CN109858448A - Item identification method and equipment under a kind of public safety - Google Patents
Item identification method and equipment under a kind of public safety Download PDFInfo
- Publication number
- CN109858448A CN109858448A CN201910101328.5A CN201910101328A CN109858448A CN 109858448 A CN109858448 A CN 109858448A CN 201910101328 A CN201910101328 A CN 201910101328A CN 109858448 A CN109858448 A CN 109858448A
- Authority
- CN
- China
- Prior art keywords
- training sample
- items
- images
- goods
- type
- 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.)
- Withdrawn
Links
Abstract
The invention discloses the item identification methods and equipment under a kind of public safety, it is related to article subdivision technical field, to solve in the prior art, under public safety, in the presence of can not be similar to a large amount of appearances, but the problem of diverse article of actual object carries out precise classification identification, the method of the present invention includes: to obtain the images of items of article, the images of items is inputted to the feature identification model constructed based on deep learning, it extracts the characteristic for mapping type of goods in the images of items and exports in vector form, according to the mapping relations based on vector sum type of goods constructed by the feature identification model, obtain the corresponding type of goods of vector of the feature identification model output.
Description
Technical field
The present invention relates to articles to segment technical field, in particular to item identification method under a kind of public safety and sets
It is standby.
Background technique
With greatly developing for scientific and technological network net, in recent years, real economy transition, electric business flow bonus disappear, business mould
Formula is constantly reformed, and the new retail epoch of on-line off-line fusion upgrading are arriving, and new retail is exactly centered on consumer experience
From " goods-field-people " to the transformation of " people-goods-field ", polynary retail form is turned to from single retail, from " article+service " to " object
Product+service+content+other " conversion, allow consumption scene it is ubiquitous, make social experience flexible and convenient, under new public safety
Article is traceable, can optimize data analysis, on-line off-line phase is allowed to melt interpromoting relation in five elements.
Under new public safety, self-service, customer only need to be by cargo automatically scanning image or label, but completes
During retail payment, the prior art can only realize rough classification to article, and with calculating by image or label technique
The development of machine vision technique, depth learning technology, at present more and more people using convolutional neural networks CNN to item pictures into
Row classification, such as when identifying food, provide the labels such as milk, melon seeds, bread, potato chips, utilization CNN to above-mentioned food into
When row identification, can commodity to above-mentioned appearance dissmilarity its carry out accurately identification and classify, such as the food series, institute
Belong to manufacturer etc., but article similar for a large amount of appearances, but can not accurately identify the tool of above-mentioned food (such as milk)
Body taste (such as original flavor, banana milk, jujube milk, almond milk), and then link is paid in article, it is unable to judge accurately out
Specific price of the article under disaggregated classification, and then very important influence is brought to the new public safety of article.
In conclusion in the prior art, under new public safety, existing can not be similar to a large amount of appearances, but actual object
Diverse article carries out the problem of precise classification identification.
Summary of the invention
The present invention provides the item identification method and equipment under a kind of public safety, to solve in the prior art, zero
It sells under scene, existing can not be similar to a large amount of appearances, but the diverse article of actual object carries out asking for precise classification identification
Topic.
In a first aspect, the present invention provides the item identification method under a kind of public safety, this method comprises:
Obtain the images of items of article；
The images of items is inputted to the feature identification model constructed based on deep learning, extracts and is reflected in the images of items
It penetrates the characteristic of type of goods and exports in vector form；
According to the mapping relations based on vector sum type of goods constructed by the feature identification model, the feature is obtained
The corresponding type of goods of vector of identification model output.
In the above method, the feature for extracting images of items exports in vector form, is easier to distinguish based on vector not of the same race
Feature difference between the article of class can carry out thinner division to type of goods.
In one possible implementation, it is based on deep learning construction feature identification model, comprising:
Obtain includes the training sample set of multiple training samples and the test sample collection including multiple test samples, Mei Gexun
Practice the images of items and the affiliated type of goods of the article that sample/test sample includes article；
The model parameter of randomization deep learning network model obtains initial Forecasting recognition model, the Forecasting recognition mould
Type includes multiple feature extraction network layers；
When trigger model training, using the training sample for the preset quantity that the training sample is concentrated, current predictive is known
Other model is trained at least once, every time after training, after the test sample concentrated using the test sample is to training
Forecasting recognition model tested, when determining that test result meets default required precision, terminate training process, will remove last
The current Forecasting recognition model output of one feature extraction network layer is the feature identification model.
In the above method, tested using deep learning network training Forecasting recognition model, and with preset test sample collection
Whether the precision of prediction of Forecasting recognition model meets the requirements, and can train the spy high to type of goods accuracy of identification and accuracy
Levy identification model.
In one possible implementation, the training sample for the preset quantity concentrated every time using the training sample,
Current predictive identification model is trained, comprising:
Images of items in the training sample of preset quantity is inputted to current Forecasting recognition model, output and the article
The corresponding prediction type of goods of image；
According to the prediction type of goods of the images of items and the images of items in training sample corresponding article kind
Class adjusts the model parameter of current Forecasting recognition model by loss function.
In one possible implementation, every time after training, determine that test result is unsatisfactory for default required precision,
Further include:
Using the training sample of training sample set described in current Forecasting recognition model discrimination, by the training sample after screening
Collection is as new training sample set, and trigger model training again.
In the above method, a large amount of sample data is screened using current Forecasting recognition model, deletes redundancy
Sample data picks out the sample data useful to the feature identification model of training a new generation, shortens feature identification model
Training time.
In one possible implementation, the training of training sample set described in current Forecasting recognition model discrimination is utilized
Sample, using the training sample set after screening as new training sample set, comprising:
Images of items in the training sample of preset quantity inputs to current Forecasting recognition model respectively, extract with it is described
The corresponding vector of images of items；
Construct the mapping relations of the vector sum type of goods of the training sample of preset quantity；
Each images of items in addition to the training sample of preset quantity is concentrated to input currently the training sample respectively
The mapping of Forecasting recognition model, the vector and the vector sum type of goods that are exported according to current Forecasting recognition model is closed
System assesses the validity of the training sample；
The training sample that all validity are unsatisfactory for preset requirement is concentrated to delete from the training sample and obtains new instruction
Practice sample set.
In one possible implementation, which is characterized in that according to the vector that current Forecasting recognition model exports, with
And the mapping relations of the vector sum type of goods assess the validity of the training sample, comprising:
Determine the corresponding vector of images of items of the training sample of current Forecasting recognition model output；
Determine that the corresponding type of goods of the images of items is corresponding in the mapping relations of the vector sum type of goods
Vector determines the effective of the training sample when being greater than preset value with the vector departure degree that current Forecasting recognition model exports
Property meets preset requirement.
In the above method, the training sample training characteristics identification model that validity meets preset requirement, Jin Erti are filtered out
The precision that feature identification model carries out feature extraction to article is risen, so that above-mentioned article is accurately classified.
In one possible implementation, the mapping based on feature identification model building vector sum type of goods is closed
System, comprising:
Images of items in the training sample of preset quantity inputs to current feature identification model respectively, extract with it is described
The corresponding vector of images of items；
According to images of items described in the corresponding vector sum of the images of items in training sample corresponding type of goods, structure
Build the mapping relations of the vector sum type of goods.
In the above method, the mapping relations of vector sum type of goods are established with the feature identification model that current training obtains,
The matched accuracy of vector sum type of goods is improved, and then promotes the accuracy for carrying out category identification to article.
Second aspect, the present invention provide the article identification equipment under a kind of public safety, which includes processor and deposit
Reservoir, wherein the memory stores executable program code, when said program code is performed, so that the processor
Execute following steps:
Obtain the images of items of article；
The images of items is inputted to the feature identification model constructed based on deep learning, extracts and is reflected in the images of items
It penetrates the characteristic of type of goods and exports in vector form；
According to the mapping relations based on vector sum type of goods constructed by the feature identification model, the feature is obtained
The corresponding type of goods of vector of identification model output.
The third aspect, the application also provide a kind of computer storage medium, are stored thereon with computer program, the program quilt
The step of first aspect the method is realized when processing unit executes.
In addition, technical effect brought by any implementation can be found in first aspect in second aspect and the third aspect
Technical effect brought by middle difference implementation, details are not described herein again.
Detailed description of the invention
To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment
Attached drawing is briefly introduced, it should be apparent that, drawings in the following description are only some embodiments of the invention, for this
For the those of ordinary skill in field, without any creative labor, it can also be obtained according to these attached drawings
His attached drawing.
Fig. 1 is the item identification method flow chart under a kind of public safety that the embodiment of the present invention one provides；
Fig. 2 is a kind of flow chart for training characteristics identification model that the embodiment of the present invention one provides；
Fig. 3 is that the article under a kind of public safety provided by Embodiment 2 of the present invention identifies equipment schematic diagram；
Fig. 4 is the item identification devices schematic diagram under a kind of public safety provided by Embodiment 2 of the present invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing to the present invention make into
It is described in detail to one step, it is clear that the described embodiments are only some of the embodiments of the present invention, rather than whole implementation
Example.Based on the embodiments of the present invention, obtained by those of ordinary skill in the art without making creative efforts
All other embodiment, shall fall within the protection scope of the present invention.
The some words occurred in text are explained below:
1, term "and/or" in the embodiment of the present invention describes the incidence relation of affiliated partner, indicates that there may be three kinds of passes
System, for example, A and/or B, can indicate: individualism A exists simultaneously A and B, these three situations of individualism B.Character "/" one
As indicate forward-backward correlation object be a kind of "or" relationship.
2, term " commodity " or " article " in the embodiment of the present invention, refer to acquire above-mentioned quotient using computer vision technique
Product/images of items, and the article that can classify, it is small to arrive fuel, furniture and appliances etc. are arrived greatly.
The application scenarios of description of the embodiment of the present invention are the technical solutions in order to more clearly illustrate the embodiment of the present invention,
The restriction for technical solution provided in an embodiment of the present invention is not constituted, those of ordinary skill in the art are it is found that with newly answering
With the appearance of scene, technical solution provided in an embodiment of the present invention is equally applicable for similar technical problem.Wherein, at this
In the description of invention, unless otherwise indicated, the meaning of " plurality " is two or more.
With the development of internet technology, the development of electric business platform and the rapid development of new retail technology, self-service cabinet/
Shop increases sharply, wherein commodity commodity image quantity present explosive growth, commodity image it is automatic classification also therefore at
For an important research direction, the extensive vision sorter research based on convolutional neural networks (CNN) in recent years achieve into
How exhibition with CNN carries out commodity classification to collected commodity, but since its feature describes single, and shallow structure classification calculation
Method nicety of grading is not high, causes the feature fine granularity for the commodity image extracted inadequate, and then influence the transaction of commodity
Journey, especially in the self-service cabinet of new public safety, commodity subdivision is many kinds of, how to increase and is sold lower commodity kind to new
Class has been subdivided into urgent problem.
Above-mentioned new public safety refers to, the scene of the general retail form of the data-driven centered on consumer experience, core
Heart value is will farthest to promote whole society's circulation retail business running efficiency, specifically, new retail is to be with consumer
Center passes through each consumption of data flowing series connection on the basis of the links such as people, commodity and service, supply chain are digitized
Scene, including smart phone, mobile terminal, computer, entity sales field and the following achievable new access etc., utilize digitizing technique
It realizes that entity delivers chain with virtual retail supply chain, transaction, the comprehensive of service chaining is merged, is supplied to consumer and covers full channel
Seamless consumption experience, the efficient general general retail structure of favour type with the characteristics of the substitution entity delivery form of logistics distribution part.
Based on above-mentioned scene, the embodiment of the present invention provides the item identification method and equipment under a kind of public safety.
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing to the present invention make into
It is described in detail to one step, it is clear that described embodiments are only a part of the embodiments of the present invention, rather than whole implementation
Example.Based on the embodiments of the present invention, obtained by those of ordinary skill in the art without making creative efforts
All other embodiment, shall fall within the protection scope of the present invention.
For above-mentioned scene, the embodiment of the present invention is described in further detail with reference to the accompanying drawings of the specification.
Embodiment one:
As shown in Figure 1, being specifically included the present embodiment provides the item identification method under a kind of public safety:
Step 101, the images of items of article is obtained；
As an alternative embodiment, can with but be not limited to using one/multiple common cameras/rotation camera shooting
Head, is fixed angle/multi-angle to above-mentioned article and takes pictures and obtain the images of items of above-mentioned article.
Step 102, above-mentioned images of items is inputted to the feature identification model constructed based on deep learning, extracts above-mentioned article
The characteristic of type of goods is mapped in image and is exported in vector form；
Step 103, it according to the mapping relations based on vector sum type of goods constructed by features described above identification model, obtains
The corresponding type of goods of vector of features described above identification model output.
It should be understood that the mapping relations of above-mentioned vector sum type of goods preserve different type of goods and corresponding vector
Between mapping relations, different type of goods should respectively correspond different vectors.
It, can be with if what is obtained is the image of multiple angles of article when above-mentioned steps 101 obtain the images of items of article
But be not limited to therefrom to select a pixel it is best/image comprising above-mentioned article appointed part is as input feature vector identification model
Images of items, do not do excessive restriction to this, those skilled in the art can be arranged according to actual needs.As a kind of optional
Embodiment is based on deep learning construction feature identification model as follows:
1) training sample set including multiple training samples and the test sample collection including multiple test samples are obtained, each
Training sample/test sample includes the images of items and the affiliated type of goods of above-mentioned article of article；
2) model parameter for being randomized deep learning network model obtains initial Forecasting recognition model, above-mentioned Forecasting recognition
Model includes multiple feature extraction network layers；
Excessive restriction is not done to above-mentioned deep learning network model, those skilled in the art can set according to actual needs
Set, in the present embodiment, above-mentioned deep learning network model can with but be not limited to include: convolutional neural networks CNN
(Convolutional Neural Network), Recognition with Recurrent Neural Network RNN (Recurrent Neural Network), depth
Neural network DNN (Deep Neural Networks) etc.；
3) when trigger model training, using the training sample for the preset quantity that above-mentioned training sample is concentrated, to current predictive
Identification model is trained at least once, and every time after training, the test sample concentrated using above-mentioned test sample is to training
Forecasting recognition model afterwards is tested, and when determining that test result meets default required precision, terminates training process, will be removed most
The current Forecasting recognition model output that later feature extracts network layer is features described above identification model.
Excessive restriction is not done to the mode of above-mentioned acquisition training sample set and test sample collection, those skilled in the art can
It is arranged according to actual needs, in the present embodiment, above-mentioned training sample set and test sample collection are acquired greatly in advance by technical staff
The images of items and type of merchandize for measuring article obtain；
Excessive restriction is not done to above-mentioned preset quantity, those skilled in the art can be arranged according to actual needs.
The training sample for the preset quantity concentrated every time using above-mentioned training sample, instructs current predictive identification model
Practice, comprising:
Images of items in the training sample of preset quantity is inputted to current Forecasting recognition model, output and the article
The corresponding prediction type of goods of image；
According to the prediction type of goods of above-mentioned images of items and above-mentioned images of items in training sample corresponding article kind
Class adjusts the model parameter of current Forecasting recognition model by loss function.
As an alternative embodiment, by loss function, calculate above-mentioned images of items prediction type of goods and
The departure degree of above-mentioned images of items corresponding type of goods in training sample；
When above-mentioned departure degree is less than default deviation threshold value, that is, think the images of items of above-mentioned Forecasting recognition model output
It predicts type of goods true type of goods close in training sample, terminates training process at this time, the last one feature will be removed
The current Forecasting recognition model output for extracting network layer is features described above identification model；
When above-mentioned departure degree is not less than default deviation threshold value, that is, think the images of items of above-mentioned Forecasting recognition model output
Prediction type of goods and type of goods true in training sample deviate it is excessive, at this time according to parameter preset method of adjustment adjust
The model parameter of current Forecasting recognition model；
Excessive restriction is not done to above-mentioned parameter preset method of adjustment, those skilled in the art can set according to actual needs
Set, can with but be not limited to staged adjusting parameter method；
As an alternative embodiment, can with but be not limited to through loss function, adjustment is worked as follows
The model parameter of preceding Forecasting recognition model:
1) end value and the corresponding method of parameter adjustment calculated in above-mentioned Forecasting recognition model construction loss function, right
When current predictive identification model is trained, according to the above-mentioned Forecasting recognition model of end value adjust automatically of loss function calculating
Model parameter；2) when being trained to current predictive identification model, the end value that loss function when each training is calculated is defeated
Out, it is investigated by those skilled in the art according to expertise or early period, the model parameter of above-mentioned Forecasting recognition model is carried out
Adjustment.
As an alternative embodiment, determining that test result is unsatisfactory for default required precision every time after training
When, further include, using the training sample of the current above-mentioned training sample set of Forecasting recognition model discrimination, by the training sample after screening
This collection is as new training sample set, and trigger model training again.
As an alternative embodiment, with the training sample of the above-mentioned above-mentioned training sample set of Forecasting recognition model discrimination
This, using the training sample set after screening as new training sample set, comprising:
Images of items in the training sample of preset quantity inputs to current Forecasting recognition model respectively, extract with it is above-mentioned
The corresponding vector of images of items；
Construct the mapping relations of the vector sum type of goods of the training sample of preset quantity；Respectively by above-mentioned training sample set
In each images of items in addition to the training sample of preset quantity input current Forecasting recognition model, according to above-mentioned Forecasting recognition
The vector of model output and the mapping relations of above-mentioned vector sum type of goods assess the validity of above-mentioned training sample；
The training sample that all validity are unsatisfactory for preset requirement is concentrated to delete from above-mentioned training sample and obtains new instruction
Practice sample set.
As an alternative embodiment, determining the article figure of the above-mentioned training sample of above-mentioned Forecasting recognition model output
As corresponding vector；
Determine that the corresponding type of goods of above-mentioned images of items is corresponding in the mapping relations of above-mentioned vector sum type of goods
Vector determines the validity of the training sample when being greater than preset value with the vector departure degree of above-mentioned Forecasting recognition model output
Meet preset requirement.
Preset requirement is met for above-mentioned validity, it should be appreciated that if above-mentioned two images of items (such as A and B) is right
When the vector departure degree answered is not more than preset value, that is, think that the corresponding type of goods of above-mentioned A and B is same class, if A is training
The training sample used when characteristic model then constructs the mapping of above-mentioned vector sum type of goods after obtaining Forecasting recognition model
When relationship, the mapping relations of above-mentioned A and its corresponding type of merchandize in the mapping relations of above-mentioned vector sum type of goods,
Therefore, the training sample comprising article B and its corresponding type of merchandize is the number of redundancy for training Forecasting recognition model
According to needing to delete the training sample comprising article B and its corresponding type of goods when carrying out the screening of above-mentioned training sample set
It removes；
In the above method, using the vector of the training sample of the preset quantity of above-mentioned Forecasting recognition model output, and in advance
If the corresponding type of goods of the training sample of quantity establishes the mapping relations of above-mentioned vector sum type of goods, and utilizes above-mentioned vector
The training sample that mapping relations screening with type of goods is concentrated for the training sample of training pattern, so that the feature trained
The precision of identification model is higher.
As an alternative embodiment, in the training using the above-mentioned above-mentioned training sample set of Forecasting recognition model discrimination
When sample, if the corresponding vector of images of items of the training sample of above-mentioned Forecasting recognition model output is not in above-mentioned vector sum article
When in the mapping relations of type, retain above-mentioned training sample.
As an alternative embodiment, after obtaining features described above identification model, respectively by the training of preset quantity
Images of items in sample inputs current feature identification model, extracts vector corresponding with above-mentioned images of items；
According to the above-mentioned images of items of the corresponding vector sum of above-mentioned images of items corresponding type of goods, structure in training sample
Build the mapping relations of above-mentioned vector sum type of goods.
As a kind of optional implementation, trained feature identification model and vector sum article kind are stated in use
When the type of the mapping relations identification article of class, for type of goods not in the mapping relations of above-mentioned vector sum type of goods
The article of new commodity type can be obtained by the way that the images of items of above-mentioned article is inputted above-mentioned trained feature identification model
To the corresponding vector of above-mentioned images of items, the corresponding vector of above-mentioned new article type that is to say, later based on obtained new article
The corresponding vector of type, in the mapping relations of above-mentioned vector sum type of goods, by the mapping of the vector sum new article type
Relationship is added, and the expansion of the mapping relations to above-mentioned vector sum type of goods can be realized, and realizes the article kind to identification
The expansion of class.
As shown in Fig. 2, a kind of entire flow of training characteristics identification model presented below, specifically includes:
Step 201, training sample set and test sample collection are obtained；
Step, 202, it is randomized the model parameter of deep learning network model, obtains initial Forecasting recognition model, it is above-mentioned
Forecasting recognition model includes multiple feature extraction network layer；
Step 203, the images of items in the training sample for preset quantity training sample concentrated inputs above-mentioned prediction and knows
Other model obtains the corresponding prediction type of goods of above-mentioned images of items；
Step 204, corresponding in training sample according to the prediction type of goods of above-mentioned images of items and above-mentioned images of items
Type of goods, adjust the model parameter of current Forecasting recognition model；
Step 205, the Forecasting recognition model after training is tested using the test sample that above-mentioned test sample is concentrated；
Step 206, judge whether test result meets default required precision, if so, entering step 207, otherwise enter step
209；
Step 207, terminate training process, the current Forecasting recognition model of the last one feature extraction network layer will be removed
Output is features described above identification model；
Step 208, according to the training sample of features described above identification model and above-mentioned preset quantity, constructs above-mentioned vector sum article
The mapping relations of type；
Step 209, using the training sample of current Forecasting recognition model and above-mentioned preset quantity, above-mentioned vector sum is constructed
The mapping relations of type of goods；
Step 210, using current Forecasting recognition model, according to the mapping relations of above-mentioned vector sum type of goods, screening
The sample data of above-mentioned training sample set obtains new training sample set, and enters step 203.
Embodiment two:
As shown in figure 3, the present embodiment provides the articles under a kind of public safety to identify equipment, which includes processor
301 and memory 302, wherein above-mentioned memory stores executable program code, when above procedure code is performed, so that
Above-mentioned processor executes following steps:
Obtain the images of items of article；
Above-mentioned images of items is inputted to the feature identification model constructed based on deep learning, extracts and is reflected in above-mentioned images of items
It penetrates the characteristic of type of goods and exports in vector form；
According to the mapping relations based on vector sum type of goods constructed by features described above identification model, features described above is obtained
The corresponding type of goods of vector of identification model output.
Above-mentioned processor is specifically used for, obtain include multiple training samples training sample set and including multiple test samples
Test sample collection, each training sample/test sample includes the images of items and the affiliated type of goods of above-mentioned article of article；
The model parameter of randomization deep learning network model obtains initial Forecasting recognition model, above-mentioned Forecasting recognition mould
Type includes multiple feature extraction network layers；
When trigger model training, using the training sample for the preset quantity that above-mentioned training sample is concentrated, current predictive is known
Other model is trained at least once, every time after training, after the test sample concentrated using above-mentioned test sample is to training
Forecasting recognition model tested, when determining that test result meets default required precision, terminate training process, will remove last
The current Forecasting recognition model output of one feature extraction network layer is features described above identification model.
Above-mentioned processor is specifically used for, and the images of items in the training sample of preset quantity is inputted to current Forecasting recognition
Model exports prediction type of goods corresponding with above-mentioned images of items；
According to the prediction type of goods of above-mentioned images of items and above-mentioned images of items in training sample corresponding article kind
Class adjusts the model parameter of current Forecasting recognition model by loss function.
Above-mentioned processor is also used to, and the training sample of above-mentioned training sample set is screened using current feature identification model,
Using the training sample set after screening as new training sample set, and trigger model training again.
Above-mentioned processor is specifically used for, and the images of items in the training sample of preset quantity is inputted to current prediction respectively
Identification model extracts vector corresponding with above-mentioned images of items, and constructs the vector sum object of the training sample of above-mentioned preset quantity
The mapping relations of kind class；
Each images of items in addition to the training sample of preset quantity is concentrated to input currently above-mentioned training sample respectively
Forecasting recognition model, according to the mapping relations of the vector of above-mentioned Forecasting recognition model output and above-mentioned vector sum type of goods
Assess the validity of the training sample；
The training sample that all validity are unsatisfactory for preset requirement is concentrated to delete from above-mentioned training sample and obtains new instruction
Practice sample set.
Above-mentioned processor is used for, and determines that the images of items of the above-mentioned training sample of above-mentioned Forecasting recognition model output is corresponding
Vector；
Determine that the corresponding type of goods of above-mentioned images of items is corresponding in the mapping relations of above-mentioned vector sum type of goods
Vector determines the validity of the training sample when being greater than preset value with the vector departure degree of above-mentioned Forecasting recognition model output
Meet preset requirement.
Above-mentioned processor is specifically used for, and the images of items in the training sample of preset quantity is inputted to current feature respectively
Identification model extracts vector corresponding with above-mentioned images of items；
According to the above-mentioned images of items of the corresponding vector sum of above-mentioned images of items corresponding type of goods, structure in training sample
Build the mapping relations of above-mentioned vector sum type of goods.
As shown in figure 4, the present embodiment provides the item identification devices under a kind of public safety, which includes:
Images of items acquisition unit 401, for obtaining the images of items of article；
Article characteristics recognition unit 402, for identifying the input of above-mentioned images of items based on the feature that deep learning constructs
Model extracts the characteristic for mapping type of goods in above-mentioned images of items and exports in vector form；
Type of goods recognition unit 403, for according to based on vector sum article kind constructed by features described above identification model
The mapping relations of class obtain the corresponding type of goods of vector of features described above identification model output.
Above-mentioned article characteristics recognition unit is used for, obtain include multiple training samples training sample set and including multiple surveys
The test sample collection of sample sheet, each training sample/test sample include the images of items and the affiliated article of above-mentioned article of article
Type；
The model parameter of randomization deep learning network model obtains initial Forecasting recognition model, above-mentioned Forecasting recognition mould
Type includes multiple feature extraction network layers；
When trigger model training, using the training sample for the preset quantity that above-mentioned training sample is concentrated, current predictive is known
Other model is trained at least once, every time after training, after the test sample concentrated using above-mentioned test sample is to training
Forecasting recognition model tested, when determining that test result meets default required precision, terminate training process, will remove last
The current Forecasting recognition model output of one feature extraction network layer is features described above identification model.
Above-mentioned article characteristics recognition unit is used for, and the images of items in the training sample of preset quantity is inputted currently pre-
Identification model is surveyed, prediction type of goods corresponding with above-mentioned images of items is exported；
According to the prediction type of goods of above-mentioned images of items and above-mentioned images of items in training sample corresponding article kind
Class adjusts the model parameter of current Forecasting recognition model by loss function.
Above-mentioned article characteristics recognition unit is also used to, and screens above-mentioned training sample set using current feature identification model
Sample data, using the training sample set after screening as new training sample set, and trigger model training again.
Above-mentioned article characteristics recognition unit is used for, and is respectively inputted the images of items in the training sample of preset quantity current
Forecasting recognition model, extract corresponding with above-mentioned images of items vector, and construct above-mentioned preset quantity training sample to
The mapping relations of amount and type of goods；
Each images of items in addition to the training sample of preset quantity is concentrated to input currently above-mentioned training sample respectively
Forecasting recognition model, according to the mapping relations of the vector of above-mentioned Forecasting recognition model output and above-mentioned vector sum type of goods
Assess the validity of the training sample；
The training sample that all validity are unsatisfactory for preset requirement is concentrated to delete from above-mentioned training sample and obtains new instruction
Practice sample set.
Above-mentioned article characteristics recognition unit is used for, and determines the article of the training sample of features described above identification model output
The corresponding vector of image；
Determine that the corresponding type of goods of above-mentioned images of items is corresponding in the mapping relations of above-mentioned vector sum type of goods
Vector determines the validity of the training sample when being greater than preset value with the vector departure degree of features described above identification model output
Meet preset requirement.
Above-mentioned article characteristics recognition unit is specifically used for, and respectively inputs the images of items in the training sample of preset quantity
Current feature identification model extracts vector corresponding with above-mentioned images of items；
According to the above-mentioned images of items of the corresponding vector sum of above-mentioned images of items corresponding type of goods, structure in training sample
Build the mapping relations of above-mentioned vector sum type of goods.
Embodiment three:
The embodiment of the present invention also provides a kind of computer-readable non-volatile memory medium, including program code, when described
For program code when running on computing terminal, said program code is for making the computing terminal execute the embodiments of the present invention
The step of one method.
Above by reference to showing according to the method, apparatus (system) of the embodiment of the present application and/or the frame of computer program product
Figure and/or flow chart describe the application.It should be understood that can realize that block diagram and or flow chart is shown by computer program instructions
The combination of the block of a block and block diagram and or flow chart diagram for figure.These computer program instructions can be supplied to logical
With computer, the processor of special purpose computer and/or other programmable data processing units, to generate machine, so that via meter
The instruction that calculation machine processor and/or other programmable data processing units execute creates for realizing block diagram and or flow chart block
In specified function action method.
Correspondingly, the application can also be implemented with hardware and/or software (including firmware, resident software, microcode etc.).More
Further, the application can take computer usable or the shape of the computer program product on computer readable storage medium
Formula has the computer realized in the medium usable or computer readable program code, to be made by instruction execution system
It is used with or in conjunction with instruction execution system.In the present context, computer can be used or computer-readable medium can be with
It is arbitrary medium, may include, stores, communicates, transmits or transmit program, is made by instruction execution system, device or equipment
With, or instruction execution system, device or equipment is combined to use.
Obviously, various changes and modifications can be made to the invention without departing from essence of the invention by those skilled in the art
Mind and range.In this way, if these modifications and changes of the present invention belongs to the range of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to include these modifications and variations.
Claims (9)
1. the item identification method under a kind of public safety, which is characterized in that this method comprises:
Obtain the images of items of article；
The images of items is inputted to the feature identification model constructed based on deep learning, extracts in the images of items and maps object
The characteristic of kind class simultaneously exports in vector form；
According to the mapping relations based on vector sum type of goods constructed by the feature identification model, the feature identification is obtained
The corresponding type of goods of vector of model output.
2. the method as described in claim 1, which is characterized in that be based on deep learning construction feature identification model, comprising:
Obtain includes the training sample set of multiple training samples and the test sample collection including multiple test samples, each trained sample
Sheet/test sample includes the images of items and the affiliated type of goods of the article of article；
The model parameter of randomization deep learning network model obtains initial Forecasting recognition model, the Forecasting recognition model packet
Include multiple feature extraction network layers；
When trigger model training, using the training sample for the preset quantity that the training sample is concentrated, mould is identified to current predictive
Type is trained at least once, and every time after training, the test sample concentrated using the test sample is to pre- after training
It surveys identification model to be tested, when determining that test result meets default required precision, terminates training process, the last one will be removed
The current Forecasting recognition model output of feature extraction network layer is the feature identification model.
3. method according to claim 2, which is characterized in that the instruction for the preset quantity concentrated every time using the training sample
Practice sample, current predictive identification model be trained, comprising:
Images of items in the training sample of preset quantity is inputted to current Forecasting recognition model, output and the images of items
Corresponding prediction type of goods；
According to the prediction type of goods of the images of items and the images of items in training sample corresponding type of goods, lead to
Cross the model parameter that loss function adjusts current Forecasting recognition model.
4. method according to claim 2, which is characterized in that every time after training, determine that test result is unsatisfactory for presetting
Required precision, further includes:
Using the training sample of training sample set described in current Forecasting recognition model discrimination, the training sample set after screening is made
For new training sample set, and trigger model training again.
5. method as claimed in claim 4, which is characterized in that utilize training sample described in current Forecasting recognition model discrimination
The training sample of collection, using the training sample set after screening as new training sample set, comprising:
Images of items in the training sample of preset quantity is inputted to current Forecasting recognition model respectively, is extracted and the article
The corresponding vector of image；
Construct the mapping relations of the corresponding vector sum type of goods of images of items in the training sample of preset quantity；
Each images of items in addition to the training sample of preset quantity is concentrated to input current prediction the training sample respectively
Identification model is assessed according to the mapping relations of the vector of Forecasting recognition model output and the vector sum type of goods
The validity of the training sample；
All validity are unsatisfactory for the training sample of preset requirement, concentrate deletion to obtain new training sample from the training sample
This collection.
6. method as claimed in claim 5, which is characterized in that according to the vector that current Forecasting recognition model exports, and
The mapping relations of the vector sum type of goods assess the validity of the training sample, comprising:
Determine the corresponding vector of images of items of the training sample of current feature identification model output；
Determine the corresponding type of goods of the images of items corresponding vector in the mapping relations of the vector sum type of goods,
When being greater than preset value with the vector departure degree that current Forecasting recognition model exports, determine that the validity of the training sample meets
Preset requirement.
7. method as claimed in claim 1 or 3, which is characterized in that construct vector sum article based on the feature identification model
The mapping relations of type, comprising:
Images of items in the training sample of preset quantity is inputted to current feature identification model respectively, is extracted and the article
The corresponding vector of image；
According to images of items described in the corresponding vector sum of the images of items in training sample corresponding type of goods, construct institute
State the mapping relations of vector sum type of goods.
8. the article under a kind of public safety identifies equipment, which is characterized in that the equipment includes processor and memory, wherein
The memory stores executable program code, when said program code is performed, so that the processor executes following step
It is rapid:
Obtain the images of items of article；
The images of items is inputted to the feature identification model constructed based on deep learning, extracts in the images of items and maps object
The characteristic of kind class simultaneously exports in vector form；
According to the mapping relations based on vector sum type of goods constructed by the feature identification model, the feature identification is obtained
The corresponding type of goods of vector of model output.
9. a kind of computer can storage medium, be stored thereon with computer program, which is characterized in that the program is executed by processor
The step of Shi Shixian such as claim 1~7 any the method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910101328.5A CN109858448A (en) | 2019-01-31 | 2019-01-31 | Item identification method and equipment under a kind of public safety |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910101328.5A CN109858448A (en) | 2019-01-31 | 2019-01-31 | Item identification method and equipment under a kind of public safety |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109858448A true CN109858448A (en) | 2019-06-07 |
Family
ID=66897320
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910101328.5A Withdrawn CN109858448A (en) | 2019-01-31 | 2019-01-31 | Item identification method and equipment under a kind of public safety |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109858448A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110378303A (en) * | 2019-07-25 | 2019-10-25 | 杭州睿琪软件有限公司 | Method and system for Object identifying |
| CN110490086A (en) * | 2019-07-25 | 2019-11-22 | 杭州睿琪软件有限公司 | A kind of method and system for Object identifying result secondary-confirmation |
-
2019
- 2019-01-31 CN CN201910101328.5A patent/CN109858448A/en not_active Withdrawn
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110378303A (en) * | 2019-07-25 | 2019-10-25 | 杭州睿琪软件有限公司 | Method and system for Object identifying |
| CN110490086A (en) * | 2019-07-25 | 2019-11-22 | 杭州睿琪软件有限公司 | A kind of method and system for Object identifying result secondary-confirmation |
| CN110378303B (en) * | 2019-07-25 | 2021-07-09 | 杭州睿琪软件有限公司 | Method and system for object recognition |
| CN110490086B (en) * | 2019-07-25 | 2021-08-20 | 杭州睿琪软件有限公司 | Method and system for secondarily confirming object recognition result |
| US11335087B2 (en) | 2019-07-25 | 2022-05-17 | Hangzhou Glority Software Limited | Method and system for object identification |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108922026B (en) | Replenishment management method and device for vending machine and user terminal | |
| CN109190705A (en) | Self-service method, apparatus and system | |
| US10380546B2 (en) | Intelligent display and self-serve experience system and method | |
| CN108537166B (en) | Method and device for determining shelf browsing amount and analyzing browsing amount | |
| US10628695B2 (en) | Fast item identification for checkout counter | |
| CN109886169B (en) | Article identification method, device, equipment and storage medium applied to unmanned container | |
| CN105893465A (en) | Automatic question answering method and device | |
| CN108805644A (en) | The commercial articles vending method and machine for vending of machine for vending | |
| WO2019127618A1 (en) | Settlement method, device and system | |
| CN107798027A (en) | A kind of heatrate Forecasting Methodology, information recommendation method and device | |
| CN112508132B (en) | Training method and device for identifying SKU | |
| WO2024027624A1 (en) | Metaverse-based intelligent resource matching method, and metaverse system | |
| CN109712324A (en) | A kind of automatic vending machine image-recognizing method, good selling method and vending equipment | |
| CN110119914A (en) | Object replenishes processing method, device and system | |
| Sarwar et al. | Smart shopping carts based on mobile computing and deep learning cloud services | |
| CN109544262A (en) | Item recommendation method, device, electronic equipment, system and readable storage medium storing program for executing | |
| CN109190706A (en) | Self-service method, apparatus and system | |
| CN113947425A (en) | Shop address determining method and device | |
| CN109858448A (en) | Item identification method and equipment under a kind of public safety | |
| CN109934569A (en) | Settlement method, device and system | |
| Maizi et al. | Building a digital twin for IoT smart stores: A case in retail and apparel industry | |
| CN110119915A (en) | Object enters library processing method, device and system | |
| CN113888254A (en) | Shelf commodity management method and electronic equipment | |
| CN111126264A (en) | Image processing method, device, equipment and storage medium | |
| CN113887884A (en) | Business-super service system |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190607 |
|
| WW01 | Invention patent application withdrawn after publication |