CN108875934A - A kind of training method of neural network, device, system and storage medium - Google Patents
A kind of training method of neural network, device, system and storage medium Download PDFInfo
- Publication number
- CN108875934A CN108875934A CN201810525605.0A CN201810525605A CN108875934A CN 108875934 A CN108875934 A CN 108875934A CN 201810525605 A CN201810525605 A CN 201810525605A CN 108875934 A CN108875934 A CN 108875934A
- Authority
- CN
- China
- Prior art keywords
- image
- level
- neural network
- classification
- image classification
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/20—Scenes; Scene-specific elements in augmented reality scenes
Abstract
The present invention provides training method, device, system and the storage medium of a kind of neural network for image recognition and a kind of image-recognizing methods neural network based.The training method includes：The sample image for being labelled with the class label of N number of image classification level is received, wherein the classification of next level in adjacent image taxonomical hierarchy is the subclass of the classification of a level thereon, and N is the integer greater than 1；Neural network is trained using sample image, wherein the class label based on each image classification level calculates separately the Classification Loss for sample image, using the level loss as the image classification level；According to the level costing bio disturbance total losses of each image classification level；Using total losses as the parameter of objective function training neural network.Technical solution of the present invention obtains the neural network that can more accurately carry out image recognition.
Description
Technical field
The present invention relates to technical field of image processing, relate more specifically to a kind of instruction of neural network for image recognition
Practice method, apparatus, system and storage medium, in addition, further relating to a kind of image-recognizing method neural network based.
Background technique
Image recognition, which refers to, is analyzed and processed image, to identify the technology of various objects in the image.Its target is
An image is given, judges automatically the object in the image belongs to which kind of specific object or which kind of specific scene.Nowadays, scheme
As identification is taken pictures in smart camera, intelligent photograph album management, image classification and retrieval, safety of image in terms of play very
Therefore important role also has received widespread attention and studies.
In existing image-recognizing method, by all target categories (may include foreground object and background) directly as mind
The parameter of output training neural network through network, to obtain the classification of trained neural network as image identification system
Device.This neural network is more and more when kind of object and/or scene type, for the image of content complexity, will be unable to obtain
Ideal image recognition result.
In order to solve the above technical problems, it is necessary to propose a kind of training skill of new neural network for image recognition
Art.
Summary of the invention
The present invention is proposed in view of the above problem.
According to an aspect of the present invention, a kind of training method of neural network for image recognition is provided, including：
The sample image for being labelled with the class label of N number of image classification level is received, wherein in adjacent image taxonomical hierarchy
Next level classification be the classification of a level thereon subclass, N is integer greater than 1；
Neural network is trained using the sample image, wherein
Based on the class label of each image classification level, the Classification Loss for the sample image is calculated separately, with
Level as the image classification level loses；
According to the level costing bio disturbance total losses of each image classification level；
Using the total losses as the parameter of the objective function training neural network.
Illustratively, the class label based on each image classification level, calculates separately for the sample image
Classification Loss, using as the image classification level level loss include：
For each image classification level, the output of class label and the neural network based on the image classification level
The classification results of the image classification level of layer output calculate the level loss of the image classification level.
Illustratively, the neural network includes：M convolutional layer and N number of full articulamentum, wherein
N number of in the M convolutional layer connect one by one with N number of full articulamentum respectively and with N number of full connection
The convolutional layer that connects one by one of layer includes M convolutional layer, the output of N number of full articulamentum respectively with N number of image classification layer
Secondary one-to-one correspondence, the output for the full articulamentum connecting with M convolutional layer correspond to the image classification of n-th layer time, and M is greater than N's
Integer.
Illustratively, the class label based on each image classification level, calculates separately for the sample image
Classification Loss, using as the image classification level level loss include：
For each image classification level, which is exported by full articulamentum corresponding with the image classification level
Secondary classification results；Class label based on the classification results He the image classification level, calculates the layer of the image classification level
Secondary loss.
Illustratively, the level costing bio disturbance total losses according to each image classification level includes：
The total losses Loss is calculated according to the following formula：
Wherein, LossiIt is the level loss of i-th of image classification level, WiIt is the layer for controlling i-th of image classification level
Secondary loss LossiEffect degree parameter, 0<i<N+1.
Illustratively, the method also includes：The parameter W is determined by searching algorithmi。
Illustratively, the method also includes：Receive the parameter W being arranged based on experience valuei。
Illustratively, the class label of N number of image classification level includes：
According to the class label for N number of image classification level that similarity degree between class is classified；Or
According to the class label for N number of image classification level that the difference for distinguishing difficulty is classified.
Illustratively, described image identification includes the scene Recognition of image.
According to an aspect of the present invention, a kind of image-recognizing method neural network based is provided, including：
Obtain images to be recognized；
Image recognition is carried out to the images to be recognized using the trained neural network, wherein the nerve net
Network leads to training method training above-mentioned and obtains.
According to a further aspect of the invention, a kind of training device of neural network for image recognition is additionally provided, including：
Receiving module receives the sample image for being labelled with the class label of N number of image classification level, wherein adjacent image point
The classification of next level in class hierarchy is the subclass of the classification of a level thereon, and N is the integer greater than 1；
Training module is trained neural network using the sample image, wherein
Based on the class label of each image classification level, the Classification Loss for the sample image is calculated separately, with
Level as the image classification level loses；
According to the level costing bio disturbance total losses of each image classification level；
Using the total losses as the parameter of the objective function training neural network.
Another aspect according to the present invention additionally provides a kind of training system of neural network for image recognition, including
Processor and memory, wherein computer program instructions are stored in the memory, the computer program instructions are described
For executing the training method for being previously described for the neural network of image recognition when processor is run.
According to a further aspect of the present invention, a kind of storage medium is additionally provided, program is stored on said storage and refers to
It enables, described program instruction is at runtime for executing the training method for the neural network for being previously described for image recognition.
Training method, device, system and the storage of neural network according to an embodiment of the present invention for image recognition are situated between
Matter, sample image are labelled with the class label of multiple images taxonomical hierarchy, and training classification of the neural network based on many levels is true
Determine image recognition result.Above-mentioned technical proposal obtains the neural network that can more accurately carry out image recognition as a result,.Especially
Identification for the image of content complexity, the neural network that above-mentioned technical proposal obtains can significantly improve compared with prior art
Image recognition accuracy.
The above description is only an overview of the technical scheme of the present invention, in order to better understand the technical means of the present invention,
And it can be implemented in accordance with the contents of the specification, and in order to allow above and other objects of the present invention, feature and advantage can
It is clearer and more comprehensible, the followings are specific embodiments of the present invention.
Detailed description of the invention
The embodiment of the present invention is described in more detail in conjunction with the accompanying drawings, the above and other purposes of the present invention,
Feature and advantage will be apparent.Attached drawing is used to provide to further understand the embodiment of the present invention, and constitutes explanation
A part of book, is used to explain the present invention together with the embodiment of the present invention, is not construed as limiting the invention.In the accompanying drawings,
Identical reference label typically represents same parts or step.
Fig. 1 show for realizing the neural network according to an embodiment of the present invention for image recognition training method and
The schematic block diagram of the exemplary electronic device of device；
Fig. 2 shows the schematic diagrames for the example classification device for being used for image recognition according to prior art；
Fig. 3 shows the schematic diagram of the example classification device according to an embodiment of the invention for image recognition；
Fig. 4 shows the signal of the training method of the neural network according to an embodiment of the invention for image recognition
Property flow chart；
Fig. 5 show it is according to an embodiment of the invention neural network is trained using sample image it is schematic
Flow chart；
Fig. 6 shows the signal of the training method of the neural network according to an embodiment of the invention for image recognition
Property block diagram；
Fig. 7 shows showing for the training method of the neural network in accordance with another embodiment of the present invention for image recognition
Meaning property block diagram；
Fig. 8 shows the training method of the neural network for image recognition accord to a specific embodiment of that present invention
Schematic block diagram；
Fig. 9 shows the schematic flow chart of image-recognizing method according to an embodiment of the invention；
Figure 10 shows showing for the training device of the neural network according to an embodiment of the invention for image recognition
Meaning property block diagram；And
Figure 11 shows showing for the training system of the neural network according to an embodiment of the invention for image recognition
Meaning property block diagram.
Specific embodiment
In order to enable the object, technical solutions and advantages of the present invention become apparent, root is described in detail below with reference to accompanying drawings
According to example embodiments of the present invention.Obviously, described embodiment is only a part of the embodiments of the present invention, rather than this hair
Bright whole embodiments, it should be appreciated that the present invention is not limited by example embodiment described herein.Based on described in the present invention
The embodiment of the present invention, those skilled in the art's obtained all other embodiment in the case where not making the creative labor
It should all fall under the scope of the present invention.
Firstly, describing referring to Fig.1 for realizing the neural network according to an embodiment of the present invention for image recognition
Training method and the exemplary electronic device of device 100.
As shown in Figure 1, electronic equipment 100 includes one or more processors 102, one or more storage devices 104.It can
Selection of land, electronic equipment 100 can also include input unit 106, output device 108 and data acquisition facility 110, these components are logical
Cross bindiny mechanism's (not shown) interconnection of bus system 112 and/or other forms.It should be noted that electronic equipment shown in FIG. 1
100 component and structure be it is illustrative, and not restrictive, as needed, the electronic equipment also can have other
Component and structure.
The processor 102 can be central processing unit (CPU), graphics processor (GPU) or have data processing
The processing unit of ability and/or the other forms of instruction execution capability, and can control other in the electronic equipment 100
Component is 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, processor 102 can run described program instruction, to realize hereafter institute
The client functionality (realized by processor) in the embodiment of the present invention stated and/or other desired functions.In the meter
Can also store various application programs and various data in calculation machine readable storage medium storing program for executing, for example, the application program use and/or
The various data etc. generated.
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 (such as image and/or sound) to external (such as user), and
It and may include one or more of display, loudspeaker etc..
The data acquisition facility 110 can acquire the various forms of data such as image, and data collected are deposited
Storage is in the storage device 104 for the use of other components.Data acquisition facility 110 can be camera etc..It should be appreciated that
Data acquisition facility 110 is only example, and electronic equipment 100 can not include data acquisition facility 110.In this case, may be used
To obtain data using other data acquisition facilities, and acquired data are sent to electronic equipment 100.
Illustratively, for realizing the training method of the neural network according to an embodiment of the present invention for image recognition and
The exemplary electronic device of device can be realized in the equipment of personal computer or remote server etc..
The training of neural network for image recognition uses supervised learning method.In order to which training is used for the mind of image recognition
Through network, it is desirable to provide sufficient amount of sample image.Each sample image is labelled with corresponding class label.Such as intelligence
The neural network of image recognition in photographing camera, it will be understood that can be to the sample image that the neural network is trained
Photo including various common photographic subjects.For example different natural land of photographic subjects：Sea, meadow, mountains and rivers, forest
Deng and different objects：People, cat, dog, bird etc..Each photo as sample image has the class label belonging to it, such as
The class label of certain sample image is " dog ".
In the prior art, usually that all foreground objects and background are straight to the training of the neural network for image recognition
The output item as neural network is connect to train the parameter of neural network.The neural network for completing parameter training may be used as image
The classifier of identifying system.Such as use classes label is respectively major part self-timer, front portrait, cat, dog, sea, meadow, day
Out, the 8 class sample images in blue sky carry out the parameter training of neural network.Only one correspondence of each sample image in the prior art
Class label, as shown in Figure 2.Firstly, sample image to be inputted to neural network to be trained, it is right which will export its
The classification results answered.Then, the parameter of neural network is adjusted according to the class label of the classification results and sample image, so that
Identification all as correct as possible for every class sample image.Housebroken neural network, i.e., the classification of one 8 classification are obtained as a result,
Device.For neural network, the status of above-mentioned each class label be it is identical, be in same level.
According to an embodiment of the invention, sample image is labelled with the class label of multiple images taxonomical hierarchy.Such as Fig. 3 institute
Show, for above-mentioned 8 class sample image, the class label of 2 image classification levels can be marked.First layer is the thick of more macroscopic view
Classification distinguishing label：{ people, animal, natural land, light environment }.The second layer is that { major part is certainly for more microcosmic subdivision class label
Bat, front portrait, cat, dog, sea, meadow, sunrise, blue sky }.The class label for taxonomical hierarchy that there are two each sample images,
Such as the label of a sample image is：The class of class label " animal " and the 2nd image classification level of 1st image classification level
Distinguishing label " dog ".
Above-mentioned example is the characteristics of can be classified from many levels in image recognition using image, to give sample image mark
Infuse the simple examples explanation of the class label of multiple images taxonomical hierarchy.Although in above-mentioned example, with 2 image classification levels
Class label is illustrated, but for hundreds and thousands of kinds of class labels in practical application, can be marked more to sample image
The class label of more image classification levels.For the class label of N number of image classification level, in adjacent image taxonomical hierarchy under
The classification of one level is the subclass of the classification of a level thereon, and N is the integer greater than 1.Optionally, N number of image classification level
Class label includes the class label classified according to similarity degree between class.Such as " animal " can be further divided into：It is winged
" reptile and the fish " of " flight animal ", " walking animal " that has four legs and not foot.Optionally, N number of image point
The class label of class hierarchy further includes according to the class label for the different classifications for distinguishing difficulty, such as the classification based on color, base
Classification in texture, classification based on shape etc..
In the following, reference Fig. 4 to be described to the training method of the neural network according to an embodiment of the present invention for being used for image recognition.
Fig. 4 shows the schematic stream of the training method 400 of the neural network according to an embodiment of the invention for image recognition
Cheng Tu.Method 400 includes the following steps：
Step S410 receives the sample image for being labelled with the class label of N number of image classification level.
In the class label of N number of image classification level, the classification of next level in adjacent image taxonomical hierarchy is it
The subclass of the classification of a upper level, N are the integer greater than 1.For example, the 2nd layer of class label is the subclass of the 1st layer of class label,
N-th layer class label is the subclass of N-1 layers of class label.N-th layer class label is the most disaggregated classification of sample image, corresponding
The final classification result of image recognition.It is appreciated that more high-level class label, corresponding to more large scale, more macroscopical
Classification；And the class label of more low level, corresponding to smaller scale, more microcosmic classification.
The received each sample image of institute is labelled with N number of class label, and this N number of class label be belonging respectively to it is N number of
One in image classification level.In other words, it for each of N number of image classification level, is all labelled with for sample image
One belongs to the class label of the image classification level.
Step S420 is trained neural network using the received sample image of step S410.Fig. 5 is shown according to this
The schematic flow chart that neural network is trained using sample image of invention one embodiment.Step S420 includes following
Step：
Step S421 calculates separately point for the sample image based on the class label of each image classification level
Class loss, using the level loss as the image classification level.
For a sample image, based on the class label of each image classification level, calculates separately and carry out the level point
Classification Loss after class, using the level loss as the image classification level.Specifically, sample image is inputted into neural network,
The neural network is directed to each image classification level, extracts the feature of sample image.Based on the feature pair for being directed to the taxonomical hierarchy
Sample image is classified and exports the classification results of the taxonomical hierarchy.Class label such as based on the first image classification level,
Export the classification results of the first image classification level.So obtain sample image, each image classification level classification results.
Current neural network is embodied to the reason of the feature of the image classification level for the classification results of each image classification level
Solution.Respectively according to the classification results of each image classification level and its class label, the level damage of the image classification level is calculated
It loses.The level of each image classification level loses, and represents and is based on the image classification to sample image using current neural network
The sorted error of the class label of level embodies current neural network to the understanding deviation of the image classification level.It is logical
The classification processing of multiple image classification levels is crossed, neural network can learn the feature of multiple images taxonomical hierarchy, thus to institute
The image to be identified has more three-dimensional comprehensive understandability.
Step S422, according to the level costing bio disturbance total losses of the obtained each image classification level of step S421.For
One sample image is lost by the level that step S421 obtains each of which image classification level.Comprehensive each image classification layer
Secondary level loses to obtain the total losses of the sample image.The total losses of the sample image, which represents, uses current neural network pair
The synthesis result that the characteristic of division of each image classification level belonging to it understands.Total losses is bigger, illustrates to use current nerve
The accuracy that network carries out image recognition is more inadequate, more has to be hoisted；Conversely, explanation carries out image using current neural network
The accuracy of identification is better.
Step S423, the total losses obtained using step S421 is as the parameter of objective function training neural network.As above
Described, the total losses of a sample image embodies and carries out image recognition to the sample image using current neural network
Accuracy.For sample image, its total losses is calculated by propagated forward in neural network.It is carried out according to the value of total losses anti-
To derivation, the adjustment of the relevant parameter of neural network is carried out.Repeatedly, by the training of sufficient amount of sample image, always
Loss converges to certain particular value or reaches frequency of training threshold value, to reach training objective.Neural network after training
It can be used as the classifier of image identification system.
The training method of the above-mentioned neural network for image recognition, sample image are labelled with multiple images taxonomical hierarchy
Class label, training neural network determine image recognition result based on multiple images taxonomical hierarchy.Class label is for example more macro
The other label of the rough sort of sight and the other label of more microcosmic disaggregated classification.Obtain can be more acurrate for the training method as a result,
The neural network of ground progress image recognition.The nerve obtained particularly with the identification of the image of content complexity, above-mentioned technical proposal
Network can significantly improve image recognition accuracy compared with prior art.
Illustratively, the training method of the neural network according to an embodiment of the present invention for image recognition can have
It is realized in the unit or system of memory and processor.
Fig. 6 shows the training method 600 of the neural network according to an embodiment of the invention for image recognition
Schematic block diagram.As shown in fig. 6, for each image classification level, being based on the image classification level for a sample image
Class label and neural network output layer output the image classification level classification results, calculate the image classification level
Level loss.Class label such as based on the 1st image classification level, the output layer of neural network export the 1st image point
The classification results of class hierarchy.According to the classification results of the 1st image classification level and its class label, the 1st image point is calculated
The level of class hierarchy loses, and is expressed as Loss1.The sides such as quadratic sum loss, intersection entropy loss can be used in the calculating of level loss
Method, circular do not influence the understanding of the present invention, are not detailed herein.Ibid, it is based on the 2nd image classification layer
Secondary class label, the output layer of neural network export the classification results of the 2nd image classification level, based on the classification results and
The level loss Loss of the 2nd image classification level is calculated in its class label2.And so on, for being labelled with N number of image
The sample image of the class label of taxonomical hierarchy is calculated separately for each image classification level, obtains N number of image classification level
Level lose Loss1、Loss2、……LossN。
According to the level costing bio disturbance total losses of above-mentioned each image classification level.Finally using total losses as target letter
The parameter of number training neural network.
Illustratively, the total losses Loss is calculated according to the following formula：
Wherein, LossiIt is the level loss of i-th of image classification level, WiIt is the layer for controlling i-th of image classification level
Secondary loss LossiEffect degree parameter, 0<i<N+1.Pass through parameter WiNeural network be can control in i-th of image classification
Level carries out influence degree of the error to last classification results of sort operation to image.
Optionally, parameter W is determined by searching algorithmi.Specifically, for each parameter Wi, setting search initial value, knot
Beam value and step-length.In search range between initial value and end value, since initial value, increasing a step-length every time is ginseng
Number WiValue, use the value as parameter WiCarry out the parameter training of neural network.By all possible in search range
WiValue trial, select the value of optimal result as parameter Wi.It is appreciated that different parameters WiIt can be set different
Search for initial value, end value and step-length.It can also be according to the parameter W for being provided with search initial value, end value and step-lengthiValue
Other parameter is set.Such as the parameter W for being 3 for NiSetting it is as follows：Parameter W1Search initial value be 0.30, end value is
0.60, step-length 0.02.Respectively using 0.30,0.32,0.34 ... 0.58,0.60 is parameter W1Value；W2=W1÷2；W3
=1.00-W1-W2.Pass through searching algorithm Selecting All Parameters WiOptimal experimental data, accuracy is related to the step-length of search, more
Small step-length is available more accurate as a result, but also bringing more calculation amounts.
Optionally, the parameter W being arranged based on experience value is receivedi.For example, can be by described in the electronic equipment 100
Input unit 106 receives the parameter, and the parameter can be preset by user.User can rule of thumb control respectively as a result,
The level of image classification level loses the effect degree in total losses, to obtain more satisfied image recognition result.
The training method 600 of the above-mentioned neural network for image recognition, haves no need to change the structure of neural network model,
It is convenient to be applied to different neural network models.Therefore versatility is very strong, suitable for the various nerves for image recognition of training
Network model, to obtain the neural network that image recognition accuracy significantly improves.
Fig. 7 shows the training method 700 of the neural network in accordance with another embodiment of the present invention for image recognition
Schematic block diagram.As shown in fig. 7, the neural network in method 700 for image recognition includes M convolutional layer and N number of connects entirely
Layer is connect, M is the integer greater than N.M convolutional layer be from top to bottom the first convolutional layer, the second convolutional layer ... M convolutional layer.Its
N number of in middle M convolutional layer connect with N number of full articulamentum one by one respectively.The convolutional layer connecting one by one with N number of full articulamentum is
K1Convolutional layer, K2Convolutional layer ... KNConvolutional layer, wherein KN=M.The output of N number of full articulamentum respectively with N number of figure
As taxonomical hierarchy one-to-one correspondence namely KiThe output of the full articulamentum of convolutional layer connection is corresponding with i-th of image classification level.
The output for the full articulamentum connecting with M convolutional layer corresponds to the image classification of n-th layer time.In other words, it is connect with M convolutional layer
Full articulamentum be neural network output layer.When carrying out image recognition using the neural network, it is connect with M convolutional layer
The result of full articulamentum output is final image recognition result.
For each image classification level, which is exported by full articulamentum corresponding with the image classification level
Secondary classification results.Such as i-th of image classification level, full articulamentum corresponding with i-th of image classification level is and KiVolume
The full articulamentum of lamination connection.Utilize the first convolutional layer to KiConvolutional layer calculates sample image, to obtain sample image
The characteristic pattern of i-th of image classification level.This feature figure passes through and K againiThe full articulamentum of convolutional layer connection calculates and exports i-th
The classification results of a image classification level.As a result, using the network structure of above-mentioned layering output, for N number of image classification level,
Obtain the classification results of layering output.Classification results and its class label based on i-th of image classification level, calculate the image
The level of taxonomical hierarchy loses Lossi.According to the level costing bio disturbance total losses of each image classification level and according to total losses
The step of training neural network parameter, is as previously mentioned, which is not described herein again.
In method 700, layering output is carried out for image classification level, to realize order training method.So that neural network
The convolutional layer of shallow-layer is trained up, preferably extraction large scale, the class another characteristic of macroscopic view；While the convolutional layer of deep layer is more
Small scale, microcosmic class another characteristic are extracted well.It can not only allow between shallow-layer and deep layer network and preferably divide the work, moreover it is possible to allow net
Network each section is timely fed back, and gradient is avoided to decay, and network parameter training is more abundant.The neural network is utilized as a result,
More accurate image recognition result can be obtained
The above method 700 is understood in order to clearer, and above-mentioned side accord to a specific embodiment of that present invention is given below
The realization process of method 700.Fig. 8 shows the instruction of the neural network for image recognition accord to a specific embodiment of that present invention
Practice the schematic block diagram of method.As shown in figure 8, the neural network for image recognition includes 16 convolutional layers and 2 full connections
Layer.The sample image of the training neural network is labelled with the class label of 2 image classification levels.Wherein, the 14th convolutional layer
The corresponding 1st image classification level of the full articulamentum of connection, exports the classification results of the 1st image classification level.Based on this point
Class result and its class label obtain the level loss Loss of the 1st image classification level1.16th convolutional layer connects complete
Articulamentum corresponds to the 2nd image classification level, exports the classification results of the 2nd image classification level.Based on the classification results and
Its class label obtains the level loss Loss of the 2nd image classification level2.Total losses Loss is calculated according to the following formula：
Loss=W1×Loss1+W2×Loss2。
W1Level for the 1st image classification level of control loses Loss1Effect degree parameter, W2For control the 2nd
The level of image classification level loses Loss2Effect degree parameter.
Optionally, parameter W1、W2It is determined by searching algorithm.For example, setting parameter W1Search initial value be 0.5, terminate
Value is 1.0, step-length 0.01.Be step-length with 0.01 in the range of 0.5~1.0, respectively using 0.50,0.51,
0.52 ... 0.99,1.00 be parameter W1Value, parameter W2=1.00-W1.Based on each group of parameter W1、W2Value, carry out mind
Parameter training through network.Select the value that can obtain optimal training result as parameter W1、W2。
Optionally, above-mentioned image recognition includes the scene Recognition of image.Scene Recognition is one in image recognition technology
Important branch.The target of scene Recognition is to give an image, judges automatically which kind of specific scene the image belongs to it.One
The scene of a image may include the image classification information of many levels abundant.In the example of the application, provides and be used for
The training method of the neural network of scene Recognition, neural network classifier for example shown in Fig. 3.As shown in figure 3, this point of training
The sample image for being labelled with the class label of 2 image classification levels is utilized in class device.Wherein first layer is the thick of more macroscopic view
Classification distinguishing label：{ people, animal, natural land, light environment }.The second layer is that { major part is certainly for more microcosmic subdivision class label
Bat, front portrait, cat, dog, sea, meadow, sunrise, blue sky }.Utilize the neural network of method training of the invention, Neng Goushi
Now more accurate scene Recognition.
According to an aspect of the present invention, a kind of image-recognizing method neural network based is provided.Fig. 9 is shown according to this
The schematic flow chart of the image-recognizing method of invention one embodiment.
As shown in figure 9, method 900 includes the following steps：
Step S910 obtains images to be recognized.
Images to be recognized can be image that is any suitable, needing to carry out image recognition, such as smart camera cellphone is clapped
The image taken the photograph.It is appreciated that the image can be the video frame in video.Images to be recognized can be the Image Acquisition such as camera
The collected original image of device is also possible to the image obtained after being pre-processed to original image.
Step S920 carries out image recognition to the images to be recognized that step 910 obtains using trained neural network,
In, the neural network is obtained by training method according to an embodiment of the present invention training.According to practical application scene, sample is given
The class label of the N number of image classification level of image labeling, wherein the class label of n-th image classification level be image recognition most
Whole class label.By above-mentioned training method according to an embodiment of the present invention, N number of image classification level is labelled with using this
The sample image training neural network of class label.Image knowledge is carried out to images to be recognized using above-mentioned trained neural network
Not, the classification results of n-th image classification level are final image recognition result.
It is used for the neural network of image recognition in method 900, significantly improves image recognition accuracy.
According to a further aspect of the invention, a kind of training device of neural network for image recognition is additionally provided.
Figure 10 shows showing for the training device of the neural network according to an embodiment of the invention for image recognition
Meaning property block diagram.
As shown in Figure 10, device 1000 includes receiving module 1010 and training module 1020.The modules can be distinguished
Execute each step/function of the training method of the above neural network for image recognition.Below only to the dress
The major function for setting 1000 each component is described, and omits the detail content having been described above.
Receiving module 1010 is used to receive the sample image for the class label for being labelled with N number of image classification level.Wherein phase
The classification of next level in adjacent image classification level is the subclass of the classification of a level thereon, and N is the integer greater than 1.It is optional
Ground, the class label of N number of image classification level may include the N number of image classification level classified according to similarity degree between class
Class label；Or the class label for the N number of image classification level classified according to the difference for distinguishing difficulty.Receiving module
1010 program instructions that can be stored in 102 Running storage device 104 of processor in electronic equipment as shown in Figure 1 come real
It is existing.
Training module 1020 is for being trained neural network using sample image.Training module 1020 can be by Fig. 1
Shown in the program instruction that stores in 102 Running storage device 104 of processor in electronic equipment realize.
According to embodiments of the present invention, training module 1020 includes level costing bio disturbance unit, total losses computing unit and mind
Through network parameter training unit.
The level costing bio disturbance unit is used for the class label based on each image classification level, calculates separately for sample
The Classification Loss of this image, using the level loss as the image classification level.
Illustratively, for each image classification level, class label and neural network based on the image classification level
Output layer output the image classification level classification results, calculate the image classification level level loss.
Illustratively, include for the neural network of image recognition：M convolutional layer and N number of full articulamentum, M are greater than N's
Integer.Wherein, N number of in the M convolutional layer connect one by one with N number of full articulamentum respectively and N number of connects entirely with described
Connecing the convolutional layer that layer connects one by one includes M convolutional layer.The output of N number of full articulamentum respectively with N number of image classification
Level corresponds.The output for the full articulamentum connecting with M convolutional layer corresponds to the image classification of n-th layer time.
Optionally, for each image classification level, which is exported by full articulamentum corresponding with the image classification level
As the classification results of taxonomical hierarchy；Class label based on the classification results He the image classification level, calculates the image classification
The level of level loses.
It is appreciated that example above shows two kinds of specific implementations of the level costing bio disturbance unit, these realizations are shown
Example is only signal, rather than limitation of the present invention.
Total losses computing unit is used for the level costing bio disturbance total losses according to each image classification level.
Optionally, total losses Loss is calculated according to the following formula：
Wherein, LossiIt is the level loss of i-th of image classification level, WiIt is the layer for controlling i-th of image classification level
Secondary loss LossiEffect degree parameter, 0<i<N+1.
Neural network parameter training unit is used for using above-mentioned total losses as the parameter of objective function training neural network.
Optionally, the parameter W is determined by searching algorithmi。
Optionally, the parameter W being arranged based on experience value is receivedi。
Illustratively, described image identification includes the scene Recognition of image.
Those of ordinary skill in the art may be aware that list described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and method and step can be realized with the combination of electronic hardware or computer software and electronic hardware.These functions are actually
It is implemented in hardware or software, the specific application and design constraint depending on technical solution.Professional technician
Each specific application can be used different methods to achieve the described function, but this realization is it is not considered that exceed
The scope of the present invention.
Figure 11 shows showing for the training system of the neural network according to an embodiment of the invention for image recognition
Meaning property block diagram.As shown in figure 11, system 1100 includes input unit 1110, memory 1120, processor 1130.
The input unit 1110 is used to receive the operational order and acquisition data of user's input.Input unit 1110 can
To include one or more of keyboard, mouse, microphone, touch screen and image collecting device etc..Described image acquisition device
It can be used for capturing sample image.
The storage of memory 1120 is for realizing the neural network according to an embodiment of the present invention for image recognition
The computer program instructions of corresponding steps in training method.Optionally, the memory 1120 can be also used for storage sample
Image.
The processor 1130 is for running the computer program instructions stored in the memory 1120, to execute basis
The corresponding steps of the training method of the neural network for image recognition of the embodiment of the present invention, and for realizing according to this hair
Receiving module 1010 and training module 1020 in the training device of the neural network for image recognition of bright embodiment.
In one embodiment, make the system when the computer program instructions are run by the processor 1130
1100 execute following steps：
The sample image for being labelled with the class label of N number of image classification level is received, wherein in adjacent image taxonomical hierarchy
Next level classification be the classification of a level thereon subclass, N is integer greater than 1；
Neural network is trained using the sample image, wherein
Based on the class label of each image classification level, the Classification Loss for the sample image is calculated separately, with
Level as the image classification level loses；
According to the level costing bio disturbance total losses of each image classification level；
Using the total losses as the parameter of the objective function training neural network.
Illustratively, the class label of N number of image classification level includes：Classified according to similarity degree between class
The class label of N number of image classification level；Or the classification for the N number of image classification level classified according to the difference for distinguishing difficulty
Label.
In one embodiment, make the system when the computer program instructions are run by the processor 1130
1100 class labels based on each image classification level executed calculate separately the Classification Loss for the sample image with
Level loss as the image classification level is further comprising the steps of：For each image classification level, based on the image point
The classification results of the class label of class hierarchy and the image classification level of the output layer of neural network output, calculate the figure
As the level of taxonomical hierarchy loses.
In one embodiment, the neural network includes：M convolutional layer and N number of full articulamentum.Wherein, the M volume
N number of convolutional layer for connecting with N number of full articulamentum and being connect one by one with N number of full articulamentum one by one respectively in lamination
Including M convolutional layer, the output of N number of full articulamentum is corresponded with N number of image classification level respectively, with the M volumes
The output of the full articulamentum of lamination connection corresponds to the image classification of n-th layer time, and M is the integer greater than N.
In the above-described embodiments, make the system when the computer program instructions are run by the processor 1130
1100 class labels based on each image classification level executed calculate separately the Classification Loss for the sample image with
Level loss as the image classification level is further comprising the steps of：For each image classification level, by dividing with the image
The corresponding full articulamentum of class hierarchy exports the classification results of the image classification level；Based on the classification results and the image classification layer
Secondary class label calculates the level loss of the image classification level.
Illustratively, execute the system 1100 when the computer program instructions are run by the processor 1130
According in the step of the level costing bio disturbance total losses of each image classification level, institute can be calculated according to the following formula
State total losses Loss：
Wherein, LossiIt is the level loss of i-th of image classification level, WiIt is the layer for controlling i-th of image classification level
Secondary loss LossiEffect degree parameter, 0<i<N+1.
Illustratively, pass through the system 1100 when the computer program instructions are run by the processor 1130
Searching algorithm determines the parameter Wi。
Illustratively, receive the system 1100 when the computer program instructions are run by the processor 1130
The parameter W being arranged based on experience valuei。
Illustratively, described image identification includes the scene Recognition of image.
In addition, according to a further aspect of the present invention, additionally providing a kind of storage medium, storing journey on said storage
Sequence instruction makes the computer or processor execute the present invention real when described program instruction is run by computer or processor
The corresponding steps of the training method of the neural network for image recognition of example are applied, and for realizing according to embodiments of the present invention
The neural network for image recognition training device in corresponding module.The storage medium for example may include intelligent electricity
The storage card of words, the storage unit of tablet computer, the hard disk of personal computer, read-only memory (ROM), erasable programmable are only
Read appointing for memory (EPROM), portable compact disc read-only memory (CD-ROM), USB storage or above-mentioned storage medium
Meaning combination.The computer readable storage medium can be any combination of one or more computer readable storage mediums.
In one embodiment, when the computer program instructions are run by computer or processor, so that the calculating
Machine or processor execute following steps：
The sample image for being labelled with the class label of N number of image classification level is received, wherein in adjacent image taxonomical hierarchy
Next level classification be the classification of a level thereon subclass, N is integer greater than 1；
Neural network is trained using the sample image, wherein
Based on the class label of each image classification level, the Classification Loss for the sample image is calculated separately, with
Level as the image classification level loses；
According to the level costing bio disturbance total losses of each image classification level；
Using the total losses as the parameter of the objective function training neural network.
Illustratively, the class label of N number of image classification level includes：Classified according to similarity degree between class
The class label of N number of image classification level；Or the classification for the N number of image classification level classified according to the difference for distinguishing difficulty
Label.
In one embodiment, when the computer program instructions are run by computer or processor, so that the calculating
The class label based on each image classification level that machine or processor execute calculates separately the classification for the sample image
It loses further comprising the steps of using the level loss as the image classification level：For each image classification level, being based on should
The classification results of the class label of image classification level and the image classification level of the output layer of neural network output, meter
Calculate the level loss of the image classification level.
In one embodiment, the neural network includes：M convolutional layer and N number of full articulamentum.Wherein, the M volume
N number of convolutional layer for connecting with N number of full articulamentum and being connect one by one with N number of full articulamentum one by one respectively in lamination
Including M convolutional layer, the output of N number of full articulamentum is corresponded with N number of image classification level respectively, with the M volumes
The output of the full articulamentum of lamination connection corresponds to the image classification of n-th layer time, and M is the integer greater than N.
In the above-described embodiments, when the computer program instructions are run by computer or processor, so that the calculating
The class label based on each image classification level that machine or processor execute calculates separately the classification for the sample image
It loses further comprising the steps of using the level loss as the image classification level：For each image classification level, by with this
The corresponding full articulamentum of image classification level exports the classification results of the image classification level；Based on the classification results and the image
The class label of taxonomical hierarchy calculates the level loss of the image classification level.
Illustratively, when the computer program instructions are run by computer or processor, so that the computer or place
Manage that device executes according in the step of the level costing bio disturbance total losses of each image classification level, can be according to following public affairs
Formula calculates the total losses Loss：
Wherein, LossiIt is the level loss of i-th of image classification level, WiIt is the layer for controlling i-th of image classification level
Secondary loss LossiEffect degree parameter, 0<i<N+1.
Illustratively, when the computer program instructions are run by computer or processor, so that the computer or place
It manages device and the parameter W is determined by searching algorithmi。
Illustratively, when the computer program instructions are run by computer or processor, so that the computer or place
Reason device receives the parameter W being arranged based on experience valuei。
Illustratively, described image identification includes the scene Recognition of image.
Each module in the training system of neural network according to an embodiment of the present invention for image recognition can pass through
The processor computer program instructions that store in memory of operation of electronic equipment according to an embodiment of the present invention realize, or
The computer that person can store in the computer readable storage medium of computer program product according to an embodiment of the present invention refers to
Realization when order is run by computer.
Training method, device, system and the storage of neural network according to an embodiment of the present invention for image recognition are situated between
Matter has the characteristics that the characteristic of division of many levels using image in image recognition, gives sample image mark multiple images classification
The class label of level, guidance neural network extract the characteristic of division of each taxonomical hierarchy, and such as more macroscopical other point of rough sort
Category feature and the other characteristic of division of more microcosmic disaggregated classification, to reach better image recognition effect.Particularly with content
The identification of complicated image, preferably can not directly extract characteristic of division, utilize the nerve of the method training of the embodiment of the present invention
Network, can extract the characteristic of division of many levels, and the characteristic of division of comprehensive many levels carries out classification judgement, may be implemented compared with
Good image recognition.
Although describing example embodiment by reference to attached drawing here, it should be understood that above example embodiment are only exemplary
, and be not intended to limit the scope of the invention to this.Those of ordinary skill in the art can carry out various changes wherein
And modification, it is made without departing from the scope of the present invention and spiritual.All such changes and modifications are intended to be included in appended claims
Within required the scope of the present invention.
Those of ordinary skill in the art may be aware that list described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and algorithm steps can be realized with the combination of electronic hardware or computer software and electronic hardware.These functions are actually
It is implemented in hardware or software, the specific application and design constraint depending on technical solution.Professional technician
Each specific application can be used different methods to achieve the described function, but this realization is it is not considered that exceed
The scope of the present invention.
In several embodiments provided herein, it should be understood that disclosed device and method can pass through it
Its mode is realized.For example, apparatus embodiments described above are merely indicative, for example, the division of the unit, only
Only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple units or components can be tied
Another equipment is closed or is desirably integrated into, or some features can be ignored or not executed.
In the instructions provided here, numerous specific details are set forth.It is to be appreciated, however, that implementation of the invention
Example can be practiced without these specific details.In some instances, well known method, structure is not been shown in detail
And technology, so as not to obscure the understanding of this specification.
Similarly, it should be understood that in order to simplify the present invention and help to understand one or more of the various inventive aspects,
To in the description of exemplary embodiment of the present invention, each feature of the invention be grouped together into sometimes single embodiment, figure,
Or in descriptions thereof.However, the method for the invention should not be construed to reflect following intention：It is i.e. claimed
The present invention claims features more more than feature expressly recited in each claim.More precisely, such as corresponding power
As sharp claim reflects, inventive point is that the spy of all features less than some disclosed single embodiment can be used
Sign is to solve corresponding technical problem.Therefore, it then follows thus claims of specific embodiment are expressly incorporated in this specific
Embodiment, wherein each, the claims themselves are regarded as separate embodiments of the invention.
It will be understood to those skilled in the art that any combination pair can be used other than mutually exclusive between feature
All features disclosed in this specification (including adjoint claim, abstract and attached drawing) and so disclosed any method
Or all process or units of equipment are combined.Unless expressly stated otherwise, this specification (is wanted including adjoint right
Ask, make a summary and attached drawing) disclosed in each feature can be replaced with an alternative feature that provides the same, equivalent, or similar purpose.
In addition, it will be appreciated by those of skill in the art that although some embodiments described herein include other embodiments
In included certain features rather than other feature, but the combination of the feature of different embodiments mean it is of the invention
Within the scope of and form different embodiments.For example, in detail in the claims, embodiment claimed it is one of any
Can in any combination mode come using.
Various component embodiments of the invention can be implemented in hardware, or to run on one or more processors
Software module realize, or be implemented in a combination thereof.It will be understood by those of skill in the art that can be used in practice
Microprocessor or digital signal processor (DSP) realize the neural network according to an embodiment of the present invention for image recognition
Training device in some modules some or all functions.The present invention is also implemented as described here for executing
Method some or all program of device (for example, computer program and computer program product).Such realization
Program of the invention can store on a computer-readable medium, or may be in the form of one or more signals.This
The signal of sample can be downloaded from an internet website to obtain, and is perhaps provided on the carrier signal or mentions in any other forms
For.
It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and ability
Field technique personnel can be designed alternative embodiment without departing from the scope of the appended claims.In the claims,
Any reference symbol between parentheses should not be configured to limitations on claims.Word "comprising" does not exclude the presence of not
Element or step listed in the claims.Word "a" or "an" located in front of the element does not exclude the presence of multiple such
Element.The present invention can be by means of including the hardware of several different elements and being come by means of properly programmed computer real
It is existing.In the unit claims listing several devices, several in these devices can be through the same hardware branch
To embody.The use of word first, second, and third does not indicate any sequence.These words can be explained and be run after fame
Claim.
The above description is merely a specific embodiment or to the explanation of specific embodiment, protection of the invention
Range is not limited thereto, and anyone skilled in the art in the technical scope disclosed by the present invention, can be easily
Expect change or replacement, should be covered by the protection scope of the present invention.Protection scope of the present invention should be with claim
Subject to protection scope.
Claims (13)
1. a kind of training method of the neural network for image recognition, including：
Receive be labelled with N number of image classification level class label sample image, wherein in adjacent image taxonomical hierarchy under
The classification of one level is the subclass of the classification of a level thereon, and N is the integer greater than 1；
Neural network is trained using the sample image, wherein
Based on the class label of each image classification level, calculate separately the Classification Loss for the sample image, using as
The level of the image classification level loses；
According to the level costing bio disturbance total losses of each image classification level；
Using the total losses as the parameter of the objective function training neural network.
2. the method for claim 1, wherein class label based on each image classification level, calculates separately
For the Classification Loss of the sample image, using the level loss as the image classification level, including：
For each image classification level, the output layer of class label and the neural network based on the image classification level is defeated
The classification results of the image classification level out calculate the level loss of the image classification level.
3. the method for claim 1, wherein the neural network includes：M convolutional layer and N number of full articulamentum,
In,
N number of in the M convolutional layer connect with N number of full articulamentum one by one respectively, and with N number of full articulamentum one
One connection convolutional layer include M convolutional layer, the output of N number of full articulamentum respectively with N number of image classification level one
One is corresponding, and the output for the full articulamentum connecting with M convolutional layer corresponds to the image classification of n-th layer time, and M is the integer greater than N.
4. method as claimed in claim 3, wherein the class label based on each image classification level calculates separately
For the Classification Loss of the sample image, using the level loss as the image classification level, including：
For each image classification level, which is exported by full articulamentum corresponding with the image classification level
Classification results；Class label based on the classification results He the image classification level calculates the level damage of the image classification level
It loses.
5. such as the described in any item methods of Claims 1-4, wherein the level according to each image classification level
Costing bio disturbance total losses includes：
The total losses Loss is calculated according to the following formula：
Wherein, LossiIt is the level loss of i-th of image classification level, WiIt is the level damage for controlling i-th of image classification level
Lose LossiEffect degree parameter, 0<i<N+1.
6. method as claimed in claim 5, wherein the method also includes：
The parameter W is determined by searching algorithmi。
7. method as claimed in claim 5, wherein the method also includes：
Receive the parameter W being arranged based on experience valuei。
8. such as the described in any item methods of Claims 1-4, wherein the class label of N number of image classification level includes：
According to the class label for N number of image classification level that similarity degree between class is classified；Or
According to the class label for N number of image classification level that the difference for distinguishing difficulty is classified.
9. such as the described in any item methods of Claims 1-4, wherein described image identification includes the scene Recognition of image.
10. a kind of image-recognizing method neural network based, including：
Obtain images to be recognized；
Image recognition is carried out to the images to be recognized using trained neural network, wherein the neural network passes through power
Benefit requires 1 to 9 described in any item method training to obtain.
11. a kind of training device of the neural network for image recognition, including：
Receiving module receives the sample image for being labelled with the class label of N number of image classification level, wherein adjacent image classification layer
The classification of next level in secondary is the subclass of the classification of a level thereon, and N is the integer greater than 1；
Training module is trained neural network using the sample image, wherein
Based on the class label of each image classification level, calculate separately the Classification Loss for the sample image, using as
The level of the image classification level loses；
According to the level costing bio disturbance total losses of each image classification level；
Using the total losses as the parameter of the objective function training neural network.
12. a kind of training system of the neural network for image recognition, including processor and memory, wherein the storage
Computer program instructions are stored in device, for executing as right is wanted when the computer program instructions are run by the processor
Seek the training method of 1 to 9 described in any item neural networks for image recognition.
13. a kind of storage medium stores program instruction on said storage, described program instruction is at runtime for holding
The training method of the row neural network as described in any one of claim 1 to 9 for image recognition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810525605.0A CN108875934A (en) | 2018-05-28 | 2018-05-28 | A kind of training method of neural network, device, system and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810525605.0A CN108875934A (en) | 2018-05-28 | 2018-05-28 | A kind of training method of neural network, device, system and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108875934A true CN108875934A (en) | 2018-11-23 |
Family
ID=64335310
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810525605.0A Pending CN108875934A (en) | 2018-05-28 | 2018-05-28 | A kind of training method of neural network, device, system and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108875934A (en) |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109784159A (en) * | 2018-12-11 | 2019-05-21 | 北京航空航天大学 | The processing method of scene image, apparatus and system |
| CN109871883A (en) * | 2019-01-24 | 2019-06-11 | 北京市商汤科技开发有限公司 | Neural network training method and device, electronic equipment and storage medium |
| CN109886143A (en) * | 2019-01-29 | 2019-06-14 | 上海鹰瞳医疗科技有限公司 | Multi-tag disaggregated model training method and equipment |
| CN109948160A (en) * | 2019-03-15 | 2019-06-28 | 智者四海（北京）技术有限公司 | Short text classification method and device |
| CN110378410A (en) * | 2019-07-16 | 2019-10-25 | 北京字节跳动网络技术有限公司 | Multi-tag scene classification method, device and electronic equipment |
| CN110490242A (en) * | 2019-08-12 | 2019-11-22 | 腾讯医疗健康(深圳)有限公司 | Training method, eye fundus image classification method and the relevant device of image classification network |
| CN110738233A (en) * | 2019-08-28 | 2020-01-31 | 北京奇艺世纪科技有限公司 | Model training method, data classification method, device, electronic equipment and storage medium |
| CN111046944A (en) * | 2019-12-10 | 2020-04-21 | 北京奇艺世纪科技有限公司 | Method and device for determining object class, electronic equipment and storage medium |
| CN111291875A (en) * | 2018-12-06 | 2020-06-16 | 意法半导体(鲁塞)公司 | Method and apparatus for determining memory size |
| CN111476309A (en) * | 2020-04-13 | 2020-07-31 | 北京字节跳动网络技术有限公司 | Image processing method, model training method, device, equipment and readable medium |
| CN111507403A (en) * | 2020-04-17 | 2020-08-07 | 腾讯科技（深圳）有限公司 | Image classification method and device, computer equipment and storage medium |
| CN111507362A (en) * | 2019-01-30 | 2020-08-07 | 中强光电股份有限公司 | Neural network training method, neural network-based classification method and device |
| CN111523351A (en) * | 2019-02-02 | 2020-08-11 | 北京地平线机器人技术研发有限公司 | Neural network training method and device and electronic equipment |
| CN111523640A (en) * | 2020-04-09 | 2020-08-11 | 北京百度网讯科技有限公司 | Training method and device of neural network model |
| CN111598182A (en) * | 2020-05-22 | 2020-08-28 | 北京市商汤科技开发有限公司 | Method, apparatus, device and medium for training neural network and image recognition |
| CN111814813A (en) * | 2019-04-10 | 2020-10-23 | 北京市商汤科技开发有限公司 | Neural network training and image classification method and device |
| CN111860548A (en) * | 2019-04-02 | 2020-10-30 | 北京京东尚科信息技术有限公司 | Image recognition method and device, computer storage medium |
| CN112052789A (en) * | 2020-09-03 | 2020-12-08 | 腾讯科技（深圳）有限公司 | Face recognition method and device, electronic equipment and storage medium |
| CN113378853A (en) * | 2021-05-25 | 2021-09-10 | 北京有竹居网络技术有限公司 | Image classification method, device, equipment and storage medium |
| CN113743535A (en) * | 2019-05-21 | 2021-12-03 | 北京市商汤科技开发有限公司 | Neural network training method and device and image processing method and device |
| WO2022042123A1 (en) * | 2020-08-25 | 2022-03-03 | 深圳思谋信息科技有限公司 | Image recognition model generation method and apparatus, computer device and storage medium |
| CN115935279A (en) * | 2023-03-14 | 2023-04-07 | 有米科技股份有限公司 | Method and device for training hierarchical classification model and hierarchical classification |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103530403A (en) * | 2013-10-23 | 2014-01-22 | 天津大学 | Structured image description method |
| CN104021224A (en) * | 2014-06-25 | 2014-09-03 | 中国科学院自动化研究所 | Image labeling method based on layer-by-layer label fusing deep network |
| CN104484322A (en) * | 2010-09-24 | 2015-04-01 | 新加坡国立大学 | Methods and systems for automated text correction |
-
2018
- 2018-05-28 CN CN201810525605.0A patent/CN108875934A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104484322A (en) * | 2010-09-24 | 2015-04-01 | 新加坡国立大学 | Methods and systems for automated text correction |
| CN103530403A (en) * | 2013-10-23 | 2014-01-22 | 天津大学 | Structured image description method |
| CN104021224A (en) * | 2014-06-25 | 2014-09-03 | 中国科学院自动化研究所 | Image labeling method based on layer-by-layer label fusing deep network |
Non-Patent Citations (3)
| Title |
|---|
| ALEX KENDALL， YARIN GAL， ROBERTO CIPOLLA: "Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics", 《ARXIV》 * |
| HICHEM BANNOUR: "Hierarchical image annotation using semantic hierarchies", 《ACM》 * |
| XINQI ZHU: "B-CNN: Branch Convolutional Neural Network for Hierarchical Classification", 《ARXIV》 * |
Cited By (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111291875B (en) * | 2018-12-06 | 2023-10-24 | 意法半导体(鲁塞)公司 | Method and apparatus for determining memory size |
| CN111291875A (en) * | 2018-12-06 | 2020-06-16 | 意法半导体(鲁塞)公司 | Method and apparatus for determining memory size |
| CN109784159A (en) * | 2018-12-11 | 2019-05-21 | 北京航空航天大学 | The processing method of scene image, apparatus and system |
| CN109871883A (en) * | 2019-01-24 | 2019-06-11 | 北京市商汤科技开发有限公司 | Neural network training method and device, electronic equipment and storage medium |
| CN109871883B (en) * | 2019-01-24 | 2022-04-05 | 北京市商汤科技开发有限公司 | Neural network training method and device, electronic equipment and storage medium |
| CN109886143A (en) * | 2019-01-29 | 2019-06-14 | 上海鹰瞳医疗科技有限公司 | Multi-tag disaggregated model training method and equipment |
| CN111507362A (en) * | 2019-01-30 | 2020-08-07 | 中强光电股份有限公司 | Neural network training method, neural network-based classification method and device |
| CN111507362B (en) * | 2019-01-30 | 2023-12-19 | 中强光电股份有限公司 | Neural network training method, neural network-based classification method and device |
| CN111523351A (en) * | 2019-02-02 | 2020-08-11 | 北京地平线机器人技术研发有限公司 | Neural network training method and device and electronic equipment |
| CN109948160B (en) * | 2019-03-15 | 2023-04-18 | 智者四海（北京）技术有限公司 | Short text classification method and device |
| CN109948160A (en) * | 2019-03-15 | 2019-06-28 | 智者四海（北京）技术有限公司 | Short text classification method and device |
| CN111860548A (en) * | 2019-04-02 | 2020-10-30 | 北京京东尚科信息技术有限公司 | Image recognition method and device, computer storage medium |
| CN111814813A (en) * | 2019-04-10 | 2020-10-23 | 北京市商汤科技开发有限公司 | Neural network training and image classification method and device |
| CN113743535A (en) * | 2019-05-21 | 2021-12-03 | 北京市商汤科技开发有限公司 | Neural network training method and device and image processing method and device |
| CN110378410A (en) * | 2019-07-16 | 2019-10-25 | 北京字节跳动网络技术有限公司 | Multi-tag scene classification method, device and electronic equipment |
| CN110490242B (en) * | 2019-08-12 | 2024-03-29 | 腾讯医疗健康(深圳)有限公司 | Training method of image classification network, fundus image classification method and related equipment |
| CN110490242A (en) * | 2019-08-12 | 2019-11-22 | 腾讯医疗健康(深圳)有限公司 | Training method, eye fundus image classification method and the relevant device of image classification network |
| CN110738233B (en) * | 2019-08-28 | 2022-07-12 | 北京奇艺世纪科技有限公司 | Model training method, data classification method, device, electronic equipment and storage medium |
| CN110738233A (en) * | 2019-08-28 | 2020-01-31 | 北京奇艺世纪科技有限公司 | Model training method, data classification method, device, electronic equipment and storage medium |
| CN111046944A (en) * | 2019-12-10 | 2020-04-21 | 北京奇艺世纪科技有限公司 | Method and device for determining object class, electronic equipment and storage medium |
| CN111523640A (en) * | 2020-04-09 | 2020-08-11 | 北京百度网讯科技有限公司 | Training method and device of neural network model |
| CN111523640B (en) * | 2020-04-09 | 2023-10-31 | 北京百度网讯科技有限公司 | Training method and device for neural network model |
| CN111476309B (en) * | 2020-04-13 | 2023-05-23 | 抖音视界有限公司 | Image processing method, model training method, device, equipment and readable medium |
| CN111476309A (en) * | 2020-04-13 | 2020-07-31 | 北京字节跳动网络技术有限公司 | Image processing method, model training method, device, equipment and readable medium |
| CN111507403A (en) * | 2020-04-17 | 2020-08-07 | 腾讯科技（深圳）有限公司 | Image classification method and device, computer equipment and storage medium |
| CN111598182B (en) * | 2020-05-22 | 2023-12-01 | 北京市商汤科技开发有限公司 | Method, device, equipment and medium for training neural network and image recognition |
| CN111598182A (en) * | 2020-05-22 | 2020-08-28 | 北京市商汤科技开发有限公司 | Method, apparatus, device and medium for training neural network and image recognition |
| WO2022042123A1 (en) * | 2020-08-25 | 2022-03-03 | 深圳思谋信息科技有限公司 | Image recognition model generation method and apparatus, computer device and storage medium |
| CN112052789A (en) * | 2020-09-03 | 2020-12-08 | 腾讯科技（深圳）有限公司 | Face recognition method and device, electronic equipment and storage medium |
| CN113378853A (en) * | 2021-05-25 | 2021-09-10 | 北京有竹居网络技术有限公司 | Image classification method, device, equipment and storage medium |
| CN113378853B (en) * | 2021-05-25 | 2024-02-13 | 北京有竹居网络技术有限公司 | Image classification method, device, equipment and storage medium |
| CN115935279A (en) * | 2023-03-14 | 2023-04-07 | 有米科技股份有限公司 | Method and device for training hierarchical classification model and hierarchical classification |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108875934A (en) | A kind of training method of neural network, device, system and storage medium | |
| CN108596882B (en) | The recognition methods of pathological picture and device | |
| CN105488515B (en) | The image processing method and image processing apparatus that a kind of pair of image is classified | |
| CN108780519A (en) | Structure learning in convolutional neural networks | |
| CN108875522A (en) | Face cluster methods, devices and systems and storage medium | |
| CN108921058A (en) | Fish identification method, medium, terminal device and device based on deep learning | |
| CN109522967A (en) | A kind of commodity attribute recognition methods, device, equipment and storage medium | |
| CN105809146B (en) | A kind of image scene recognition methods and device | |
| CN110414377A (en) | A kind of remote sensing images scene classification method based on scale attention network | |
| CN109299716A (en) | Training method, image partition method, device, equipment and the medium of neural network | |
| CN107077624A (en) | Track hand/body gesture | |
| CN107563439A (en) | A kind of model for identifying cleaning food materials picture and identification food materials class method for distinguishing | |
| CN109117897A (en) | Image processing method, device and readable storage medium storing program for executing based on convolutional neural networks | |
| CN109559300A (en) | Image processing method, electronic equipment and computer readable storage medium | |
| CN109637207A (en) | A kind of preschool education interactive teaching device and teaching method | |
| CN110363233A (en) | A kind of the fine granularity image-recognizing method and system of the convolutional neural networks based on block detector and Fusion Features | |
| CN106529503A (en) | Method for recognizing face emotion by using integrated convolutional neural network | |
| CN108875932A (en) | Image-recognizing method, device and system and storage medium | |
| Blečić et al. | Towards automatic assessment of perceived walkability | |
| CN108875769A (en) | Data mask method, device and system and storage medium | |
| CN110222728A (en) | The training method of article discrimination model, system and article discrimination method, equipment | |
| CN108509833A (en) | A kind of face identification method, device and equipment based on structured analysis dictionary | |
| CN109635833A (en) | A kind of image-recognizing method and system based on cloud platform and model intelligent recommendation | |
| CN109886153A (en) | A kind of real-time face detection method based on depth convolutional neural networks | |
| CN110516734A (en) | A kind of image matching method, device, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181123 |