CN108874900A - A kind of acquisition methods and system of samples pictures data acquisition system - Google Patents
A kind of acquisition methods and system of samples pictures data acquisition system Download PDFInfo
- Publication number
- CN108874900A CN108874900A CN201810506155.0A CN201810506155A CN108874900A CN 108874900 A CN108874900 A CN 108874900A CN 201810506155 A CN201810506155 A CN 201810506155A CN 108874900 A CN108874900 A CN 108874900A
- Authority
- CN
- China
- Prior art keywords
- image data
- pictures
- data
- confidence level
- positive sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- 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
- G06F18/211—Selection of the most significant subset of features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention provides the acquisition methods and system of a kind of samples pictures data acquisition system, method includes:The cleaning process of image data to be cleaned specifically includes:Obtain positive sample pictures and negative sample pictures;The characteristic information of image data in positive sample pictures and the characteristic information of Target Photo are identical;The characteristic information of image data in negative sample pictures and the characteristic information of Target Photo be not identical;According to positive sample pictures and negative sample pictures, training obtains neural network sorter;Image data to be cleaned is classified according to neural network sorter to obtain several confidence level set;The acquisition process of samples pictures data acquisition system specifically includes:It obtains confidence level and reaches the image data in the confidence level set of predetermined level, obtain samples pictures data acquisition system.The present invention realizes that automatic screening and classification obtain samples pictures data acquisition system, improves sifting sort efficiency and accuracy.
Description
Technical field
The present invention relates to data processing field, the espespecially a kind of acquisition methods and system of samples pictures data acquisition system.
Background technique
It is well known that in the training in deep learning convolutional neural networks, it would be desirable to the data of magnanimity.And one at
The data volume of ripe neural network can be readily achieved tb grades, and by taking convolutional neural networks as an example, the input source of data is generally
Picture, bigger picture, one big about several million, even if the generally several hundred k of smaller picture, are counted according to tb grades of data volume
It calculates, this will be very large workload.
The way of industry is at present, using crawling for web crawlers magnanimity, then all by manually being screened and being classified
The problem of acquisition samples pictures data acquisition system, this processing mode is brought is that workload is extremely huge, and the selection result subjectivity is big,
The selection result is easy error.Meanwhile the later period is trained neural network using the samples pictures data acquisition system of mistake, can bring
The classification results of mistake.
Summary of the invention
The object of the present invention is to provide the acquisition methods and system of a kind of samples pictures data acquisition system, realize automatic screening and
Classification obtains samples pictures data acquisition system, improves sifting sort efficiency and accuracy.
Technical solution provided by the invention is as follows:
The present invention provides a kind of acquisition methods of samples pictures data acquisition system, including step:
The cleaning process of image data to be cleaned specifically includes:
Obtain positive sample pictures and negative sample pictures;The characteristic information of image data in the positive sample pictures
It is identical as the characteristic information of Target Photo;The characteristic information of image data in the negative sample pictures and the spy of Target Photo
Reference breath is not identical;
According to the positive sample pictures and the negative sample pictures, training obtains neural network sorter;
Image data to be cleaned is classified according to the neural network sorter to obtain several confidence level set;
The acquisition process of samples pictures data acquisition system specifically includes:
It obtains confidence level and reaches the image data in the confidence level set of predetermined level, obtain samples pictures data set
It closes;The characteristic information of image data in the samples pictures data acquisition system is identical as the characteristic information of Target Photo and described
The image data quantity of samples pictures data acquisition system is greater than the image data quantity of the positive sample pictures.
Further, the acquisition positive sample pictures and negative sample pictures include step:
Concentrate the positive sample image data for obtaining the first preset number as the positive sample pictures from source image data;
The positive sample image data is image data identical with the Target Photo characteristic information;
Concentrate the negative sample image data for obtaining the second preset number as the negative sample pictures from source image data;
The negative sample image data be and the different image data of Target Photo characteristic information.
Further, described according to the positive sample pictures and the negative sample pictures, training obtains neural network
Sorter includes step:
Delete the last one full articulamentum of the neural network model of pre-training;
A full articulamentum and an active coating are successively added at the last one full articulamentum deleting;
It is obtained according to the positive sample pictures and the newly added full articulamentum of negative sample pictures training and active coating
To the neural network sorter.
Further, described that image data to be cleaned is classified to obtain several according to the neural network sorter
Confidence level set includes step:
All image datas to be cleaned are inputted into the neural network sorter, obtain setting for each image data to be cleaned
Reliability;
Range is divided according to the confidence level of the image data to be cleaned and default confidence interval, it will be described to be cleaned
Image data is referred to corresponding confidence level set.
Further, described according to the confidence level according to the image data to be cleaned and default confidence interval
Range is divided, it includes step that the image data to be cleaned, which is referred to after corresponding confidence level set,:
It obtains confidence level and reaches the image data in the confidence level set of predetermined level as positive sample image data
It is added to the positive sample pictures.
Further, the acquisition confidence level reaches the image data in the confidence level set of predetermined level, obtains
It include step after samples pictures data acquisition system:
The number of pictures of image data of the statistical confidence grade in the confidence level set of predetermined level range;
When the number of pictures reaches target requirement quantity, stop the cleaning process of image data to be cleaned;
When the number of pictures miss the mark quantity required, continue the cleaning process of image data to be cleaned.
The present invention also provides a kind of acquisition systems of samples pictures data acquisition system, including:Cleaning module and acquisition module;Institute
Cleaning module is stated to connect with the acquisition module;
The cleaning module obtains positive sample pictures and negative sample pictures;Picture in the positive sample pictures
The characteristic information of data and the characteristic information of Target Photo are identical;The characteristic information of image data in the negative sample pictures
It is not identical as the characteristic information of Target Photo;According to the positive sample pictures and the negative sample pictures, training obtains mind
Through network distribution device;Image data to be cleaned is classified according to the neural network sorter to obtain several confidence level collection
It closes;
The acquisition module obtains confidence level and reaches the image data in the confidence level set of predetermined level, obtains
Samples pictures data acquisition system;The characteristic information of image data in the samples pictures data acquisition system and the feature of Target Photo are believed
Manner of breathing is same, and the image data quantity of the samples pictures data acquisition system is greater than the image data number of the positive sample pictures
Amount.
Further, the cleaning module includes:Positive sample data capture unit, negative sample data capture unit, sorting
Device training unit and classified storage unit;The sorter training unit is described respectively with the positive sample data capture unit
Negative sample data capture unit is connected with the classified storage unit;
The positive sample data capture unit concentrates the positive sample picture number for obtaining the first preset number from source image data
According to as the positive sample pictures;The positive sample image data is picture number identical with the Target Photo characteristic information
According to;
The negative sample data capture unit concentrates the negative sample picture number for obtaining the second preset number from source image data
According to as the negative sample pictures;The negative sample image data be and the different picture of Target Photo characteristic information
Data;
The sorter training unit deletes the last one full articulamentum of the neural network model of pre-training;It is deleting
A full articulamentum and an active coating are successively added at the last one full articulamentum;According to positive sample pictures and described
The newly added full articulamentum of negative sample pictures training and active coating obtain the neural network sorter;
All image datas to be cleaned are inputted the neural network sorter, obtained each by the classified storage unit
The confidence level of image data to be cleaned;Model is divided according to the confidence level of the image data to be cleaned and default confidence interval
It encloses, the image data to be cleaned is referred to corresponding confidence level set.
Further, the positive sample data capture unit is connect with the classified storage unit;
The positive sample data capture unit, also acquisition confidence level reach the figure in the confidence level set of predetermined level
Sheet data is added to the positive sample pictures as positive sample image data.
Further, further include:Statistical module and control module;The statistical module respectively with the acquisition module and institute
Control module connection is stated, the control module is connect with the cleaning module;
The statistical module, the figure of image data of the statistical confidence grade in the confidence level set of predetermined level range
Piece number;
When the number of pictures reaches target requirement quantity, the control module controls the cleaning module and stops to clear
Wash the cleaning process of image data;
When the number of pictures miss the mark quantity required, the control module control the cleaning module continue to
Clean the cleaning process of image data.
The acquisition methods and system of a kind of samples pictures data acquisition system provided through the invention, can bring it is following at least
A kind of beneficial effect:
1) present invention classifies to picture to be cleaned by neural network sorter, reduces the labour of artificial screening classification
Power and subjectivity promote sifting sort efficiency, thus when the quantity of the corresponding image data of part picture characteristic information is inadequate,
Efficiently, the screening of high quality expands sample of the corresponding image data as neural network model.
2) present invention obtains samples pictures data acquisition system and provides a kind of image data to be cleaned in the process cleans automatically
Auxiliary data cleaning strategy for improving artificial cleaning efficiency, or even full automation cleaning, and solves artificial cleaning bring
Variety of problems reduces the subjectivity of artificial cleaning sifting sort, promotes the accuracy and efficiency of sifting sort.
3) present invention is sorted while sorting image data to be cleaned using neural network sorter using neural network
The image data for the high confidence level that device sifting sort obtains is added in positive sample, is trained to neural network sorter, can
To effectively improve model accuracy, accelerate the speed of neural network sorter training.
Detailed description of the invention
Below by clearly understandable mode, preferred embodiment is described with reference to the drawings, to a kind of samples pictures data set
The acquisition methods of conjunction and above-mentioned characteristic, technical characteristic, advantage and its implementation of system are further described.
Fig. 1 is a kind of flow chart of one embodiment of the acquisition methods of samples pictures data acquisition system of the present invention;
Fig. 2 is a kind of flow chart of an example of the acquisition methods of samples pictures data acquisition system of the present invention;
Fig. 3 is a kind of structural schematic diagram of one embodiment of the acquisition system of samples pictures data acquisition system of the present invention;
Fig. 4 is a kind of structural schematic diagram of one embodiment of the acquisition system of samples pictures data acquisition system of the present invention.
Specific embodiment
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, Detailed description of the invention will be compareed below
A specific embodiment of the invention.It should be evident that drawings in the following description are only some embodiments of the invention, for
For those of ordinary skill in the art, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing, and obtain other embodiments.
To make simplified form, part related to the present invention is only schematically shown in each figure, they are not represented
Its practical structures as product.In addition, there is identical structure or function in some figures so that simplified form is easy to understand
Component only symbolically depicts one of those, or has only marked one of those.Herein, "one" is not only indicated
" only this ", can also indicate the situation of " more than one ".
Transfer learning (Transfer learning), as the term suggests being exactly is exactly to move trained model parameter has been learned
New model is moved on to help new model training.In view of there are correlations for most of data or task, so by moving
Move study we can by the model parameter acquired (also be understood as model acquire knowledge) by certain mode come point
It enjoys and not having to as most of networks to accelerate the learning efficiency of simultaneously Optimized model from zero study to new model.
Data cleansing just refers to:The data of needs are picked out in the data set of magnanimity.Such as classification problem, it lifts
A example:Screen magnanimity tomato omelette/omelet picture, cleaning task be exactly by these pictures be tomato omelette/omelet picture screening
Out.
First embodiment of the invention, as shown in Figure 1:
A kind of acquisition methods of samples pictures data acquisition system, including:
The cleaning process of image data to be cleaned specifically includes:
Obtain positive sample pictures and negative sample pictures;The characteristic information of image data in the positive sample pictures
It is identical as the characteristic information of Target Photo;The characteristic information of image data in the negative sample pictures and the spy of Target Photo
Reference breath is not identical;
Specifically, characteristic information includes but is not limited to image content, the picture features such as picture classification.It can be to from network
On the picture that crawls by artificial screening preset number image data identical with the characteristic information of Target Photo as positive sample
The different image data negative sample pictures of the characteristic information of pictures, artificial screening preset number and Target Photo;It can also
To open image data identical with the characteristic information of Target Photo by artificial screening one to the picture crawled from network, and with
The identical image data of the characteristic information with Target Photo is that template picture carries out data augmentation, such as pixel transform data augmentation
It carries out data augmentation with any one or a variety of pairs of template pictures in geometric transformation data augmentation and obtains positive sample pictures.
Similarly, negative sample pictures can also be obtained according to the mode of data augmentation.Any acquisition positive sample pictures and negative sample
This pictures is not limited thereto, and is all belonged to the scope of protection of the present invention.
Wherein, pixel transform includes:1, increase noise and filtering, the mode of noise include but is not limited to salt-pepper noise,
Gaussian noise, median filtering;2, channel is converted, the sequence in tri- channels RBG is adjusted;3, contrast, brightness and saturation degree are adjusted,
Color jitter.
Geometric transformation includes:1, it overturns, such as:Flip horizontal, it is vertical to overturn, it overturns according to the actual situation, for example, close
In face, the face reformed into down has been spun upside down, has been overturn without practical significance;2, it translates, simulates real-life
Picture situation not placed in the middle, occurs the transformation of position;3, it rotates;4, it sets black, simulates the data sample being at least partially obscured;5, it cuts out
It cuts;6, it scales.
According to the positive sample pictures and the negative sample pictures, training obtains neural network sorter;
Specifically, the image data in positive sample image data collection and negative sample pictures can be inputted to pre-training mind
Through being finely adjusted to the parameter of all levels in pre-training neural network model in network model, then carry out to all layers
Grade is trained to obtain neural network sorter;It can also be by the picture number in positive sample image data collection and negative sample pictures
According to inputting in pre-training neural network model, to the parameter of the full articulamentum of the last layer in pre-training neural network model into
Row fine tuning, then carries out being trained to obtain neural network sorter to the full articulamentum of the last layer.
Image data to be cleaned is classified according to the neural network sorter to obtain several confidence level set;
Specifically, the neural network sorter obtained by training carries out screening point to the image data to be cleaned of acquisition
Class, such as 100,000 image datas to be cleaned are crawled from network, this 100,000 image datas to be cleaned are input to nerve net
In network sorter, corresponding figure to be cleaned is calculated by carrying out in neural network sorter to each image data to be cleaned
The predicted value of sheet data identifies the classification of each picture to be cleaned according to the corresponding predicted value of each image data to be cleaned, from
And classify to all image datas to be cleaned.
The acquisition process of samples pictures data acquisition system specifically includes:
It obtains confidence level and reaches the image data in the confidence level set of predetermined level, obtain samples pictures data set
It closes;The characteristic information of image data in the samples pictures data acquisition system is identical as the characteristic information of Target Photo and described
The image data quantity of samples pictures data acquisition system is greater than the image data quantity of the positive sample pictures.
Specifically, in the present embodiment, the image data that confidence level reaches in the confidence level set of predetermined level is exactly
Similarity reaches the picture of default similarity threshold, these pictures between Target Photo corresponding with the characteristic information of Target Photo
Data are exactly the picture of user demand.The present invention classifies to picture to be cleaned by neural network sorter, reduces artificial
The labour of sifting sort and subjectivity promote sifting sort efficiency, thus in the corresponding picture number of part picture characteristic information
According to quantity it is inadequate when, efficiently, the screening of high quality expands the sample of corresponding image data as neural network model.
Second embodiment of the invention, the present embodiment are the preferred embodiments of first embodiment, with above-mentioned first embodiment phase
Than advanced optimizing, obtaining positive sample pictures and negative sample pictures include:
Concentrate the positive sample image data for obtaining the first preset number as the positive sample pictures from source image data;
The positive sample image data is image data identical with the Target Photo characteristic information;
Concentrate the negative sample image data for obtaining the second preset number as the negative sample pictures from source image data;
The negative sample image data be and the different image data of Target Photo characteristic information.
Specifically, in many occasions, it is not necessary that train entire neural network model (random initializtion ginseng from the beginning
Number) because without data set abundant enough, and training is also very time-consuming, cost source process.Therefore, direct basis
The neural network model of the pre-training of trained mistake has been classified before this source image data, which is concentrated, obtains certain amount such as the
Then the positive sample image data of one preset quantity is concentrated from source image data as positive sample pictures and obtains certain amount such as
The positive sample image data of second preset quantity is as negative sample pictures.First preset quantity can phase with the second preset quantity
Deng can also be unequal.Such as the pre- instruction being obtained corresponding to vegetable picture is trained excessively to various vegetable pictures before this
Experienced neural network model, in order to obtain the sub-neural network model of tomato omelette/omelet vegetable, then just directly according to pre-training
The source image data that the classifier of neural network model is classified concentrates the positive sample image data for filtering out tomato omelette/omelet, directly
Positive sample pictures after having reached the first preset quantity to quantity as the sub-neural network model of tomato omelette/omelet vegetable, and
It is not tomato omelette/omelet that the source image data concentration classified according to the classifier of the neural network model of pre-training, which filters out,
Negative sample image data, the sub-neural network model after quantity has reached the second preset quantity as tomato omelette/omelet vegetable
Negative sample pictures.
Third embodiment of the invention, the present embodiment are the preferred embodiments of first embodiment, with above-mentioned first embodiment phase
Than advanced optimizing, according to the positive sample pictures and the negative sample pictures, training obtains neural network sorter
Including step:
Delete the last one full articulamentum of the neural network model of pre-training;
A full articulamentum and an active coating are successively added at the last one full articulamentum deleting;
It is obtained according to the positive sample pictures and the newly added full articulamentum of negative sample pictures training and active coating
To the neural network sorter.
Specifically, the operation of transfer learning training sorter, the actually neural network model of deletion pre-training are such as
(mobilenetv1) after the full articulamentum of the last one, then a full connection is successively added at the last one full articulamentum
Layer and an active coating.It is obtained according to positive sample pictures and the newly added full articulamentum of negative sample pictures training and active coating
Neural network sorter, neural network sorter are to the already present nerve according to above-mentioned deletion addition treated pre-training
Network model is trained fine tuning to neural network dimension and neural network model parameter, is correctly tied so that network is fitted
Fruit.Assuming that there is 400 vegetable image datas to need to clean, newly-increased full articulamentum is 1 × 1001 × 401 filter, is increased newly
A full articulamentum stochastic parameter initialization (such as gaussian random initialization), and after training, reverse conduction obtains fixed ginseng
Number.SoftMax can be used in a newly-increased active coating, guarantees that neural network during forward and reverse conducts, is conducive to
Calculate the regression fit with mathematics.This transfer learning model for being added to full an articulamentum and an active coating, because not
It is related to the modification to entire model parameter, only it needs to be determined that two layers of newly-increased this during forward-propagating and backpropagation
Parameter, and the parameter of other layers can be multiplexed, therefore can greatly be saved the training time, be saved in the case where guaranteeing precision
Time of deep learning.
Fourth embodiment of the invention, the present embodiment are the preferred embodiments of first embodiment, with above-mentioned first embodiment phase
Than advanced optimizing, being classified to obtain several confidences to image data to be cleaned according to the neural network sorter
Degree set includes step:
All image datas to be cleaned are inputted into the neural network sorter, obtain setting for each image data to be cleaned
Reliability;
Range is divided according to the confidence level of the image data to be cleaned and default confidence interval, it will be described to be cleaned
Image data is referred to corresponding confidence level set.
Specifically, all image datas to be cleaned are just inputted neural network after training obtains neural network sorter
Sorter is predicted to obtain corresponding confidence level, each be set by neural network sorter to each image data to be cleaned
What reliability represented is the image data to be cleaned as the probability of Target Photo characteristic information, and confidence level is bigger, this figure to be cleaned
A possibility that sheet data is consistent with the characteristic information of Target Photo is bigger, i.e., image data to be cleaned is more similar to Target Photo.
The present invention by real time to acquisition obtain image data to be cleaned carry out calculating confidence level, so as to according to confidence level will each to
Cleaning image data classification is divided in corresponding confidence level set, convenient for the subsequent image data in each confidence level set
Screened the samples pictures data acquisition system needed.
Fifth embodiment of the invention, the present embodiment are the preferred embodiments of fourth embodiment, with above-mentioned fourth embodiment phase
Than advanced optimizing, being divided according to the confidence level according to the image data to be cleaned and default confidence interval
Range, it includes step that the image data to be cleaned, which is referred to after corresponding confidence level set,:
It obtains confidence level and reaches the image data in the confidence level set of predetermined level as positive sample image data
It is added to the positive sample pictures.
Specifically, the sample size of the positive sample picture of model is on the low side, in fact, instructing when transfer learning constructs model
Model is in the state for being similar to poor fitting after having practiced, and at this moment, the confidence level sorted out reaches predetermined level
Image data in confidence level set is added to positive sample pictures as positive sample image data, and confidence level is reached pre-
Neural network sorter is carried out again if the image data in the confidence level set of grade is added to again in positive sample pictures
Secondary training.In this way after first time artificial screening obtains positive sample pictures, during continuous training neural network sorter,
The image data in confidence level set for directly reaching predetermined level using confidence level changes as positive sample image data
Generation training, this iterative training can effectively improve model progress.It can be improved the precision of neural network sorter
While, additionally it is possible to the artificial screening for reducing positive sample pictures and negative sample pictures is greatly lowered labor workload, subtracts
Data classification caused by the subjectivity of few artificial screening identification malfunctions, and improves the robustness of neural network model.
Sixth embodiment of the invention, the present embodiment are the preferred embodiments of the first to the 5th embodiment, with above-mentioned first to
5th embodiment is compared, and is advanced optimized, and the acquisition confidence level reaches the figure in the confidence level set of predetermined level
Sheet data, obtaining samples pictures data acquisition system includes later step:
The number of pictures of image data of the statistical confidence grade in the confidence level set of predetermined level range;
When the number of pictures reaches target requirement quantity, stop the cleaning process of image data to be cleaned;
When the number of pictures miss the mark quantity required, continue the cleaning process of image data to be cleaned.
Specifically, it is preferred that after the confidence level that each image data to be cleaned can be calculated, automated by data acquisition
Script, data acquisition automatized script can classify to image data to be cleaned according to confidence interval division range is preset,
Different confidence level set are respectively put into, image data of the statistical confidence grade in the confidence level set of predetermined level range
Number of pictures stops the cleaning process of image data to be cleaned when number of pictures reaches target requirement quantity;Work as number of pictures
When miss the mark quantity required, continue the cleaning process of image data to be cleaned.Such as high confidence level set (such as 0.95
Or more confidence level) in, be basically required picture, cleaning at this moment can be fairly simple, it is only necessary to figure
Piece collection is swept and mistake, removes the picture of apparent error, interior for low confidence set (such as 0.3 and confidence level below),
The data set of the inside substantially the cleaning of unwanted picture at this moment can be fairly simple, it is only necessary to pictures are swept
And mistake, select obvious correctly picture.Screening sorts out neural network sorter to high confidence level collection through the above way
The image data of identification classification error is closed, remaining figure in the high confidence level set after then counting the image data of deletion error
The number of pictures of sheet data stops the cleaning process of image data to be cleaned if number of pictures reaches target requirement quantity,
Conversely, continuing the cleaning process of image data to be cleaned.Similarly, aforesaid way can also be taken to carry out low confidence set
Screening.The cleaning efficiency of image data to be cleaned can be thus promoted, the workload of the screening of artificial eye subjectivity is reduced, mentions
High working efficiency.
Further, it is also possible to range is divided according to the confidence level of the image data to be cleaned and default confidence interval,
After the image data to be cleaned is referred to corresponding confidence level set, thinking that Internet picture data set is boundless
Under assuming that, image data of the confidence level in the confidence level set outside predetermined level range can be directly abandoned, network is climbed
Worm carries out constantly crawling image data, constantly accumulates the picture of these high confidence levels, the figure in high confidence level set
Until the number of pictures of sheet data has reached target requirement quantity.Such as when needing the image data in high confidence level set,
The image data of confidence level set and low confidence set in can directly abandoning only retains the picture number in high confidence level set
According to, and the crawler of front is constantly crawled, and the picture of these high confidence levels is constantly accumulated, until picture number reaches mesh
Until marking quantity required.Simultaneously using neural network sorter sorting data, it is obtained using neural network sorter category filter
Confidence level set in image data precision raising has been carried out to neural network sorter, strengthen neural network sorter
Precision.
Based on the above embodiment, example cites an actual example, as shown in Fig. 2, including:
S1, a small amount of positive sample picture is obtained;
S2, transfer learning training neural network sorter;
S3, data acquisition automatized script sort image data to be cleaned, obtain the image data of three confidence levels;
Enter step S4 or S5;
S4, artificial screening high confidence level image data;
S5, judge whether image data is high confidence level;If so, entering step S6 and S7;Otherwise, S8 is entered step;
S6, discarding;
S7, high confidence level image data is stored to target data set;
S8, using high confidence level image data as positive sample picture, and enter step S2;
S9, judge that target data concentrates whether the number of pictures of high confidence level image data reaches target requirement quantity;If
It is to enter step S10;Otherwise, return step S3;
S10, samples pictures data acquisition system is obtained.
Specifically, in the present embodiment, it is assumed that after the processing of neural network sorter, image data to be cleaned is divided into
Three classes.The first kind, the data set of high confidence level, the second class, the data set of middle confidence level, the picture in moderate confidence level data set
Data, that is, the picture for needing manually to check one by one as conventional cleaning task, it is however generally that, the data set of middle confidence level
The image data quantity of conjunction is generally not too large.The data set of third class low confidence.The picture taken off that swashes from network is past
It is past to have very more interference pictures, even completely unrelated picture.The data set to be cleaned for obtaining magnanimity, filters out
A small amount of positive sample picture and negative sample picture, it is however generally that, tens are just much of that, for the figure of tens of thousands of hundreds of thousands easily
Sheet data amount, this is extremely micro work.Transfer learning is carried out using positive sample picture and negative sample picture, training obtains one
The neural network model of pre-training is deleted the last layer, and successively that is, in a manner of transfer learning by a neural network sorter
A full articulamentum and an active coating is newly added, using positive sample picture and negative sample picture to the full articulamentum being newly added
And active coating is trained, and has the training time short, and can achieve relatively high accuracy.The nerve obtained using training
Network distribution device predicts image data to be cleaned to obtain corresponding confidence level.This confidence level represents whether picture is me
Required for targeted species probability.
It, can be by data acquisition automatized script to be cleaned after obtaining the confidence level of each image data to be cleaned
Image data carry out classified storage, data acquisition automatized script can according to low middle high three sections to image data to be cleaned into
Row classified storage is respectively put into three different confidence level set.Substantially for the image data in high confidence level set
It is exactly required picture.At this moment cleaning can be fairly simple, it is only necessary to carry out to the image data in high confidence level set
Sweep and mistake, remove be clearly not Target Photo image data.It is basic for the image data in low confidence set
Upper is exactly unwanted picture, and cleaning at this moment can be fairly simple, it is only necessary to the image data in low confidence set
Swept and mistake, remove be obviously Target Photo image data.Both modes all greatly improve work effect
Rate.
After obtaining the confidence level of each image data to be cleaned, can also by neural network sorter calculate each to
It cleans the confidence level of image data, and is divided after range classified according to confidence level and default confidence interval, thought
Under the boundless hypothesis of Internet picture data set, the image data of middle confidence level and low confidence can be directly discarded,
Only retain the image data of high confidence level, and constantly crawled image data to be cleaned by web crawlers, constantly tires out
The image data of product high confidence level, until the number of pictures of the image data of high confidence level reaches target requirement quantity.?
While sorting image data to be cleaned using neural network sorter, the height that is obtained using neural network sorter sifting sort
The image data of confidence level is added in positive sample, is trained to neural network sorter, and model essence can be effectively improved
Degree.
The present invention picture quality is very good and picture quality is excessively poor when can play the role of it is extraordinary, very
Extremely to the perfect condition that can not need manually to clean substantially.But, it is however generally that, for conservative, it is desired nonetheless to manually most
Inspection afterwards, the even so also very big efficiency for improving data cleansing task and the subjectivity for avoiding data cleansing
With tired bring accuracy.The present invention is to provide a kind of picture number to be cleaned during obtaining samples pictures data acquisition system
According to the auxiliary data cleaning strategy cleaned automatically, for improving artificial cleaning efficiency, or even full automation cleaning, and people is solved
Work cleans bring variety of problems, reduces the subjectivity of artificial cleaning sifting sort, promotes the accuracy and efficiency of sifting sort.
Seventh embodiment of the invention, as shown in figure 3, a kind of acquisition system of samples pictures data acquisition system includes:
Data cleansing module 100 and sample acquisition module 200;The data cleansing module 100 and the sample
This acquisition module 200 connects;
The data cleansing module 100 obtains positive sample pictures and negative sample pictures;In the positive sample pictures
Image data characteristic information it is identical as the characteristic information of Target Photo;The spy of image data in the negative sample pictures
Reference breath is not identical as the characteristic information of Target Photo;According to the positive sample pictures and the negative sample pictures, training
Obtain neural network sorter;Image data to be cleaned is classified to obtain several according to the neural network sorter and is set
Reliability set;
The sample acquisition module 200 obtains confidence level and reaches the picture number in the confidence level set of predetermined level
According to obtaining samples pictures data acquisition system;The characteristic information and Target Photo of image data in the samples pictures data acquisition system
Characteristic information it is identical, and the image data quantity of the samples pictures data acquisition system be greater than the positive sample pictures picture
Data bulk.
Specifically, the present embodiment is the corresponding system embodiment of above method embodiment, specific effect is referring to above-mentioned first
Embodiment, this is no longer going to repeat them.
Eighth embodiment of the invention, as shown in figure 4, the present embodiment is the preferred embodiment of the 7th embodiment, with above-mentioned
Seven embodiments are compared, and the data cleansing module 100 includes:Positive sample data capture unit, negative sample data capture unit, point
Pick device training unit and classified storage unit;The sorter training unit respectively with the positive sample data capture unit, institute
Negative sample data capture unit is stated to connect with the classified storage unit;
The positive sample data capture unit concentrates the positive sample picture number for obtaining the first preset number from source image data
According to as the positive sample pictures;The positive sample image data is picture number identical with the Target Photo characteristic information
According to;
The negative sample data capture unit concentrates the negative sample picture number for obtaining the second preset number from source image data
According to as the negative sample pictures;The negative sample image data be and the different picture of Target Photo characteristic information
Data;
The sorter training unit deletes the last one full articulamentum of the neural network model of pre-training;It is deleting
A full articulamentum and an active coating are successively added at the last one full articulamentum;According to positive sample pictures and described
The newly added full articulamentum of negative sample pictures training and active coating obtain the neural network sorter;
All image datas to be cleaned are inputted the neural network sorter, obtained each by the classified storage unit
The confidence level of image data to be cleaned;Model is divided according to the confidence level of the image data to be cleaned and default confidence interval
It encloses, the image data to be cleaned is referred to corresponding confidence level set.
Specifically, the present embodiment is the corresponding system embodiment of above method embodiment, specific effect is referring to above-mentioned second
Embodiment, 3rd embodiment and fourth embodiment, this is no longer going to repeat them.
Ninth embodiment of the invention, the present embodiment are the preferred embodiments of the 8th embodiment, with above-mentioned 8th embodiment phase
Than the positive sample data capture unit is connect with the classified storage unit;
The positive sample data capture unit, also acquisition confidence level reach the figure in the confidence level set of predetermined level
Sheet data is added to the positive sample pictures as positive sample image data.
Specifically, the present embodiment is the corresponding system embodiment of above method embodiment, specific effect is referring to the above-mentioned 5th
Embodiment, this is no longer going to repeat them.
Tenth embodiment of the invention, the present embodiment are the preferred embodiments of the 7th to the 8th embodiment, with the above-mentioned 7th to
8th embodiment is compared, and further includes:Statistical module and control module;The statistical module respectively with the sample acquisition module
200 connect with the control module, and the control module is connect with the data cleansing module 100;
The statistical module, the figure of image data of the statistical confidence grade in the confidence level set of predetermined level range
Piece number;
When the number of pictures reaches target requirement quantity, the control module controls the data cleansing module 100
Stop the cleaning process of image data to be cleaned;
When the number of pictures miss the mark quantity required, the control module controls the data cleansing module
100 continue the cleaning process of image data to be cleaned.
Specifically, the present embodiment is the corresponding system embodiment of above method embodiment, specific effect is referring to the above-mentioned 6th
Embodiment, this is no longer going to repeat them.
It should be noted that above-described embodiment can be freely combined as needed.The above is only of the invention preferred
Embodiment, it is noted that for those skilled in the art, in the premise for not departing from the principle of the invention
Under, several improvements and modifications can also be made, these modifications and embellishments should also be considered as the scope of protection of the present invention.
Claims (10)
1. a kind of acquisition methods of samples pictures data acquisition system, which is characterized in that including step:
The cleaning process of image data to be cleaned specifically includes:
Obtain positive sample pictures and negative sample pictures;The characteristic information and mesh of image data in the positive sample pictures
Mark on a map piece characteristic information it is identical;The characteristic information of image data in the negative sample pictures and the feature of Target Photo are believed
It ceases not identical;
According to the positive sample pictures and the negative sample pictures, training obtains neural network sorter;
Image data to be cleaned is classified according to the neural network sorter to obtain several confidence level set;
The acquisition process of samples pictures data acquisition system specifically includes:
It obtains confidence level and reaches the image data in the confidence level set of predetermined level, obtain samples pictures data acquisition system;
The characteristic information of image data in the samples pictures data acquisition system is identical as the characteristic information of Target Photo, and the sample
The image data quantity of image data set is greater than the image data quantity of the positive sample pictures.
2. the acquisition methods of samples pictures data acquisition system according to claim 1, which is characterized in that the acquisition positive sample
Pictures and negative sample pictures include step:
Concentrate the positive sample image data for obtaining the first preset number as the positive sample pictures from source image data;It is described
Positive sample image data is image data identical with the Target Photo characteristic information;
Concentrate the negative sample image data for obtaining the second preset number as the negative sample pictures from source image data;It is described
Negative sample image data be and the different image data of Target Photo characteristic information.
3. the acquisition methods of samples pictures data acquisition system according to claim 1, which is characterized in that it is described according to it is described just
Samples pictures collection and the negative sample pictures, it includes step that training, which obtains neural network sorter,:
Delete the last one full articulamentum of the neural network model of pre-training;
A full articulamentum and an active coating are successively added at the last one full articulamentum deleting;
Institute is obtained according to the positive sample pictures and the newly added full articulamentum of negative sample pictures training and active coating
State neural network sorter.
4. the acquisition methods of samples pictures data acquisition system according to claim 1, which is characterized in that described according to the mind
Image data to be cleaned is classified to obtain several confidence level set to include step through network distribution device:
All image datas to be cleaned are inputted into the neural network sorter, obtain the confidence of each image data to be cleaned
Degree;
Range is divided according to the confidence level of the image data to be cleaned and default confidence interval, by the picture to be cleaned
Data are referred to corresponding confidence level set.
5. the acquisition methods of samples pictures data acquisition system according to claim 4, which is characterized in that described according to described
Range is divided according to the confidence level of the image data to be cleaned and default confidence interval, the image data to be cleaned is returned
It include step after class to corresponding confidence level set:
It obtains confidence level and reaches the image data in the confidence level set of predetermined level as the addition of positive sample image data
To the positive sample pictures.
6. the acquisition methods of samples pictures data acquisition system according to claim 1-5, which is characterized in that described to obtain
Confidence level is taken to reach the image data in the confidence level set of predetermined level, obtain samples pictures data acquisition system includes later
Step:
The number of pictures of image data of the statistical confidence grade in the confidence level set of predetermined level range;
When the number of pictures reaches target requirement quantity, stop the cleaning process of image data to be cleaned;
When the number of pictures miss the mark quantity required, continue the cleaning process of image data to be cleaned.
7. a kind of acquisition system of samples pictures data acquisition system, which is characterized in that including:Cleaning module and acquisition module;It is described
Cleaning module is connect with the acquisition module;
The cleaning module obtains positive sample pictures and negative sample pictures;Image data in the positive sample pictures
Characteristic information it is identical as the characteristic information of Target Photo;The characteristic information and mesh of image data in the negative sample pictures
Mark on a map piece characteristic information it is not identical;According to the positive sample pictures and the negative sample pictures, training obtains nerve net
Network sorter;Image data to be cleaned is classified according to the neural network sorter to obtain several confidence level set;
The acquisition module obtains confidence level and reaches the image data in the confidence level set of predetermined level, obtains sample
Image data set;The characteristic information phase of the characteristic information of image data in the samples pictures data acquisition system and Target Photo
Together, and the image data quantity of the samples pictures data acquisition system be greater than the positive sample pictures image data quantity.
8. the acquisition system of samples pictures data acquisition system according to claim 7, which is characterized in that the cleaning module packet
It includes:Positive sample data capture unit, negative sample data capture unit, sorter training unit and classified storage unit;Described point
Pick device training unit respectively with the positive sample data capture unit, the negative sample data capture unit and the classified storage
Unit connection;
The positive sample data capture unit concentrates the positive sample image data for obtaining the first preset number to make from source image data
For the positive sample pictures;The positive sample image data is image data identical with the Target Photo characteristic information;
The negative sample data capture unit concentrates the negative sample image data for obtaining the second preset number to make from source image data
For the negative sample pictures;The negative sample image data be and the different picture number of the Target Photo characteristic information
According to;
The sorter training unit deletes the last one full articulamentum of the neural network model of pre-training;It is last deleting
A full articulamentum and an active coating are successively added at one full articulamentum;According to the positive sample pictures and the negative sample
The newly added full articulamentum of this pictures training and active coating obtain the neural network sorter;
All image datas to be cleaned are inputted the neural network sorter, obtained each to clear by the classified storage unit
Wash the confidence level of image data;Range is divided according to the confidence level of the image data to be cleaned and default confidence interval,
The image data to be cleaned is referred to corresponding confidence level set.
9. the acquisition system of samples pictures data acquisition system according to claim 8, which is characterized in that the positive sample data
Acquiring unit is connect with the classified storage unit;
The positive sample data capture unit, also acquisition confidence level reach the picture number in the confidence level set of predetermined level
The positive sample pictures are added to according to as positive sample image data.
10. according to the acquisition system of the described in any item samples pictures data acquisition systems of claim 7-9, which is characterized in that also wrap
It includes:Statistical module and control module;The statistical module is connect with the acquisition module and the control module respectively, the control
Molding block is connect with the cleaning module;
The statistical module, the picture number of image data of the statistical confidence grade in the confidence level set of predetermined level range
Mesh;
When the number of pictures reaches target requirement quantity, the control module controls the cleaning module and stops figure to be cleaned
The cleaning process of sheet data;
When the number of pictures miss the mark quantity required, it is to be cleaned that the control module controls the cleaning module continuation
The cleaning process of image data.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810506155.0A CN108874900A (en) | 2018-05-24 | 2018-05-24 | A kind of acquisition methods and system of samples pictures data acquisition system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810506155.0A CN108874900A (en) | 2018-05-24 | 2018-05-24 | A kind of acquisition methods and system of samples pictures data acquisition system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108874900A true CN108874900A (en) | 2018-11-23 |
Family
ID=64334141
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810506155.0A Pending CN108874900A (en) | 2018-05-24 | 2018-05-24 | A kind of acquisition methods and system of samples pictures data acquisition system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108874900A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110083728A (en) * | 2019-04-03 | 2019-08-02 | 上海联隐电子科技合伙企业(有限合伙) | A kind of methods, devices and systems of optimization automation image data cleaning quality |
CN110210536A (en) * | 2019-05-22 | 2019-09-06 | 北京邮电大学 | A kind of the physical damnification diagnostic method and device of optical interconnection system |
CN111160406A (en) * | 2019-12-10 | 2020-05-15 | 北京达佳互联信息技术有限公司 | Training method of image classification model, and image classification method and device |
CN111652259A (en) * | 2019-04-16 | 2020-09-11 | 上海铼锶信息技术有限公司 | Method and system for cleaning data |
CN111652257A (en) * | 2019-03-27 | 2020-09-11 | 上海铼锶信息技术有限公司 | Sample data cleaning method and system |
CN111651433A (en) * | 2019-03-27 | 2020-09-11 | 上海铼锶信息技术有限公司 | Sample data cleaning method and system |
CN112287923A (en) * | 2020-12-24 | 2021-01-29 | 德联易控科技(北京)有限公司 | Card information identification method, device, equipment and storage medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101814149A (en) * | 2010-05-10 | 2010-08-25 | 华中科技大学 | Self-adaptive cascade classifier training method based on online learning |
US20140355871A1 (en) * | 2012-06-15 | 2014-12-04 | Vufind, Inc. | System and method for structuring a large scale object recognition engine to maximize recognition accuracy and emulate human visual cortex |
CN104850832A (en) * | 2015-05-06 | 2015-08-19 | 中国科学院信息工程研究所 | Hierarchical iteration-based large-scale image sample marking method and system |
CN105224947A (en) * | 2014-06-06 | 2016-01-06 | 株式会社理光 | Sorter training method and system |
CN107705256A (en) * | 2017-09-13 | 2018-02-16 | 西南交通大学 | A kind of forward direction Vehicular video image enchancing method semantic based on contact net |
CN108062341A (en) * | 2016-11-08 | 2018-05-22 | 中国移动通信有限公司研究院 | The automatic marking method and device of data |
-
2018
- 2018-05-24 CN CN201810506155.0A patent/CN108874900A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101814149A (en) * | 2010-05-10 | 2010-08-25 | 华中科技大学 | Self-adaptive cascade classifier training method based on online learning |
US20140355871A1 (en) * | 2012-06-15 | 2014-12-04 | Vufind, Inc. | System and method for structuring a large scale object recognition engine to maximize recognition accuracy and emulate human visual cortex |
CN105224947A (en) * | 2014-06-06 | 2016-01-06 | 株式会社理光 | Sorter training method and system |
CN104850832A (en) * | 2015-05-06 | 2015-08-19 | 中国科学院信息工程研究所 | Hierarchical iteration-based large-scale image sample marking method and system |
CN108062341A (en) * | 2016-11-08 | 2018-05-22 | 中国移动通信有限公司研究院 | The automatic marking method and device of data |
CN107705256A (en) * | 2017-09-13 | 2018-02-16 | 西南交通大学 | A kind of forward direction Vehicular video image enchancing method semantic based on contact net |
Non-Patent Citations (1)
Title |
---|
HOO-CHANG SHIN ET AL.: "Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures,Dataset Characteristics and Transfer Learning", 《IEEE TRANSACTIONS ON MEDICAL IMAGING》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111652257A (en) * | 2019-03-27 | 2020-09-11 | 上海铼锶信息技术有限公司 | Sample data cleaning method and system |
CN111651433A (en) * | 2019-03-27 | 2020-09-11 | 上海铼锶信息技术有限公司 | Sample data cleaning method and system |
CN111651433B (en) * | 2019-03-27 | 2023-05-12 | 上海铼锶信息技术有限公司 | Sample data cleaning method and system |
CN110083728A (en) * | 2019-04-03 | 2019-08-02 | 上海联隐电子科技合伙企业(有限合伙) | A kind of methods, devices and systems of optimization automation image data cleaning quality |
CN110083728B (en) * | 2019-04-03 | 2021-08-20 | 上海铼锶信息技术有限公司 | Method, device and system for optimizing automatic picture data cleaning quality |
CN111652259A (en) * | 2019-04-16 | 2020-09-11 | 上海铼锶信息技术有限公司 | Method and system for cleaning data |
CN111652259B (en) * | 2019-04-16 | 2024-03-08 | 上海铼锶信息技术有限公司 | Method and system for cleaning data |
CN110210536A (en) * | 2019-05-22 | 2019-09-06 | 北京邮电大学 | A kind of the physical damnification diagnostic method and device of optical interconnection system |
CN111160406A (en) * | 2019-12-10 | 2020-05-15 | 北京达佳互联信息技术有限公司 | Training method of image classification model, and image classification method and device |
CN112287923A (en) * | 2020-12-24 | 2021-01-29 | 德联易控科技(北京)有限公司 | Card information identification method, device, equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108874900A (en) | A kind of acquisition methods and system of samples pictures data acquisition system | |
CN106682704B (en) | A kind of disease geo-radar image recognition methods of integrating context information | |
CN106238342B (en) | Panoramic vision potato sorts and defect detecting device and its sorting detection method | |
CN107622233A (en) | A kind of Table recognition method, identifying system and computer installation | |
CN104899871B (en) | A kind of IC elements solder joint missing solder detection method | |
CN103985182B (en) | A kind of bus passenger flow automatic counting method and automatic counter system | |
CN108596338A (en) | A kind of acquisition methods and its system of neural metwork training collection | |
CN108686978A (en) | The method for sorting and system of fruit classification and color and luster based on ARM | |
CN109102515A (en) | A kind of method for cell count based on multiple row depth convolutional neural networks | |
Gyawali et al. | Comparative analysis of multiple deep CNN models for waste classification | |
CN106841209A (en) | One kind is based on big data self study chemical fiber wire ingot intelligence outward appearance detecting system and method | |
CN110963209A (en) | Garbage sorting device and method based on deep reinforcement learning | |
Jin et al. | Design and implementation of anti-leakage planting system for transplanting machine based on fuzzy information | |
CN107506793A (en) | Clothes recognition methods and system based on weak mark image | |
CN107194418A (en) | A kind of Aphids in Rice Field detection method based on confrontation feature learning | |
CN107832780A (en) | Low confidence sample processing method and system are sorted based on artificial intelligence plank | |
CN104867145B (en) | IC element welding point defect detection methods based on VIBE models | |
CN107121436B (en) | The Intelligent detecting method and identification device of a kind of silicon material quality | |
CN108617480A (en) | A kind of gardens intelligent irrigation system | |
CN108038415A (en) | A kind of unmanned plane based on machine vision detects automatically and tracking | |
CN111652326A (en) | Improved fruit maturity identification method and identification system based on MobileNet v2 network | |
CN107358176A (en) | Sorting technique based on high score remote sensing image area information and convolutional neural networks | |
CN109241397A (en) | A kind of method and apparatus for cleaning data | |
CN106250911A (en) | A kind of picture classification method based on convolutional neural networks | |
CN108510739A (en) | A kind of road traffic state recognition methods, system and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181123 |