CN109544496A - Generation method, the training method and device of object detection model of training data - Google Patents
Generation method, the training method and device of object detection model of training data Download PDFInfo
- Publication number
- CN109544496A CN109544496A CN201811382077.4A CN201811382077A CN109544496A CN 109544496 A CN109544496 A CN 109544496A CN 201811382077 A CN201811382077 A CN 201811382077A CN 109544496 A CN109544496 A CN 109544496A
- Authority
- CN
- China
- Prior art keywords
- data
- training
- image
- diagram data
- stingy
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration by the use of more than one image, e.g. averaging, subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
Abstract
The present invention provides a kind of generation method of training data, the training method of object detection model and devices；Wherein, the generation method of the training data includes: to obtain comprising there are many image datas of object；A variety of objects are taken from image data, the stingy diagram data of the object taken；According to preset synthetic parameters, the stingy diagram data of the object taken is blended on corresponding background image, composograph is obtained；Wherein, synthetic parameters include one of the type for scratching the corresponding object of diagram data, the dimensional information for scratching diagram data, rotation angle, coverage extent and background image or a variety of；Using composograph as the training data of object detection model.The present invention can be synthesized to obtain a large amount of composograph by a small amount of image data by adjusting a variety of synthetic parameters, to improve the convenience for obtaining a large amount of training datas, reduce the human cost to training data mark processing.
Description
Technical field
The present invention relates to object detection technique fields, generation method, object detection more particularly, to a kind of training data
The training method and device of model.
Background technique
Object detection (object detection) is the key areas of computer vision.More popular object detection side
There are two types of methods, and one is using Faster R-CNN as two-stage (two-stage) method of representative, another kind is using YOLO as generation
One stage of table (one-stage) method.However, which kind of method for checking object to require instruction magnanimity and with mark regardless of
Practice data to be trained model, can just obtain the higher object detection model of detection effect.
In practical applications, obtain magnanimity and the training data with mark be not easy to.It, can by taking commodity detect as an example
Directly to arrive supermarket or certain specific commodity selling place (such as pharmacy) shootings, but this mode feasibility is not high；?
Commodity can be bought back, be shot in laboratory environments.But since the classification of commodity is very more, a pharmacy may just have up to ten thousand
Kind drug, and the arrangement of these drugs spatially has infinite a variety of possibility, so wanting in laboratory environment as far as possible
Possible arrangement variation, needs to expend more manpower and time, mode is relatively complicated in exhaustive practical application scene.In addition,
Even if having got a large amount of image by way of shooting mass picture, information labeling is carried out to these images and is also needed greatly
The manpower of amount and time, cost are higher.
Summary of the invention
In view of this, the purpose of the present invention is to provide a kind of generation methods of training data, the instruction of object detection model
Practice method and apparatus, to improve the convenience for obtaining a large amount of training datas, and reduce to the manpower of training data mark processing at
This.
In a first aspect, the embodiment of the invention provides a kind of generation method of training data, this method comprises: acquisition includes
There are many image datas of object；Wherein, every kind of object is labeled with the type of object；It is a variety of right to take from image data
As the stingy diagram data of the object taken；According to preset synthetic parameters, the stingy diagram data of the object taken is blended into pair
On the background image answered, composograph is obtained；Wherein, synthetic parameters include the type for scratching the corresponding object of diagram data, scratch figure number
According to dimensional information, rotation angle, one of coverage extent and background image or a variety of；Using composograph as object detection
The training data of model.
In preferred embodiments of the present invention, above-mentioned acquisition include there are many object image data the step of: obtain from
Multiple images of the object of multiple angle shots.
In preferred embodiments of the present invention, if including the dimensional information for scratching diagram data in synthetic parameters, by what is taken
The stingy diagram data of object is blended into the step on corresponding background image, comprising: according to dimensional information, adjusts the object taken
Scratch the scale of diagram data；Wherein, the stingy diagram data scale being blended on same background image is identical；By scratching after rescaling
Diagram data is blended on corresponding background image.
In preferred embodiments of the present invention, if including the rotation angle for scratching diagram data in synthetic parameters, by what is taken
The stingy diagram data of object is blended into the step on corresponding background image, comprising: revolves the stingy diagram data of each object taken
Go to corresponding rotation angle；Postrotational stingy diagram data is blended on corresponding background image.
In preferred embodiments of the present invention, the stingy diagram data of the above-mentioned object that will be taken is blended into corresponding background image
On step, comprising: by random manner, position arrangement is carried out to the stingy diagram data of the object taken, will be scratched after arrangement
Diagram data is blended on corresponding background image.
In preferred embodiments of the present invention, if including the coverage extent for scratching diagram data in synthetic parameters, to what is taken
After the stingy diagram data of object carries out the step of position arrangement, method further include: according to coverage extent, judge the object taken
Scratching diagram data whether there is the stingy diagram data that should be blocked；If so, the mobile stingy diagram data being blocked is to predeterminated position；?
On predeterminated position, the stingy diagram data being blocked is hidden by the stingy diagram data in addition to the stingy diagram data being blocked according to coverage extent
Gear.
In preferred embodiments of the present invention, using composograph as the step of the training data of object detection model it
Afterwards, method includes: whether the quantity of training of judgement data is greater than or equal to preset amount threshold；If not, continue to execute by
According to preset synthetic parameters, the stingy diagram data of the object taken is blended into the step on corresponding background image, until training
The quantity of data is greater than or equal to amount threshold.
In preferred embodiments of the present invention, above-mentioned object is rigid body object.
Second aspect, the embodiment of the invention provides a kind of training method of object detection model, method includes: acquisition pair
The training data of elephant；Training data is generated by the generation method of above-mentioned training data；Training data is input to preset first
It is trained in beginning model, obtains object detection model.
In preferred embodiments of the present invention, above-mentioned be input to training data is trained in preset initial model
After step, method includes: acquisition test data；Test data includes that multiple include the image data of object；Number will be tested
According in the initial model being input to after training, output test result；The position letter of object is labeled in test result comprising multiple
The image data of breath and type；From the location information and the correct image data of type of test result screening object；Pass through sieve
The image data selected is finely adjusted processing to the initial model after training, obtains final object detection model.
The third aspect, the embodiment of the invention provides a kind of generating means of training data, which includes: image data
Module is obtained, for obtaining comprising there are many image datas of object；Wherein, every kind of object is labeled with the type of object；It is right
As taking module, for taking a variety of objects from image data, the stingy diagram data of the object taken；Image synthesizes mould
Block, for the stingy diagram data of the object taken being blended on corresponding background image, is closed according to preset synthetic parameters
At image；Wherein, synthetic parameters include scratch the corresponding object of diagram data type, scratch diagram data dimensional information, rotation angle,
One of coverage extent and background image are a variety of；Using composograph as the training data of object detection model.
Fourth aspect, the embodiment of the invention provides a kind of training device of object detection model, which includes: training
Data acquisition module, for obtaining the training data of object；Training data is generated by the generation method of above-mentioned training data；Instruction
Practice module, is trained for training data to be input in preset initial model, obtains object detection model.
5th aspect, the embodiment of the invention provides a kind of electronic system, electronic system includes: image capture device, place
Manage equipment and storage device；Image capture device, for obtaining preview video frame or image data；Meter is stored on storage device
Calculation machine program, computer program executes the generation method such as above-mentioned training data when equipment processed is run, or executes such as
The training method of above-mentioned object detection model.
6th aspect, the embodiment of the invention provides a kind of computer readable storage medium, computer readable storage mediums
On be stored with computer program, when computer program equipment operation processed, executes the generation method such as above-mentioned training data, or
Person executes the step of training method such as above-mentioned object detection model.
The embodiment of the present invention bring it is following the utility model has the advantages that
The generation method of training data provided in an embodiment of the present invention, the training method of object detection model, device, electronics
System and computer readable storage medium are got comprising being taken from the image data there are many after the image data of object
A variety of objects, the stingy diagram data of the object taken；And then according to preset synthetic parameters, by the stingy figure number of the object taken
According to being blended on corresponding background image, composograph is obtained, to obtain the training data of object detection model；Which
In, by adjusting a variety of synthetic parameters, can be synthesized to obtain a large amount of composograph by a small amount of image data, to improve
Obtain the convenience of a large amount of training datas.In addition, since image has before synthesis marked each object, thus nothing
Processing need to be labeled to composograph energetically again, reduce the human cost to training data mark processing.
Other features and advantages of the present invention will illustrate in the following description, also, partly become from specification
It obtains it is clear that understand through the implementation of the invention.The objectives and other advantages of the invention are in specification, claims
And specifically noted structure is achieved and obtained in attached drawing.
To enable the above objects, features and advantages of the present invention to be clearer and more comprehensible, preferred embodiment is cited below particularly, and cooperate
Appended attached drawing, is described in detail below.
Detailed description of the invention
It, below will be to specific in order to illustrate more clearly of the specific embodiment of the invention or technical solution in the prior art
Embodiment or attached drawing needed to be used in the description of the prior art be briefly described, it should be apparent that, it is described below
Attached drawing is some embodiments of the present invention, for those of ordinary skill in the art, before not making the creative labor
It puts, is also possible to obtain other drawings based on these drawings.
Fig. 1 is a kind of structural schematic diagram of electronic system provided in an embodiment of the present invention；
Fig. 2 is a kind of flow chart of the generation method of training data provided in an embodiment of the present invention；
Fig. 3 is a kind of flow chart of the training method of object detection model provided in an embodiment of the present invention；
Fig. 4 is a kind of structural schematic diagram of the generating means of training data provided in an embodiment of the present invention；
Fig. 5 is a kind of structural schematic diagram of the training device of object detection model provided in an embodiment of the present invention.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with attached drawing to the present invention
Technical solution be clearly and completely described, it is clear that described embodiments are some of the embodiments of the present invention, rather than
Whole embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art are not making creative work premise
Under every other embodiment obtained, shall fall within the protection scope of the present invention.
In view of the training data of existing object detection model, acquisition modes are cumbersome and the cost of information labeling is higher
Problem, the embodiment of the invention provides a kind of generation method of training data, the training method of object detection model, device and electricity
Subsystem；The technology can be applied in the multiple terminals equipment such as server, computer, mobile phone, tablet computer, and to medicine
In the detection of a variety of commodity such as product, food, tobacco and wine, daily necessities, books；The technology can be used corresponding software and hardware and realize,
It describes in detail below to the embodiment of the present invention.
Embodiment one:
Firstly, describing generation method, the object detection mould of the training data for realizing the embodiment of the present invention referring to Fig.1
Training method, the example electronic system 100 of device and electronic system of type.
A kind of structural schematic diagram of electronic system as shown in Figure 1, electronic system 100 include one or more processing equipments
102, one or more storage devices 104, input unit 106, output device 108 and one or more image capture devices
110, these components pass through the interconnection of bindiny mechanism's (not shown) of bus system 112 and/or other forms.It should be noted that Fig. 1 institute
The component and structure for the electronic system 100 shown be it is illustrative, and not restrictive, as needed, the electronic system
It can have other assemblies and structure.
The processing equipment 102 can be gateway, or intelligent terminal, or include central processing unit
It (CPU) or the equipment of the processing unit of the other forms with data-handling capacity and/or instruction execution capability, can be to institute
The data for stating other components in electronic system 100 are handled, and other components in the electronic system 100 can also be controlled
To execute desired function.
The storage device 104 may include one or more computer program products, and the computer program product can
To include various forms of computer readable storage mediums, such as volatile memory and/or nonvolatile memory.It is described easy
The property lost memory for example may include random access memory (RAM) and/or cache memory (cache) etc..It is described non-
Volatile memory for example may include read-only memory (ROM), hard disk, flash memory etc..In the computer readable storage medium
On can store one or more computer program instructions, processing equipment 102 can run described program instruction, to realize hereafter
The client functionality (realized by processing equipment) in the embodiment of the present invention and/or other desired functions.Institute
Various application programs and various data can also be stored by stating in computer readable storage medium, such as the application program uses
And/or various data generated etc..
The input unit 106 can be the device that user is used to input instruction, and may include keyboard, mouse, wheat
One or more of gram wind and touch screen etc..
The output device 108 can export various information (for example, image or sound) to external (for example, user), and
It and may include one or more of display, loudspeaker etc..
Described image acquisition equipment 110 can acquire preview video frame or image data, and collected preview is regarded
Frequency frame or image data are stored in the storage device 104 for the use of other components.
Illustratively, for realizing the generation method of training data according to an embodiment of the present invention, object detection model
Each device in the example electronic system of training method, device and electronic system can integrate setting, can also be all with scattering device
It is such as that processing equipment 102, storage device 104, input unit 106 and output device 108 is integrally disposed in one, and image is adopted
Collection equipment 110 is set to the designated position that can collect target image.When each device in above-mentioned electronic system is integrally disposed
When, which may be implemented as the intelligent terminals such as camera, smart phone, tablet computer, computer.
Embodiment two:
A kind of generation method of training data is present embodiments provided, this method is by the processing equipment in above-mentioned electronic system
It executes；The processing equipment can be any equipment or chip with data-handling capacity.The processing equipment can be docked independently
The information received is handled, and can also be connected with server, is analyzed and processed jointly to information, and processing result is uploaded
To cloud.
As shown in Fig. 2, the generation method of the training data includes the following steps:
Step S202 is obtained comprising there are many image datas of object；Wherein, every kind of object is labeled with the kind of object
Class；
Training data in the present embodiment can be used for the training of the detection model of rigid body object or quasi- rigid body object；Rigid body
Object can be understood as cannot actively the object that deformation occurs, the visual cosmetic variation of rigid body object is due to putting angle mostly
The objective factors such as degree, ambient light are shone, background changes cause.Quasi- rigid body object can be understood as the limited object of deformation, such as walk
Pedestrian under state；And for the people under other states, animal these objects, can actively deformation occurs, and deformation is more acute
It is strong, it is difficult to which that by changing the deformation of all kinds of possibilities of objective factor simulated object, thus the training data in the present embodiment is more
Add the training of the detection model suitable for above-mentioned rigid body object or quasi- rigid body object.
It can only include a picture in the image data of above-mentioned acquisition, include a variety of objects in the picture；Also it can wrap
Plurality of pictures is included, includes one or more object in every picture；If be appreciated that in image data only comprising a kind of right
As obtaining a large amount of different pictures of the object after being replaced background, adjustment scale, rotation angle etc. reason to the object；
But since these pictures include same object, the training data of these pictures composition has stronger tendentiousness, easily leads to logical
When crossing other kind of class object of model inspection that the training data trains, poor effect；Thus, no matter picture in image data
Quantity is how many, is required in image data comprising there are many objects.
It should be noted that can have biggish difference between different types of object, it is possible to have subtleer difference
It is different.Such as beverage, cigarette, be properly termed as different types of object between drug, for another example cold drug, dermic, between medicine for enterogastritis
Be referred to as different types of object, for another example A brand cold drug, B brand cold drug, can also claim between C brand cold drug
It is third for cold drug, the cold drug that main component is second and main component that different types of object or main component are first
Cold drug between be referred to as different types of object.It specifically can be true according to the detection range of detection model to be trained
Which object fixed above-mentioned a variety of objects specifically include.
After getting above-mentioned image data, the kind of every kind of object marking object can be directed to by way of manual identification
Class；In actual implementation, the position of object can be determined by callout box, then input the type of the corresponding object of the callout box；
The callout box can be fixed shape, such as rectangle, circle, ellipse, and object can also be obtained by way of edge detection
Edge lines, using edge lines as the callout box of object.Alternatively, it is also possible to pass through marking software automatically to above-mentioned picture number
Object in is labeled, after the completion of mark, then by manually being manually adjusted.
Step S204 takes a variety of objects from image data, the stingy diagram data of the object taken；
Take object, it is understood that from image data to be partitioned into the corresponding image-region of object from image data
Process, which can pass through object detection method FPN (Feature Pyramid Network, feature pyramid network)
It realizes；Image data can be specifically input in trained FPN model, export the image district of each object in the image data
Domain, the i.e. stingy diagram data of object.
In a further mode of operation, it is contemplated that the corresponding image-region of object and background area are generally comprised in image data,
Wherein, relative to background area, the corresponding image-region of object is usually foreground area；Based on the difference, image can be passed through
Masking-out channel in data obtains the masking-out image of the image data, due to the picture of foreground area and background area in masking-out image
Plain value differs greatly, thus the edge of available foreground area, and then obtains the edge of the corresponding image-region of object, the side
The image-region that edge is surrounded is the stingy diagram data of object.
It is, of course, also possible to object be taken from above-mentioned image data by scratching figure software (such as Photoshop), by scratching
Various stingy figure tools in figure software, if figure is scratched in channel, mask scratches figure, and pen tool scratches figure etc., and it is each right in image data to obtain
The stingy diagram data of elephant.
The stingy diagram data of the object taken is blended into corresponding Background according to preset synthetic parameters by step S206
As upper, composograph is obtained；Wherein, which includes the type for scratching the corresponding object of diagram data, the scale for scratching diagram data
One of information, rotation angle, coverage extent and background image are a variety of.
Specifically, for the type of the corresponding object of stingy diagram data in synthetic parameters, can be used for characterizing to be synthesized
The type for the object for including in composograph, the type for such as scratching the corresponding object of diagram data includes object 1, object 2 and object 3,
It then include object 1, object 2 and object 3 totally three kinds of objects in corresponding composograph；This scratches the type of the corresponding object of diagram data
Can also only include number of species, such as three kinds, then in corresponding composograph can at random or according to certain rule from take to
Stingy diagram data in select three kinds of objects.
The dimensional information of stingy diagram data in above-mentioned synthetic parameters, can be used for characterizing the object for including in composograph
Scale.For the detection model that convolutional neural networks are built, the full articulamentum in convolutional neural networks usually requires that the figure of input
As being fixed size.And size and shooting distance, the shooting angle by object itself are influenced, the stingy figure number of above-mentioned object
It may be of different sizes according to scale.The each of convolutional neural networks is input in order to be conveniently adjusted in the training process of detection model
The scale of the corresponding image-region of object generally requires before synthesizing above-mentioned composograph, will be each in same composograph
The stingy diagram data of a object is adjusted to identical scale, but the scale not necessarily requires the fixed ruler with above-mentioned full articulamentum requirement
It spends identical.The dimensional information of stingy diagram data in above-mentioned synthetic parameters can be used to provide each right in current composograph
The scale of the stingy diagram data of elephant.In addition, can have between different composographs different to improve the rich of scale
Scale parameter, to improve detection robustness of the detection model in terms of scale.
Rotation angle in above-mentioned synthetic parameters, can be used for characterizing the shooting angle for the object for including in composograph；
It in actual implementation, may only include the image data of the limited angle of object in the image data got；By by object
Stingy diagram data rotate to above-mentioned rotation angle degree, the rich of the shooting angle of object can be increased.The rotation angle can be
Rotation angle in two-dimensional surface, or three-dimensional rotation angle in three-dimensional space.It is each right in same composograph
The rotation angle of elephant may be the same or different.
Coverage extent in above-mentioned synthetic parameters, may include multiple parameters, as include in composograph object whether
The position and occlusion area be blocked, blocked account for the percentage etc. of the total stingy diagram data of the object；By default, same
The each object in composograph is opened not block mutually each other；If symbolized in above-mentioned coverage extent certain an object need by
It blocks, occluded object can be carried out by other objects blocking processing, the position of other such as mobile objects, or movement should
The position of occluded object, so that the situation of blocking of the occluded object meets parameters specified in above-mentioned coverage extent.
Background image in above-mentioned synthetic parameters can be used for characterizing background image used in current composograph.It can
To obtain a variety of background images in advance, when needing composograph, it is random or according to preset rules from the Background obtained in advance
Corresponding background image is selected as in.
During synthesizing above-mentioned composograph, a part of parameter in above-mentioned a variety of synthetic parameters can be used only,
Other parameters do not use or use its default value.In the process being blended into the stingy diagram data of object on corresponding background image
In, specifically can pixel value on the position of background image locating for the stingy diagram data by object, be substituted for the stingy figure number of object
Pixel value in, to realize the synthesis process of image.
Step S208, using above-mentioned composograph as the training data of object detection model.
In general, the training effect in order to guarantee object detection model, it usually needs a large amount of training data is trained；Cause
And a large amount of different composograph each other can be obtained by changing synthetic parameters, the set of these composographs is
It can be used as the training data of object detection model.
The generation method of training data provided in an embodiment of the present invention is got comprising there are many image datas of object
Afterwards, a variety of objects are taken from the image data, the stingy diagram data of the object taken；And then join according to preset synthesis
Number, the stingy diagram data of the object taken is blended on corresponding background image, composograph is obtained, to obtain object detection
The training data of model；In which, by adjusting a variety of synthetic parameters, it can be synthesized to obtain by a small amount of image data a large amount of
Composograph, to improve the convenience for obtaining a large amount of training datas.
In addition, since image has before synthesis marked each object, thus no longer need to synthesis energetically
Image is labeled processing, reduces the human cost to training data mark processing.
Embodiment three:
The generation method of another training data is present embodiments provided, this method is real on the basis of the above embodiments
It is existing；In above-described embodiment, it is noted that synthetic parameters include the type for scratching the corresponding object of diagram data, the scale letter for scratching diagram data
One of breath, rotation angle, coverage extent and background image are a variety of；In the present embodiment, emphasis description passes through various synthesis
The detailed process of the stingy diagram data of the multiple objects of parameter processing.The treatment process of every kind of synthetic parameters is retouched respectively below
It states.
(1) type about the corresponding object of stingy diagram data
It may include the object of one or more types for a composograph, therefore, when image is opened in synthesis one,
Pair that can include from the image for selecting this to be synthesized in a variety of objects got randomly or according to default rule
The type of elephant.For example, a variety of objects got share ten kinds, respectively object 1 to object 10；It is random from this ten kinds of objects
Ground selecting object, the quantity of object type and specific type can be randomly selected.
In addition, can preset every opening and closing with image should include if having preset the quantity of composograph
Object type, such as preset composograph are 100, wherein 10 composographs include 2 objects, 10 composograph packets
Containing 3 objects, and so on.It, can be sequentially from above-mentioned object 1 as the object in every image specifically including which type
It is selected to object 10, e.g., for composograph A, selecting object 1 and object 2, for composograph B, selecting object 3 and object
4, currently specific object can also be selected according to preset interval.
(2) dimensional information about stingy diagram data
If adjusting the object taken according to the dimensional information including scratching the dimensional information of diagram data in synthetic parameters
Scratch the scale of diagram data；Stingy diagram data after rescaling is blended on corresponding background image.Wherein, it is blended into same
Stingy diagram data scale on background image is identical.
In view of training object detection model needs a large amount of training data, thus a dimensional information can be pre-established
Set or dimensional information range, include a variety of scales in the dimensional information set or range, when one image of every synthesis, from the ruler
It spends random in information aggregate or sequentially obtains a dimensional information, can also believe according to preset scale interval from above-mentioned scale
A dimensional information is obtained in breath, and the scale for scratching diagram data is adjusted with the dimensional information got.Since convolutional neural networks are taken
In the detection model built, it is fixed size that full articulamentum, which requires the image of input, for the ease of in same composograph of adjustment
The scale of the corresponding image-region of each object, generally requires the stingy diagram data being blended on same background image being adjusted to phase
With scale, but the scale not necessarily require it is identical as the fixed size of above-mentioned full articulamentum requirement.
If the scale of the stingy diagram data of object is less than scale specified in above-mentioned dimensional information, need to scratch this figure number
According to being stretched, stretched location of pixels can be filled in drawing process by way of interpolation；If the stingy figure number of object
According to scale be greater than above-mentioned dimensional information specified in scale, then need to scratch this diagram data carry out compression or screenshot processing；Its
In, in compression process, pixel column or pixel column can be deleted according to preset interval, thus to obtain meeting scratching for dimensional information
Diagram data.
(3) about the rotation angle of stingy diagram data
If in synthetic parameters include scratch diagram data rotation angle, by the stingy diagram data of each object taken rotate to
Corresponding rotation angle；Postrotational stingy diagram data is blended on corresponding background image.
It is similar with above-mentioned dimensional information, a rotation angle set or rotation angle range, the rotation can be pre-established
It include a variety of rotation angles in angle set or range, it is random or suitable from the dimensional information set when one image of every synthesis
One or more rotation angles are obtained to sequence, one can also be obtained from above-mentioned rotation angle range according to preset angle interval
A or multiple rotation angles, to adjust the rotation angle of multiple objects of to be synthesized to one image.It is appreciated that same synthesis
The rotation angle of object in image may be the same or different.Above-mentioned rotation angle can be the rotation under two-dimensional surface
Angle, for example, rotating clockwise 10 degree, 90 degree, 20 degree of rotation etc. counterclockwise；Or the rotation angle under three-dimensional space,
10 degree are such as rotated clockwise in x/y plane, are rotated by 90 ° in z-axis direction.
(4) coverage extent about stingy diagram data
It is above-mentioned when stingy diagram data is blended into corresponding background image, it usually needs to preset each stingy diagram data in background
Position on image；One way in which is, by random manner, carries out position row to the stingy diagram data of the object taken
Column, then the stingy diagram data after arrangement is blended on corresponding background image.It, specifically can be according to right in the alignment processes of position
The stingy diagram data of each object is evenly distributed stingy diagram data by the quantity of elephant.As for the neighbouring relations between object, i.e., which is right
Which, as adjacent with object, can be arranged by random manner.
It is above-mentioned install in place during, can be not blocked mutually between default objects, therefore initial position arranges knot
Fruit is that the stingy diagram data complementation of each object is blocked；And if in synthetic parameters including the coverage extent for scratching diagram data, basis
Coverage extent judges that the stingy diagram data of the object taken whether there is the stingy diagram data that should be blocked；If so, movement is hidden
The stingy diagram data of gear is to predeterminated position；On predeterminated position, the stingy diagram data being blocked is by addition to the stingy diagram data being blocked
Stingy diagram data, blocked according to coverage extent.
Above-mentioned coverage extent, each object that can be treated in composograph are configured respectively；For example, can build in advance
The parameter matrix of a vertical coverage extent, the parameter matrix may include the letter for blocking percentage and blocking two dimensions in direction
Breath；Wherein, blocking percentage may be a range；Block direction can for left side, right side, upside, downside, upper left side,
Lower right side, or more fine-grained direction divides, and side direction as above rotates clockwise 10 degree of direction etc..Every synthesis one
When image, it is directed to each object respectively, corresponding parameter is selected from the parameter matrix, for example, object A is corresponding to block percentage
Than being zero, illustrate the stingy diagram data of object A without being blocked, the corresponding percentage that blocks of object B is 50%, and blocking direction is
Upside then illustrates that object B needs are blocked, and it is right to block this by object A or other objects for being not necessarily to be blocked at this time
As B.Specifically, the stingy figure number of the occluded object can be moved after each object is arranged in the form that do not block mutually
According to the stingy diagram data of object B is moved near the stingy diagram data of object A by the i.e. stingy diagram data of object B.Due to coverage extent
In block direction be upside, illustrate to need to be blocked on the upside of the stingy diagram data of object B, at this time can be by the stingy figure of object B
Data are moved to the downside of the stingy diagram data of object A, and continue to move up, until object A hides according to the above-mentioned percentage that blocks
The stingy diagram data of object B is kept off, object B reaches above-mentioned predeterminated position at this time.
In another implementation, it can be hidden before each object is arranged in background image for each object setting
Gear degree carries out position arrangement according still further to stingy diagram data of the coverage extent to each object.
It should be noted that for the ease of the arrangement of each object, when can preset one image of every synthesis, the composite diagram
As at least part in corresponding all objects is without being blocked, the position of object is excessively caused to avoid the object being blocked
The case where arrangement is excessively complicated or cannot achieve at all.For example, 3 therein can be preset if the sum of object is 5
Without being blocked, 2 objects can be blocked object.
By the way that this synthetic parameters of coverage extent are arranged, object detection model can be improved to the detection robust slightly blocked
Property.
(5) about background image
The background image is usually no object images, such as desktop picture, paper image, wall image；It equally can be pre-
First establish the set of a background image, it is random from this collection or sequentially obtain a background when one image of every synthesis
Image, and then by the synthesis of the stingy diagram data of each object on the background image.
It should be noted that, although mainly model is trained by the stingy diagram data of object in each composograph, but
The background image of synthetic object also will affect the effect of model training, if composograph all in training data uses equally
Background then will lead to the background in serious over-fitting training data such as munpy array, i.e. plain white background so that final instruction
The detection effect of the object detection model got is poor.
Every kind of synthetic parameters are described above to the detailed process of the stingy diagram data of process object, when image is opened in synthesis one
When, if there are many synthetic parameters include, it can be according to random or preset sequence, respectively according to every kind of synthetic parameters to right
The stingy diagram data of elephant is handled, and the generation method of following training datas provides an example, with rigid body object in the example
For, but not as the restriction to synthesis parameter processing sequence and process object type, this method comprises the following steps:
Step 302, multiple images from the object of multiple angle shots are obtained；Wherein, every kind of object is labeled with object
Type；
By taking rigid body object as an example, above-mentioned multiple angles can for the front of rigid body object, left side, right side, upside, downside,
Upper left side, lower left side, upper right side, lower right side etc.；It, as far as possible can be from the angle shot to the rigid body object in selected angle
Feature, as the rigid body object be drug when, brand, trade mark, the drug name of the drug can be embodied from the image of each angle shot
Claim etc.；If a certain angle, such as left side angle, are only capable of taking the box body side of the medicine box, can embody this without any
The text of drug characteristic is then not necessarily to from the left side angle shot drug at this time.
In actual implementation, for a kind of rigid body object, one image of each angle shot, and then the rigid body object is obtained
Multiple images.It is of course also possible to be shot simultaneously for multiple rigid body objects simultaneously, multiple rigid bodies pair in same image
The shooting angle of elephant is identical or different.After all rigid body object shootings, each rigid body object mark can be somebody's turn to do one by one
The type of object；When the type of the object of mark can be used for subsequent composograph, corresponding object type is selected, can also be used
When training object detection model, training result is tested.
Step 304, a variety of objects are taken from image data, the stingy diagram data of the object taken；
Step 306, judge in preset synthetic parameters whether to include the type for scratching the corresponding object of diagram data, if so,
Step 308 is executed, if not, executing step 310；
Step 308, according to the type of the corresponding object of stingy diagram data, this kind is selected from the stingy diagram data of the object taken
The stingy diagram data of the corresponding object of class；
Step 310, judge in preset synthetic parameters whether to include the dimensional information for scratching diagram data, if so, executing step
Rapid 312, if not, executing step 314；
Step 312, according to the dimensional information, the scale of the stingy diagram data of the object taken is adjusted；Wherein, it is blended into same
The stingy diagram data scale opened on background image is identical.
Step 314, judge in preset synthetic parameters whether to include the rotation angle for scratching diagram data, if so, executing step
Rapid 316, if not, executing step 318；
Step 316, the stingy diagram data of each object taken is rotated to corresponding rotation angle；
Step 318, by random manner, position arrangement is carried out to the stingy diagram data of the object taken；
Step 320, judge in preset synthetic parameters whether to include the coverage extent for scratching diagram data, if so, executing step
Rapid 322, if not, executing step 324；
Step 322, according to coverage extent, judge that the stingy diagram data of the object taken whether there is the stingy figure that should be blocked
Data；If so, the mobile stingy diagram data being blocked is to predeterminated position；On predeterminated position, the stingy diagram data being blocked is removed
Stingy diagram data other than the stingy diagram data being blocked, is blocked according to coverage extent.
Step 324, the stingy diagram data of the object taken is blended on the background image in preset synthetic parameters included,
Obtain composograph；Using the composograph as the training data of object detection model.
Step 326, whether the quantity of training of judgement data is greater than or equal to preset amount threshold；If so, terminating stream
Journey, if not, step 306 is continued to execute, to continue to synthesize the stingy diagram data of the object taken according to preset synthetic parameters
On to corresponding background image, until the quantity of training data is greater than or equal to amount threshold.
For example, if the object type got is ten kinds, the image of corresponding ten different angles of every kind of object, by certainly
It is dynamic that above-mentioned various synthetic parameters are adjusted (various synthetic parameters can be adjusted in a random way), it is available
Tens of thousands of composographs, these composographs form the training data of object detection model.Since the quantity of training data can be with
The detection accuracy of model is directly affected, the model inspection accuracy that the training picture training of a such as thousand sheets obtains may be than 10,000
The model inspection accuracy that Zhang Xunlian picture training obtains low about 10%.A large amount of training data is thus obtained through this embodiment
It can help improve the accuracy of object detection model.
The generation method of above-mentioned training data can be by a small amount of image data by a variety of synthetic parameters of adjust automatically
Synthesis obtains a large amount of composograph, to improve the convenience for obtaining a large amount of training datas, while no longer needing to energetically
Composograph is labeled processing, reduces the human cost to training data mark processing.
Example IV:
A kind of training method of object detection model is present embodiments provided, this method is in training provided by the above embodiment
It is realized on the basis of the generation method of data；As shown in figure 3, this method comprises the following steps:
Step S302 obtains the training data of object；The training data is generated by the generation method of above-mentioned training data；
Training data is input in preset initial model and is trained by step S304, obtains object detection model.
Wherein, which is specifically as follows convolutional neural networks, and the convolutional neural networks are exported by FPN
Characteristic carries out further feature extraction and analysis.In addition, in the training process of model, it is also necessary to pass through test data
The initial training effect of test model, and further finely tuned based on model of the test result to initial training, it is therefore, above-mentioned
After training data is input to the step of being trained in preset initial model, the above method further includes following steps:
Step 1, test data is obtained；The test data includes that multiple include the image data of object；The test data
Be specifically as follows really take pictures include object image data.
Step 2, test data is input in the initial model after training, is outputed test result；It is wrapped in the test result
The location information of object and the image data of type are labeled with containing multiple；
Wherein, the location information of object can be identified by rectangle frame, and the type of object can be with the shape of text or symbol
Formula identifies near rectangle frame, and corresponds with rectangle frame.
Step 3, from the location information and the correct image data of type of test result screening object；
Step 4, by the image data filtered out, processing is finely adjusted to the initial model after training, is obtained final
Object detection model.
It is above-mentioned that test data is input to the process in the initial model after training and to output test result, it can be understood as
The process that the object in test data is labeled by the initial model after training, if the initial model after training is to survey
Location information and the type for trying the mark of certain image datas in data are correct, illustrate that these image datas are more in line with instruction
The detection of initial model after white silk is inclined to, thus, according to these image datas, place is finely adjusted to the initial model after training
Reason, can make final object detection model be more in line with the desired effect of object detection.
Specifically, above-mentioned trim process is referred to as fine-tune process；Again by the above-mentioned image data filtered out
The secondary initial model being input to after training, can adjust out the weight of each network layer in initial model according to these image datas,
And the relevant parameter of last several layers of network layers；Due to above-mentioned test data be true environment under shoot include object figure
As data, model is finely adjusted by the part image data in these test datas, can be made up in above-mentioned training data
Difference of the composograph in terms of illumination and background with actual measurement environment, makes the detection effect of object detection model with more robust
Property.Tests prove that the object detection model that training data training obtains is obtained using the generation method of above-mentioned training data, then
After above-mentioned trim process, the accuracy in detection of model can achieve 98%.
The training method of above-mentioned object detection model provided in this embodiment, is obtained by the generation method of above-mentioned training data
To after training data, training data is input in preset initial model by this and is trained, obtains object detection model.It should
Mode, as training data, improves the convenience for obtaining a large amount of training datas by using composograph, thus guaranteeing mould
While type detection accuracy, the efficiency of model training is improved.
Embodiment five:
Corresponding to above method embodiment, a kind of structural schematic diagram of the generating means of training data shown in Figure 4,
The device includes:
Image data acquisition module 40, for obtaining comprising there are many image datas of object；Wherein, every kind of object is marked
It is marked with the type of object；
Object takes module 41, for taking a variety of objects from image data, the stingy diagram data of the object taken；
The stingy diagram data of the object taken is blended into pair for according to preset synthetic parameters by image synthesis module 42
On the background image answered, composograph is obtained；Wherein, synthetic parameters include the type for scratching the corresponding object of diagram data, scratch figure number
According to dimensional information, rotation angle, one of coverage extent and background image or a variety of；Using composograph as object detection
The training data of model.
The generating means of training data provided in an embodiment of the present invention are got comprising there are many image datas of object
Afterwards, a variety of objects are taken from the image data, the stingy diagram data of the object taken；And then join according to preset synthesis
Number, the stingy diagram data of the object taken is blended on corresponding background image, composograph is obtained, to obtain object detection
The training data of model；In which, by adjusting a variety of synthetic parameters, it can be synthesized to obtain by a small amount of image data a large amount of
Composograph, to improve the convenience for obtaining a large amount of training datas.In addition, since image is before synthesis to each
Object is marked, thus no longer needs to be labeled composograph energetically processing, is reduced at training data mark
The human cost of reason.
Further, above-mentioned image data acquisition module, is also used to: obtaining multiple figures from the object of multiple angle shots
Picture.
Further, if including the dimensional information for scratching diagram data in synthetic parameters, above-mentioned image synthesis module is also used to:
According to dimensional information, the scale of the stingy diagram data of the object taken is adjusted；Wherein, the stingy figure being blended on same background image
Data scale is identical；Stingy diagram data after rescaling is blended on corresponding background image.
Further, if including the rotation angle for scratching diagram data in synthetic parameters, above-mentioned image synthesis module is also used to:
The stingy diagram data of each object taken is rotated to corresponding rotation angle；Postrotational stingy diagram data is blended into corresponding
On background image.
Further, above-mentioned image synthesis module is also used to: by random manner, to the stingy diagram data of the object taken
Position arrangement is carried out, the stingy diagram data after arrangement is blended on corresponding background image.
Further, if including the coverage extent for scratching diagram data in synthetic parameters, above-mentioned image synthesis module is also used to:
According to coverage extent, judge that the stingy diagram data of the object taken whether there is the stingy diagram data that should be blocked；If so, mobile
The stingy diagram data being blocked is to predeterminated position；On predeterminated position, the stingy diagram data being blocked is by except the stingy diagram data being blocked
Stingy diagram data in addition, is blocked according to coverage extent.
Further, above-mentioned apparatus further includes judgment models, and whether the quantity for training of judgement data is greater than or equal to
Preset amount threshold；Model is continued to execute, if the quantity for training data is less than preset amount threshold, continues to trigger
Above-mentioned image synthesis module operation, until the quantity of training data is greater than or equal to amount threshold.
Further, above-mentioned object is rigid body object.
The technical effect of device provided by the present embodiment, realization principle and generation is identical with previous embodiment, for letter
It describes, Installation practice part does not refer to place, can refer to corresponding contents in preceding method embodiment.
A kind of structural schematic diagram of the training device of object detection model shown in Figure 5, the device include:
Training data obtains module 50, for obtaining the training data of object；Training data passes through above-mentioned training data
Generation method operation generates；
Training module 51 is trained for training data to be input in preset initial model, obtains object detection
Model.
The training device of above-mentioned object detection model provided in this embodiment, is obtained by the generation method of above-mentioned training data
To after training data, training data is input in preset initial model by this and is trained, obtains object detection model.It should
Mode, as training data, improves the convenience for obtaining a large amount of training datas by using composograph, thus guaranteeing mould
While type detection accuracy, the efficiency of model training is improved.
Further, above-mentioned apparatus further include:
Test data obtains module, for obtaining test data；Test data includes that multiple include the picture number of object
According to；
Data input module outputs test result for test data to be input in the initial model after training；Test
As a result the location information of object and the image data of type are labeled with comprising multiple in；
Screening module, for the location information and the correct image data of type from test result screening object；
Module is finely tuned, for the image data by filtering out, processing is finely adjusted to the initial model after training, is obtained
Final object detection model.
The technical effect of device provided by the present embodiment, realization principle and generation is identical with previous embodiment, for letter
It describes, Installation practice part does not refer to place, can refer to corresponding contents in preceding method embodiment.
Embodiment six:
The embodiment of the invention provides a kind of electronic system, the electronic system include: image capture device, processing equipment and
Storage device；Image capture device, for obtaining preview video frame or image data；
Computer program is stored on storage device, computer program executes above-mentioned trained number when equipment processed is run
According to generation method or above-mentioned object detection model training method.
It is apparent to those skilled in the art that for convenience and simplicity of description, the electronics of foregoing description
The specific work process of system, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.
Further, the present embodiment additionally provides a kind of computer readable storage medium, on computer readable storage medium
It is stored with computer program, computer program equipment processed executes the generation method such as above-mentioned training data when running, or
The step of executing the training method of above-mentioned object detection model.
The generation method of training data provided by the embodiment of the present invention, the training method of object detection model, device and
The computer program product of electronic system, the computer readable storage medium including storing program code, said program code
Including instruction can be used for executing previous methods method as described in the examples, specific implementation can be found in embodiment of the method, herein
It repeats no more.
In addition, in the description of the embodiment of the present invention unless specifically defined or limited otherwise, term " installation ", " phase
Even ", " connection " shall be understood in a broad sense, for example, it may be being fixedly connected, may be a detachable connection, or be integrally connected；It can
To be mechanical connection, it is also possible to be electrically connected；It can be directly connected, can also can be indirectly connected through an intermediary
Connection inside two elements.For the ordinary skill in the art, above-mentioned term can be understood at this with concrete condition
Concrete meaning in invention.
It, can be with if the function is realized in the form of SFU software functional unit and when sold or used as an independent product
It is stored in a computer readable storage medium.Based on this understanding, technical solution of the present invention is substantially in other words
The part of the part that contributes to existing technology or the technical solution can be embodied in the form of software products, the meter
Calculation machine software product is stored in a storage medium, including some instructions are used so that a computer equipment (can be a
People's computer, server or network equipment etc.) it performs all or part of the steps of the method described in the various embodiments of the present invention.
And storage medium above-mentioned includes: that USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), arbitrary access are deposited
The various media that can store program code such as reservoir (RAM, Random Access Memory), magnetic or disk.
In the description of the present invention, it should be noted that term " center ", "upper", "lower", "left", "right", "vertical",
The orientation or positional relationship of the instructions such as "horizontal", "inner", "outside" be based on the orientation or positional relationship shown in the drawings, merely to
Convenient for description the present invention and simplify description, rather than the device or element of indication or suggestion meaning must have a particular orientation,
It is constructed and operated in a specific orientation, therefore is not considered as limiting the invention.In addition, term " first ", " second ",
" third " is used for descriptive purposes only and cannot be understood as indicating or suggesting relative importance.
Finally, it should be noted that embodiment described above, only a specific embodiment of the invention, to illustrate the present invention
Technical solution, rather than its limitations, scope of protection of the present invention is not limited thereto, although with reference to the foregoing embodiments to this hair
It is bright to be described in detail, those skilled in the art should understand that: anyone skilled in the art
In the technical scope disclosed by the present invention, it can still modify to technical solution documented by previous embodiment or can be light
It is readily conceivable that variation or equivalent replacement of some of the technical features；And these modifications, variation or replacement, do not make
The essence of corresponding technical solution is detached from the spirit and scope of technical solution of the embodiment of the present invention, should all cover in protection of the invention
Within the scope of.Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (14)
1. a kind of generation method of training data, which is characterized in that the described method includes:
It obtains comprising there are many image datas of object；Wherein, every kind of object is labeled with the type of the object；
A variety of objects are taken from described image data, the stingy diagram data of the object taken；
According to preset synthetic parameters, the stingy diagram data of the object taken is blended on corresponding background image, is obtained
Composograph；Wherein, the synthetic parameters include the type of the corresponding object of the stingy diagram data, the stingy diagram data
One of dimensional information, rotation angle, coverage extent and background image are a variety of；
Using the composograph as the training data of object detection model.
2. the method according to claim 1, wherein the step of acquisition includes the image data there are many object:
Obtain multiple images from the object of multiple angle shots.
3. the method according to claim 1, wherein if in the synthetic parameters including the stingy diagram data
The stingy diagram data of dimensional information, the object that will be taken is blended into the step on corresponding background image, comprising:
According to the dimensional information, the scale of the stingy diagram data of the object taken is adjusted；Wherein, it is blended into same background
The stingy diagram data scale on image is identical；
The stingy diagram data after rescaling is blended on corresponding background image.
4. the method according to claim 1, wherein if in the synthetic parameters including the stingy diagram data
Angle is rotated, the stingy diagram data of the object that will be taken is blended into the step on corresponding background image, comprising:
The stingy diagram data of each object taken is rotated to corresponding rotation angle；
The postrotational stingy diagram data is blended on corresponding background image.
5. the method according to claim 1, wherein the stingy diagram data of the object that will be taken is blended into
Step on corresponding background image, comprising:
By random manner, position arrangement is carried out to the stingy diagram data of the object taken, by the stingy figure after arrangement
On Data Synthesis to corresponding background image.
6. according to the method described in claim 5, it is characterized in that, if in the synthetic parameters including the stingy diagram data
Coverage extent, after the step of carrying out position arrangement to the stingy diagram data of the object taken, the method also includes:
According to the coverage extent, judge that the stingy diagram data of the object taken whether there is the stingy figure number that should be blocked
According to；
If so, the stingy diagram data being blocked described in mobile is to predeterminated position；On the predeterminated position, it is described be blocked scratch
Diagram data is blocked by the stingy diagram data in addition to the stingy diagram data being blocked according to the coverage extent.
7. the method according to claim 1, wherein using the composograph as the training of object detection model
After the step of data, which comprises
Judge whether the quantity of the training data is greater than or equal to preset amount threshold；
If not, continuing to execute according to preset synthetic parameters, the stingy diagram data of the object taken is blended into corresponding
Step on background image, until the quantity of the training data is greater than or equal to the amount threshold.
8. method according to claim 1-7, which is characterized in that the object is rigid body object.
9. a kind of training method of object detection model, which is characterized in that the described method includes:
Obtain the training data of object；The training data is generated by the described in any item the methods of claim 1-8；
The training data is input in preset initial model and is trained, object detection model is obtained.
10. according to the method described in claim 9, it is characterized in that, the training data is input to preset initial model
In after the step of being trained, which comprises
Obtain test data；The test data includes that multiple include the image data of object；
The test data is input in the initial model after training, is outputed test result；It is wrapped in the test result
The location information of the object and the image data of type are labeled with containing multiple；
The location information and the correct image data of type of the object are screened from the test result；
By the described image data filtered out, processing is finely adjusted to the initial model after training, obtains final pair
As detection model.
11. a kind of generating means of training data, which is characterized in that described device includes:
Image data acquisition module, for obtaining comprising there are many image datas of object；Wherein, every kind of object marks
There is the type of the object；
Object takes module, and for taking a variety of objects from described image data, the object taken is scratched
Diagram data；
Image synthesis module, for according to preset synthetic parameters, the stingy diagram data of the object taken to be blended into correspondence
Background image on, obtain composograph；Wherein, the synthetic parameters include the kind of the corresponding object of the stingy diagram data
One of class, the dimensional information of the stingy diagram data, rotation angle, coverage extent and background image are a variety of；By the conjunction
Training data at image as object detection model.
12. a kind of training device of object detection model, which is characterized in that described device includes:
Training data obtains module, for obtaining the training data of object；The training data passes through any one of claim 1-8
Described the method generates；
Training module is trained for the training data to be input in preset initial model, obtains object detection mould
Type.
13. a kind of electronic system, which is characterized in that the electronic system includes: image capture device, processing equipment and storage dress
It sets；
Described image acquires equipment, for obtaining preview video frame or image data；
Computer program is stored on the storage device, the computer program executes such as when being run by the processing equipment
The described in any item methods of claim 1 to 8, or execute method as claimed in claim 8 or 9.
14. a kind of computer readable storage medium, computer program, feature are stored on the computer readable storage medium
It is, the computer program equipment processed executes method as claimed in any one of claims 1 to 8 when running, or holds
The step of capable method as described in claim 9 or 10.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811382077.4A CN109544496A (en) | 2018-11-19 | 2018-11-19 | Generation method, the training method and device of object detection model of training data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811382077.4A CN109544496A (en) | 2018-11-19 | 2018-11-19 | Generation method, the training method and device of object detection model of training data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109544496A true CN109544496A (en) | 2019-03-29 |
Family
ID=65848436
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811382077.4A Pending CN109544496A (en) | 2018-11-19 | 2018-11-19 | Generation method, the training method and device of object detection model of training data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109544496A (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110111238A (en) * | 2019-04-24 | 2019-08-09 | 薄涛 | Image processing method, device, equipment and its storage medium |
| CN110163285A (en) * | 2019-05-23 | 2019-08-23 | 阳光保险集团股份有限公司 | Ticket recognition training sample synthetic method and computer storage medium |
| CN110298833A (en) * | 2019-06-28 | 2019-10-01 | 百度在线网络技术（北京）有限公司 | Image processing method and device |
| CN110992297A (en) * | 2019-11-11 | 2020-04-10 | 北京百度网讯科技有限公司 | Multi-commodity image synthesis method and device, electronic equipment and storage medium |
| CN111210681A (en) * | 2020-03-03 | 2020-05-29 | 北京文香信息技术有限公司 | Multi-functional wisdom blackboard |
| CN111724392A (en) * | 2020-05-25 | 2020-09-29 | 浙江工业大学 | Data processing method for deep learning feature attention transfer |
| CN111951154A (en) * | 2020-08-14 | 2020-11-17 | 中国工商银行股份有限公司 | Method and device for generating picture containing background and medium |
| CN112419214A (en) * | 2020-10-28 | 2021-02-26 | 深圳市优必选科技股份有限公司 | Method and device for generating labeled image, readable storage medium and terminal equipment |
| CN112505049A (en) * | 2020-10-14 | 2021-03-16 | 上海互觉科技有限公司 | Mask inhibition-based method and system for detecting surface defects of precision components |
| CN113012054A (en) * | 2019-12-20 | 2021-06-22 | 舜宇光学（浙江）研究院有限公司 | Sample enhancement method and training method based on sectional drawing, system and electronic equipment thereof |
| CN113065607A (en) * | 2021-04-20 | 2021-07-02 | 平安国际智慧城市科技股份有限公司 | Image detection method, image detection device, electronic device, and medium |
| WO2021196551A1 (en) * | 2020-03-31 | 2021-10-07 | 北京旷视科技有限公司 | Image retrieval method and apparatus, computer device, and storage medium |
| CN113688887A (en) * | 2021-08-13 | 2021-11-23 | 百度在线网络技术（北京）有限公司 | Training and image recognition method and device of image recognition model |
| WO2023151527A1 (en) * | 2022-02-09 | 2023-08-17 | 维沃移动通信有限公司 | Image photographing method and apparatus |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682023A (en) * | 2015-11-10 | 2017-05-17 | 杭州华为数字技术有限公司 | Method and device for generating data sets |
| CN108230343A (en) * | 2018-01-05 | 2018-06-29 | 厦门华联电子股份有限公司 | A kind of image processing method and device |
| CN108320290A (en) * | 2017-12-29 | 2018-07-24 | 中国银联股份有限公司 | Target Photo extracts antidote and device, computer equipment and recording medium |
| CN108460414A (en) * | 2018-02-27 | 2018-08-28 | 北京三快在线科技有限公司 | Generation method, device and the electronic equipment of training sample image |
| CN108492343A (en) * | 2018-03-28 | 2018-09-04 | 东北大学 | A kind of image combining method for the training data expanding target identification |
| US20180260668A1 (en) * | 2017-03-10 | 2018-09-13 | Adobe Systems Incorporated | Harmonizing composite images using deep learning |
| CN108537859A (en) * | 2017-03-02 | 2018-09-14 | 奥多比公司 | Use the image masks of deep learning |
| CN108564103A (en) * | 2018-01-09 | 2018-09-21 | 众安信息技术服务有限公司 | Data processing method and device |
| CN108573254A (en) * | 2017-03-13 | 2018-09-25 | 北京君正集成电路股份有限公司 | The generation method of characters on license plate gray-scale map |
| CN108805201A (en) * | 2018-06-08 | 2018-11-13 | 湖南宸瀚信息科技有限责任公司 | Destination image data set creation method and its device |
-
2018
- 2018-11-19 CN CN201811382077.4A patent/CN109544496A/en active Pending
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682023A (en) * | 2015-11-10 | 2017-05-17 | 杭州华为数字技术有限公司 | Method and device for generating data sets |
| CN108537859A (en) * | 2017-03-02 | 2018-09-14 | 奥多比公司 | Use the image masks of deep learning |
| US20180260668A1 (en) * | 2017-03-10 | 2018-09-13 | Adobe Systems Incorporated | Harmonizing composite images using deep learning |
| CN108573254A (en) * | 2017-03-13 | 2018-09-25 | 北京君正集成电路股份有限公司 | The generation method of characters on license plate gray-scale map |
| CN108320290A (en) * | 2017-12-29 | 2018-07-24 | 中国银联股份有限公司 | Target Photo extracts antidote and device, computer equipment and recording medium |
| CN108230343A (en) * | 2018-01-05 | 2018-06-29 | 厦门华联电子股份有限公司 | A kind of image processing method and device |
| CN108564103A (en) * | 2018-01-09 | 2018-09-21 | 众安信息技术服务有限公司 | Data processing method and device |
| CN108460414A (en) * | 2018-02-27 | 2018-08-28 | 北京三快在线科技有限公司 | Generation method, device and the electronic equipment of training sample image |
| CN108492343A (en) * | 2018-03-28 | 2018-09-04 | 东北大学 | A kind of image combining method for the training data expanding target identification |
| CN108805201A (en) * | 2018-06-08 | 2018-11-13 | 湖南宸瀚信息科技有限责任公司 | Destination image data set creation method and its device |
Non-Patent Citations (1)
| Title |
|---|
| 杨天祺 等: "改进卷积神经网络在分类与推荐中的实例应用", 《计算机应用研究》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110111238A (en) * | 2019-04-24 | 2019-08-09 | 薄涛 | Image processing method, device, equipment and its storage medium |
| CN110163285A (en) * | 2019-05-23 | 2019-08-23 | 阳光保险集团股份有限公司 | Ticket recognition training sample synthetic method and computer storage medium |
| CN110298833A (en) * | 2019-06-28 | 2019-10-01 | 百度在线网络技术（北京）有限公司 | Image processing method and device |
| CN110298833B (en) * | 2019-06-28 | 2021-08-31 | 百度在线网络技术（北京）有限公司 | Picture processing method and device |
| CN110992297A (en) * | 2019-11-11 | 2020-04-10 | 北京百度网讯科技有限公司 | Multi-commodity image synthesis method and device, electronic equipment and storage medium |
| CN113012054A (en) * | 2019-12-20 | 2021-06-22 | 舜宇光学（浙江）研究院有限公司 | Sample enhancement method and training method based on sectional drawing, system and electronic equipment thereof |
| CN113012054B (en) * | 2019-12-20 | 2023-12-05 | 舜宇光学（浙江）研究院有限公司 | Sample enhancement method and training method based on matting, system and electronic equipment thereof |
| CN111210681A (en) * | 2020-03-03 | 2020-05-29 | 北京文香信息技术有限公司 | Multi-functional wisdom blackboard |
| WO2021196551A1 (en) * | 2020-03-31 | 2021-10-07 | 北京旷视科技有限公司 | Image retrieval method and apparatus, computer device, and storage medium |
| CN111724392A (en) * | 2020-05-25 | 2020-09-29 | 浙江工业大学 | Data processing method for deep learning feature attention transfer |
| CN111951154A (en) * | 2020-08-14 | 2020-11-17 | 中国工商银行股份有限公司 | Method and device for generating picture containing background and medium |
| CN111951154B (en) * | 2020-08-14 | 2023-11-21 | 中国工商银行股份有限公司 | Picture generation method and device containing background and medium |
| CN112505049A (en) * | 2020-10-14 | 2021-03-16 | 上海互觉科技有限公司 | Mask inhibition-based method and system for detecting surface defects of precision components |
| CN112505049B (en) * | 2020-10-14 | 2021-08-03 | 上海互觉科技有限公司 | Mask inhibition-based method and system for detecting surface defects of precision components |
| CN112419214A (en) * | 2020-10-28 | 2021-02-26 | 深圳市优必选科技股份有限公司 | Method and device for generating labeled image, readable storage medium and terminal equipment |
| CN113065607A (en) * | 2021-04-20 | 2021-07-02 | 平安国际智慧城市科技股份有限公司 | Image detection method, image detection device, electronic device, and medium |
| CN113688887A (en) * | 2021-08-13 | 2021-11-23 | 百度在线网络技术（北京）有限公司 | Training and image recognition method and device of image recognition model |
| WO2023151527A1 (en) * | 2022-02-09 | 2023-08-17 | 维沃移动通信有限公司 | Image photographing method and apparatus |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109544496A (en) | Generation method, the training method and device of object detection model of training data | |
| CN108764024A (en) | Generating means, method and the computer readable storage medium of human face recognition model | |
| JP6976270B2 (en) | Remote determination of the amount stored in a container in a geographic area | |
| CN109242961A (en) | A kind of face modeling method, apparatus, electronic equipment and computer-readable medium | |
| CN109376667A (en) | Object detection method, device and electronic equipment | |
| CN109325437A (en) | Image processing method, device and system | |
| CN109740571A (en) | The method of Image Acquisition, the method, apparatus of image procossing and electronic equipment | |
| CN104427252B (en) | Method and its electronic equipment for composograph | |
| CN109727264A (en) | Image generating method, the training method of neural network, device and electronic equipment | |
| CN109690620A (en) | Threedimensional model generating means and threedimensional model generation method | |
| CN108805047A (en) | A kind of biopsy method, device, electronic equipment and computer-readable medium | |
| CN106575450A (en) | Augmented reality content rendering via albedo models, systems and methods | |
| CN106797460A (en) | The reconstruction of 3 D video | |
| CN106663126A (en) | Video processing for motor task analysis | |
| CN109815843A (en) | Object detection method and Related product | |
| CN109376631A (en) | A kind of winding detection method and device neural network based | |
| CN107563978A (en) | Face deblurring method and device | |
| CN108961149A (en) | Image processing method, device and system and storage medium | |
| CN107918767B (en) | Object detection method, device, electronic equipment and computer-readable medium | |
| CN110428412A (en) | The evaluation of picture quality and model generating method, device, equipment and storage medium | |
| US20190102654A1 (en) | Generation of Training Data for Image Classification | |
| CN106131405A (en) | Image processing method and device and for its image capturing system | |
| CN110064200A (en) | Object construction method, device and readable storage medium storing program for executing based on virtual environment | |
| CN107403160A (en) | Image detecting method, equipment and its storage device in a kind of intelligent driving scene | |
| Kumar et al. | High-throughput 3D reconstruction of plant shoots for phenotyping |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190329 |