild
(RCTW-17) data set is in conjunction with training text detection model, Text region model；
S3: it after above-mentioned two model training, is taken in real time firstly, capturing multi-source fusion video monitoring system video flowing
Then frame carries out series of preprocessing to picture is obtained using digital image processing techniques: so as to subsequent detection identification；
S4: step S3 treated picture is inputted in trained text detection model, to the textual data in detection image
According to region and the location information of text block is saved, cuts text data region to be identified；
S5: the text data region picture after step S4 cutting is input to trained Text region model and carries out text data
Content recognition；
S6: post-processing the content of text of identification, filters out undesirable character, and using at NLP natural language
Adjustment method carries out missing information prediction, is lacked with this completion text data region, covered lteral data information.
2. the intelligent identification Method of the lteral data under power industry complex scene according to claim 1, feature exist
In the step S1 is specifically included:
S11: a large amount of video data is saved from the acquisition of Utilities Electric Co.'s multi-source fusion video monitoring system and takes frame；
S12: the picture saved to step S11 first does deduplication processing, then unifies dimension of picture, then carry out picture luminance tune
Whole and picture contrast enhancing processing is successively carrying out direction of rotation correction and perspective transform processing to picture later；
S13: text detection model is carried out to the picture of step S12 respectively and text identification model marks twice:
Text detection model mark: mark the rectangle frame position of the entire text data in all pictures coordinate (x1, y1),
(x2, y2), (x3, y3), (x4, y4) are simultaneously saved；
Text identification model mark: on the basis of text detection model mark, text is cut according to the position coordinates of mark
Notebook data, and mark the content of text data.
3. the intelligent identification Method of the lteral data under power industry complex scene according to claim 1, feature exist
In the step S2 is specifically included:
S21: training text detection model, text detection mode input is picture, and output is lteral data in picture
Position:
Using training method end to end, text picture is obtained using preceding 5 convolution stages of convolutional neural networks VGG16 model
Feature Mapping W*H*C, wherein C is the number in Feature Mapping or channel, and W × H is space layout, in the spy of the 5th convolutional layer
The feature that the window of 3*3*C is taken on each position of mapping is levied, the feature will be used to predict the corresponding class of the anchor point of position k
Other information and location information；The feature W*3*3*C of the corresponding 3*3*C of all windows of every a line is input to Recognition with Recurrent Neural Network
In, the output of W*256 is obtained, the W*256 of Recognition with Recurrent Neural Network is input to the full articulamentum of 512 dimensions, by full articulamentum feature
It is input to three classification or returns in layer；Finally with simple line of text construction algorithm, the cell for the text that classification is obtained
Domain is merged into text filed；
S22: three kinds of loss functions are introduced when training text detection model:Text/non-textual loss function is using soft
Max function,The loss function of text vertical coordinate using L1 norm regularization,The loss letter of text edges refinement
It counts using L1 norm regularization, specific formula is as follows:
In formula, each described anchor point is a training sample, and i is the rope of an anchor point in a small lot data
Draw, siIt is text/non-textual prediction probability of anchor point i,It is true value；vjWithIt is anchor point j prediction and true vertical
Coordinate value；K is the index of edge anchor point, is defined as in the left or right side horizontal distance of actual text row bound frame
One group of anchor point, okWithIt is the prediction and practical true excursions amount of the associated x-axis of k-th of anchor point；Parameter lambda1=1, parameter lambda2=2,
Parameter Ns, NvAnd NoIt is normalizing parameter, indicatesThe anchor point sum used respectively；
S23: training Text region model, the network architecture of Text region model include:
1. convolutional layer extracts characteristic sequence from input picture；
2. circulation layer predicts the label distribution of each frame；
3. transcribing layer, the label that each frame is predicted is become into final sequence label；
The training first step is that the picture that will be trained is sent to convolutional layer progress convolution extraction feature, obtains Feature Mapping feature
Then map is used as a time series input feature vector by each column of Feature Mapping feature map or per several column, is sent into
Circulation layer, when obtain length memory network LSTM after soft max function obtains character probabilities score matrix, through transcription be
Obtain output result；During carrying out Text region model training, need to do picture altitude once normalizing, by step
The picture that S13 has been marked switchs to LMDB database file, wherein saving two kinds of data, one is image datas, and one is marks
Sign data；Timing class classification (CTC) neural network based is the calculation method of Text region model loss function, with based on mind
Timing class classification (CTC) through network replaces Soft max loss function.
4. the intelligent identification Method of the lteral data under power industry complex scene according to claim 1, feature exist
In the step S3 is specifically included:
S31: gaussian filtering process is carried out to picture to be identified；
S32: interfering for the light under natural environment, carries out limitation Contrast-limited adaptive histogram equalization algorithm to picture and comes
Brightness is redistributed to change picture contrast and brightness；
The limitation Contrast-limited adaptive histogram equalization algorithm comprises the following steps that
1) image block first calculates histogram in blocks, then trims histogram, and last equilibrium sets clipped value as Clip
Limit, ask in histogram higher than the value part and total Excess, it is assumed that total Excess is given institute
There is gray level, the height L=total Excess/N that histogram rises overally is found out, using upper=Clip Limit-L as boundary
Histogram is handled as follows in limit:
1. being directly set to Clip Limit if amplitude is higher than Clip Limit；
2. being padded to Clip Limit if amplitude is between Upper and Clip Limit；
3. directly filling up L pixel if amplitude is lower than Upper；
2) linear interpolation between, the value that each pixel is pointed out carry out bilinear interpolation by the mapping function value of 4 sub-blocks around it and obtain
It arrives, transformation directly is done with the mapping function of a sub-block for the pixel of boundary or two sub-blocks do mapping function and make line
Property interpolation；
S33: the processing of figure layer colour filter hybrid algorithm function is carried out to image:
F (a, b)=1- (1-a) * (1-b).
5. the intelligent identification Method of the lteral data under power industry complex scene according to claim 1, feature exist
In the step S4 is specifically included:
S41: carry out the text in detection image using step S3 treated picture as inputting in trained text detection model
Data area；
S42: the coordinate (x1, y1) of testing result text box, (x2, y2), (x3, y3), (x4, y4) are saved and is sat according to text box
Mark cutting text block, the text block after cutting need corresponding in saving location information of the text block in original image.
6. the intelligent identification Method of the lteral data under power industry complex scene according to claim 1, feature exist
In the step S5 is specifically included:
S51: multiple text blocks that S42 is obtained carry out perspective transform, to correct text；
S52: the text block after perspective transform is denoised, light correction process；
S53: by treated, text block is used as input to enter Text region model；
S54: extracting image convolution feature by convolutional neural networks first, then further by length memory network LSTM
The sequence signature in image convolution feature is extracted, timing class classification CTC neural network based is finally introducing and solves word when training
The problem of symbol can not be aligned saves output recognition result.
7. the intelligent identification Method of the lteral data under power industry complex scene according to claim 1, feature exist
In the step S6 is specifically included:
Word content after S61: read step S5 identification；
S62: filtering undesirable character, then handles text information and lacks problem:
The text of missing is predicted using the continuous bag of words CBOW of word2vec, firstly, using open Chinese corpus and
Electrical system correlation corpus trains CBOW model, makes the missing text of this model prediction electric utility；Mode input was
Content of text after filter exports the content of text of the missing for prediction, finally by the missing content of text of prediction and existing text
Content is integrated, and word content is restored；
The objective function of the CBOW model are as follows:
Training process optimizes it using stochastic gradient rise method；
S63: the recognition result of step S42 location information and step S62 is mapped, and a text filed coordinate pair should be corresponding
Text identification content, i.e., completion identification mission.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910318881.4A CN110059694B (en) | 2019-04-19 | 2019-04-19 | Intelligent identification method for character data in complex scene of power industry |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910318881.4A CN110059694B (en) | 2019-04-19 | 2019-04-19 | Intelligent identification method for character data in complex scene of power industry |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110059694A true CN110059694A (en) | 2019-07-26 |
| CN110059694B CN110059694B (en) | 2020-02-11 |
Family
ID=67319656
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910318881.4A Active CN110059694B (en) | 2019-04-19 | 2019-04-19 | Intelligent identification method for character data in complex scene of power industry |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110059694B (en) |
Cited By (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110399845A (en) * | 2019-07-29 | 2019-11-01 | 上海海事大学 | Continuously at section text detection and recognition methods in a kind of image |
| CN110519485A (en) * | 2019-09-09 | 2019-11-29 | Oppo广东移动通信有限公司 | Image processing method, device, storage medium and electronic equipment |
| CN110796082A (en) * | 2019-10-29 | 2020-02-14 | 上海眼控科技股份有限公司 | Nameplate text detection method and device, computer equipment and storage medium |
| CN110826560A (en) * | 2019-11-06 | 2020-02-21 | 山东省计算中心（国家超级计算济南中心） | Esophageal cancer pathological image labeling method |
| CN110852324A (en) * | 2019-08-23 | 2020-02-28 | 上海撬动网络科技有限公司 | Deep neural network-based container number detection method |
| CN110909737A (en) * | 2019-11-14 | 2020-03-24 | 武汉虹旭信息技术有限责任公司 | Picture character recognition method and system |
| CN110909728A (en) * | 2019-12-03 | 2020-03-24 | 中国太平洋保险（集团）股份有限公司 | Control algorithm and device for multilingual policy automatic identification |
| CN110956171A (en) * | 2019-11-06 | 2020-04-03 | 广州供电局有限公司 | Automatic nameplate identification method and device, computer equipment and storage medium |
| CN110991448A (en) * | 2019-11-27 | 2020-04-10 | 云南电网有限责任公司电力科学研究院 | Text detection method and device for nameplate image of power equipment |
| CN111191649A (en) * | 2019-12-31 | 2020-05-22 | 上海眼控科技股份有限公司 | Method and equipment for identifying bent multi-line text image |
| CN111221995A (en) * | 2019-10-10 | 2020-06-02 | 南昌市微轲联信息技术有限公司 | Sequence matching method based on big data and set theory |
| CN111666931A (en) * | 2020-05-21 | 2020-09-15 | 平安科技（深圳）有限公司 | Character and image recognition method, device and equipment based on mixed convolution and storage medium |
| CN111900694A (en) * | 2020-07-07 | 2020-11-06 | 贵州电网有限责任公司 | Relay protection equipment information acquisition method and system based on automatic identification |
| CN111985484A (en) * | 2020-08-11 | 2020-11-24 | 云南电网有限责任公司电力科学研究院 | CNN-LSTM-based temperature instrument digital identification method and device |
| CN112183233A (en) * | 2020-09-09 | 2021-01-05 | 上海鹰觉科技有限公司 | Ship board identification method and system based on deep learning |
| CN112230876A (en) * | 2020-10-13 | 2021-01-15 | 华南师范大学 | Artificial intelligence reading accompanying method and reading accompanying robot |
| CN112364821A (en) * | 2020-11-30 | 2021-02-12 | 国网湖南省电力有限公司 | Self-recognition method and device for power mode data of relay protection device |
| CN112396057A (en) * | 2019-08-13 | 2021-02-23 | 上海高德威智能交通系统有限公司 | Character recognition method and device and electronic equipment |
| CN112836632A (en) * | 2021-02-02 | 2021-05-25 | 浪潮云信息技术股份公司 | Method and system for realizing user-defined template character recognition |
| CN112990212A (en) * | 2021-02-05 | 2021-06-18 | 开放智能机器（上海）有限公司 | Reading method and device of thermal imaging temperature map, electronic equipment and storage medium |
| CN113159086A (en) * | 2020-12-31 | 2021-07-23 | 南京太司德智能电气有限公司 | Efficient power signal description model training method |
| CN113449728A (en) * | 2021-07-21 | 2021-09-28 | 北京有竹居网络技术有限公司 | Character recognition method and related equipment thereof |
| CN113505741A (en) * | 2021-07-27 | 2021-10-15 | 京东科技控股股份有限公司 | Text image processing method and device, electronic equipment and storage medium |
| CN113688815A (en) * | 2021-06-01 | 2021-11-23 | 无锡启凌科技有限公司 | Medicine packaging text computer recognition algorithm and device for complex illumination environment |
| CN114092947A (en) * | 2022-01-04 | 2022-02-25 | 湖南师范大学 | Text detection method and device, electronic equipment and readable storage medium |
| CN116912845A (en) * | 2023-06-16 | 2023-10-20 | 广东电网有限责任公司佛山供电局 | Intelligent content identification and analysis method and device based on NLP and AI |
| WO2023231987A1 (en) * | 2022-05-30 | 2023-12-07 | 华为技术有限公司 | Text recognition method and electronic device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103390194A (en) * | 2012-05-07 | 2013-11-13 | 北京三星通信技术研究有限公司 | Method, device and system for predicating user intention and recommending suggestion |
| US20180336185A1 (en) * | 2017-05-17 | 2018-11-22 | International Business Machines Corporation | Natural language processing of formatted documents |
| CN108898131A (en) * | 2018-05-23 | 2018-11-27 | 郑州金惠计算机系统工程有限公司 | It is a kind of complexity natural scene under digital instrument recognition methods |
-
2019
- 2019-04-19 CN CN201910318881.4A patent/CN110059694B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103390194A (en) * | 2012-05-07 | 2013-11-13 | 北京三星通信技术研究有限公司 | Method, device and system for predicating user intention and recommending suggestion |
| US20180336185A1 (en) * | 2017-05-17 | 2018-11-22 | International Business Machines Corporation | Natural language processing of formatted documents |
| CN108898131A (en) * | 2018-05-23 | 2018-11-27 | 郑州金惠计算机系统工程有限公司 | It is a kind of complexity natural scene under digital instrument recognition methods |
Non-Patent Citations (1)
| Title |
|---|
| 蔡成涛等: "《海洋环境下的计算机视觉技术》", 31 October 2015 * |
Cited By (36)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110399845A (en) * | 2019-07-29 | 2019-11-01 | 上海海事大学 | Continuously at section text detection and recognition methods in a kind of image |
| CN112396057A (en) * | 2019-08-13 | 2021-02-23 | 上海高德威智能交通系统有限公司 | Character recognition method and device and electronic equipment |
| CN110852324A (en) * | 2019-08-23 | 2020-02-28 | 上海撬动网络科技有限公司 | Deep neural network-based container number detection method |
| CN110519485A (en) * | 2019-09-09 | 2019-11-29 | Oppo广东移动通信有限公司 | Image processing method, device, storage medium and electronic equipment |
| CN110519485B (en) * | 2019-09-09 | 2021-08-31 | Oppo广东移动通信有限公司 | Image processing method, image processing device, storage medium and electronic equipment |
| CN111221995B (en) * | 2019-10-10 | 2023-10-03 | 南昌市微轲联信息技术有限公司 | Sequence matching method based on big data and set theory |
| CN111221995A (en) * | 2019-10-10 | 2020-06-02 | 南昌市微轲联信息技术有限公司 | Sequence matching method based on big data and set theory |
| CN110796082A (en) * | 2019-10-29 | 2020-02-14 | 上海眼控科技股份有限公司 | Nameplate text detection method and device, computer equipment and storage medium |
| CN110796082B (en) * | 2019-10-29 | 2020-11-24 | 上海眼控科技股份有限公司 | Nameplate text detection method and device, computer equipment and storage medium |
| CN110826560A (en) * | 2019-11-06 | 2020-02-21 | 山东省计算中心（国家超级计算济南中心） | Esophageal cancer pathological image labeling method |
| CN110826560B (en) * | 2019-11-06 | 2023-10-27 | 山东省计算中心（国家超级计算济南中心） | Esophageal cancer pathological image labeling method |
| CN110956171A (en) * | 2019-11-06 | 2020-04-03 | 广州供电局有限公司 | Automatic nameplate identification method and device, computer equipment and storage medium |
| CN110909737A (en) * | 2019-11-14 | 2020-03-24 | 武汉虹旭信息技术有限责任公司 | Picture character recognition method and system |
| CN110991448A (en) * | 2019-11-27 | 2020-04-10 | 云南电网有限责任公司电力科学研究院 | Text detection method and device for nameplate image of power equipment |
| CN110909728A (en) * | 2019-12-03 | 2020-03-24 | 中国太平洋保险（集团）股份有限公司 | Control algorithm and device for multilingual policy automatic identification |
| CN111191649A (en) * | 2019-12-31 | 2020-05-22 | 上海眼控科技股份有限公司 | Method and equipment for identifying bent multi-line text image |
| CN111666931A (en) * | 2020-05-21 | 2020-09-15 | 平安科技（深圳）有限公司 | Character and image recognition method, device and equipment based on mixed convolution and storage medium |
| CN111666931B (en) * | 2020-05-21 | 2024-05-28 | 平安科技（深圳）有限公司 | Mixed convolution text image recognition method, device, equipment and storage medium |
| CN111900694B (en) * | 2020-07-07 | 2022-12-27 | 贵州电网有限责任公司 | Relay protection equipment information acquisition method and system based on automatic identification |
| CN111900694A (en) * | 2020-07-07 | 2020-11-06 | 贵州电网有限责任公司 | Relay protection equipment information acquisition method and system based on automatic identification |
| CN111985484A (en) * | 2020-08-11 | 2020-11-24 | 云南电网有限责任公司电力科学研究院 | CNN-LSTM-based temperature instrument digital identification method and device |
| CN112183233A (en) * | 2020-09-09 | 2021-01-05 | 上海鹰觉科技有限公司 | Ship board identification method and system based on deep learning |
| CN112230876A (en) * | 2020-10-13 | 2021-01-15 | 华南师范大学 | Artificial intelligence reading accompanying method and reading accompanying robot |
| CN112364821A (en) * | 2020-11-30 | 2021-02-12 | 国网湖南省电力有限公司 | Self-recognition method and device for power mode data of relay protection device |
| CN113159086B (en) * | 2020-12-31 | 2024-04-30 | 南京太司德智能电气有限公司 | Efficient electric power signal description model training method |
| CN113159086A (en) * | 2020-12-31 | 2021-07-23 | 南京太司德智能电气有限公司 | Efficient power signal description model training method |
| CN112836632A (en) * | 2021-02-02 | 2021-05-25 | 浪潮云信息技术股份公司 | Method and system for realizing user-defined template character recognition |
| CN112990212A (en) * | 2021-02-05 | 2021-06-18 | 开放智能机器（上海）有限公司 | Reading method and device of thermal imaging temperature map, electronic equipment and storage medium |
| CN113688815A (en) * | 2021-06-01 | 2021-11-23 | 无锡启凌科技有限公司 | Medicine packaging text computer recognition algorithm and device for complex illumination environment |
| CN113449728A (en) * | 2021-07-21 | 2021-09-28 | 北京有竹居网络技术有限公司 | Character recognition method and related equipment thereof |
| CN113505741A (en) * | 2021-07-27 | 2021-10-15 | 京东科技控股股份有限公司 | Text image processing method and device, electronic equipment and storage medium |
| CN113505741B (en) * | 2021-07-27 | 2024-04-09 | 京东科技控股股份有限公司 | Text image processing method and device, electronic equipment and storage medium |
| CN114092947A (en) * | 2022-01-04 | 2022-02-25 | 湖南师范大学 | Text detection method and device, electronic equipment and readable storage medium |
| WO2023231987A1 (en) * | 2022-05-30 | 2023-12-07 | 华为技术有限公司 | Text recognition method and electronic device |
| CN116912845B (en) * | 2023-06-16 | 2024-03-19 | 广东电网有限责任公司佛山供电局 | Intelligent content identification and analysis method and device based on NLP and AI |
| CN116912845A (en) * | 2023-06-16 | 2023-10-20 | 广东电网有限责任公司佛山供电局 | Intelligent content identification and analysis method and device based on NLP and AI |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110059694B (en) | 2020-02-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110059694A (en) | The intelligent identification Method of lteral data under power industry complex scene | |
| CN111723860B (en) | Target detection method and device | |
| CN108710865B (en) | Driver abnormal behavior detection method based on neural network | |
| CN109670429B (en) | Method and system for detecting multiple targets of human faces of surveillance videos based on instance segmentation | |
| Sun et al. | Traffic sign detection and recognition based on convolutional neural network | |
| CN104050471B (en) | Natural scene character detection method and system | |
| CN109740413A (en) | Pedestrian recognition methods, device, computer equipment and computer storage medium again | |
| CN104408449B (en) | Intelligent mobile terminal scene literal processing method | |
| CN104091341B (en) | A kind of image fuzzy detection method based on conspicuousness detection | |
| CN110298343A (en) | A kind of hand-written blackboard writing on the blackboard recognition methods | |
| CN101615244A (en) | Handwritten plate blank numbers automatic identifying method and recognition device | |
| CN114333070A (en) | Examinee abnormal behavior detection method based on deep learning | |
| CN111507227B (en) | Multi-student individual segmentation and state autonomous identification method based on deep learning | |
| CN107038416A (en) | A kind of pedestrian detection method based on bianry image modified HOG features | |
| CN107945210A (en) | Target tracking algorism based on deep learning and environment self-adaption | |
| CN110929746A (en) | Electronic file title positioning, extracting and classifying method based on deep neural network | |
| CN112232371A (en) | American license plate recognition method based on YOLOv3 and text recognition | |
| CN113297956B (en) | Gesture recognition method and system based on vision | |
| CN110533100A (en) | A method of CME detection and tracking is carried out based on machine learning | |
| CN110135446A (en) | Method for text detection and computer storage medium | |
| CN109472733A (en) | Image latent writing analysis method based on convolutional neural networks | |
| CN112364883A (en) | American license plate recognition method based on single-stage target detection and deptext recognition network | |
| CN116597438A (en) | Improved fruit identification method and system based on Yolov5 | |
| CN111126155B (en) | Pedestrian re-identification method for generating countermeasure network based on semantic constraint | |
| CN110991374B (en) | Fingerprint singular point detection method based on RCNN |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20190726 Address after: Licheng Alexander Road in Ji'nan City, Shandong province 250199 No. 27 Applicant after: SHANDONG University Applicant after: SHANDONG ZHIYANG ELECTRIC Co.,Ltd. Address before: Licheng Alexander Road in Ji'nan City, Shandong province 250199 No. 27 Applicant before: Shandong University |
|
| TA01 | Transfer of patent application right | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB02 | Change of applicant information |
Address after: Licheng Alexander Road in Ji'nan City, Shandong province 250199 No. 27 Applicant after: SHANDONG University Applicant after: Zhiyang Innovation Technology Co.,Ltd. Address before: Licheng Alexander Road in Ji'nan City, Shandong province 250199 No. 27 Applicant before: Shandong University Applicant before: SHANDONG ZHIYANG ELECTRIC Co.,Ltd. |
|
| CB02 | Change of applicant information | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20200110 Address after: Licheng Alexander Road in Ji'nan City, Shandong province 250199 No. 27 Applicant after: SHANDONG University Applicant after: Zhiyang Innovation Technology Co.,Ltd. Applicant after: STATE GRID ZHEJIANG ELECTRIC POWER Co.,Ltd. Address before: Licheng Alexander Road in Ji'nan City, Shandong province 250199 No. 27 Applicant before: Shandong University Applicant before: Zhiyang Innovation Technology Co.,Ltd. |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |