CN108229497A - Image processing method, device, storage medium, computer program and electronic equipment - Google Patents
Image processing method, device, storage medium, computer program and electronic equipment Download PDFInfo
- Publication number
- CN108229497A CN108229497A CN201710632941.0A CN201710632941A CN108229497A CN 108229497 A CN108229497 A CN 108229497A CN 201710632941 A CN201710632941 A CN 201710632941A CN 108229497 A CN108229497 A CN 108229497A
- Authority
- CN
- China
- Prior art keywords
- characteristic pattern
- feature
- convolutional layer
- network
- branching networks
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Abstract
An embodiment of the present invention provides a kind of image processing method, device, storage medium, computer program and electronic equipment, wherein, described image processing method includes：Obtain the characteristic pattern of image to be detected；At least two kinds of different scales are based on by neural network, feature extraction is carried out to the characteristic pattern, obtain at least two other characteristic patterns；Merge the characteristic pattern and other each described characteristic patterns, obtain the fisrt feature figure of described image to be detected.Using the technical solution of the embodiment of the present invention, neural network learning can be utilized and extract the feature of different scale, improve the accuracy and robustness of feature extraction.
Description
Technical field
The present embodiments relate to technical field of computer vision more particularly to a kind of image processing method, device, storages
Medium, computer program and electronic equipment.
Background technology
Human body attitude estimation mainly positions the position of human body each section in given image or video, is meter
One important research topic of calculation machine visual field is mainly used in action recognition, Activity recognition, clothes parsing, task pair
Than, human-computer interaction etc..
At present, estimation method of human posture depends on the feature that object detector is detected, existing object detector one
As be trained to obtain on fixed size.
Invention content
An embodiment of the present invention provides a kind of image procossing schemes.
It is according to embodiments of the present invention in a first aspect, provide a kind of image processing method, including：Obtain image to be detected
Characteristic pattern；At least two kinds of different scales are based on by neural network, feature extraction is carried out to the characteristic pattern, obtain at least two
Other a characteristic patterns；Merge the characteristic pattern and other each described characteristic patterns, obtain the fisrt feature figure of described image to be detected.
Optionally, it further includes：The target object in described image to be detected is carried out according to the fisrt feature figure crucial
Point detection.
Optionally, it is described that critical point detection is carried out to the target object according to the fisrt feature figure, including：According to institute
State the shot chart that fisrt feature figure obtains an at least key point for the target object respectively；It is wrapped according in each shot chart
The score of the pixel included determines the position of the corresponding key point of the target object.
Optionally, the neural network includes at least one feature pyramid sub-network, the feature pyramid sub-network
At least one second branching networks in parallel with first branching networks including the first branching networks and respectively；It is described other
Characteristic pattern includes second feature figure or third feature figure；The original scale of first branching networks based on the characteristic pattern is to institute
It states characteristic pattern and carries out feature extraction, obtain the second feature figure；Each second branching networks are based respectively on different from described
Other scales of original scale carry out feature extraction to the characteristic pattern, obtain the third feature figure.
Optionally, first branching networks include the second convolutional layer, third convolutional layer and Volume Four lamination；Described second
Convolutional layer reduces the dimension of the characteristic pattern；After the original scale of the third convolutional layer based on the characteristic pattern is to reducing dimension
Characteristic pattern carry out process of convolution；The Volume Four lamination promotes the dimension of the characteristic pattern Jing Guo process of convolution, acquisition described the
Two characteristic patterns.
Optionally, at least 1 second branching networks include the 5th convolutional layer, down-sampled layer, the 6th convolutional layer, on adopt
Sample layer and the 7th convolutional layer；5th convolutional layer reduces the dimension of the characteristic pattern；The down-sampled layer is adopted according to setting drop
Sample ratio is down-sampled to the characteristic pattern progress after reducing dimension, wherein, the scale of the characteristic pattern after down-sampled is less than described
The original scale of characteristic pattern；6th convolutional layer carries out process of convolution to described by down-sampled characteristic pattern；It is adopted on described
Sample layer up-samples the characteristic pattern Jing Guo convolution according to up-sampling ratio is set, wherein, the characteristic pattern after up-sampling
Scale be equal to the characteristic pattern original scale；7th convolutional layer promotes the dimension of the characteristic pattern after up-sampling,
Obtain the third feature figure.
Optionally, second branching networks have multiple；The down-sampled ratio of setting of at least two second branching networks
Example it is different and/or, the down-sampled ratio of setting of at least two second branching networks is identical.
Optionally, second branching networks have multiple；The 6th convolution of at least two second branching networks
Layer shared parameter.
Optionally, second branching networks include the 5th convolutional layer, expansion convolutional layer and the 7th convolutional layer；Described 5th
Convolutional layer reduces the dimension of the characteristic pattern；The expansion convolutional layer carries out expansion convolution to the characteristic pattern after reducing dimension
Processing, the 7th convolutional layer promote the dimension of the characteristic pattern after expanding convolution, obtain the third feature figure.
Optionally, second branching networks have multiple；At least two second branching networks share described volume five
Lamination and/or the 7th convolutional layer；And/or at least two second branching networks have respective 5th convolution
Layer and/or the 7th convolutional layer.
Optionally, the feature pyramid sub-network further includes the first output merging layer；It is right that first output merges layer
Share the 7th convolutional layer at least two second branching networks before the 7th convolutional layer respectively export into
Row merges and exports amalgamation result to shared the 7th convolutional layer.
Optionally, the neural network includes at least two feature pyramid sub-networks；Each feature pyramid subnet
Network, the fisrt feature figure exported using the previous feature pyramid sub-network being connect with current signature pyramid sub-network are input,
And according to the fisrt feature figure of input, the fisrt feature figure based on different scale extraction current signature pyramid sub-network.
Optionally, the neural network be hourglass HOURGLASS neural networks, the hourglass HOURGLASS neural networks
Including an at least hourglass module include at least one feature pyramid sub-network.
Optionally, the initialization network parameter of an at least network layer for the neural network, from according to the initialization net
It is obtained, and the mean value of the initialization network parameter is zero in the network parameter distribution that the mean value and variance of network parameter determine.
Optionally, if existing in neural network including the situation that at least two identical mappings are added, needing what is be added
Output adjustment module in an at least identical mapping branch is set, adjusts what the identical mapping branch exported by output adjustment module
Fisrt feature figure.
Second aspect according to embodiments of the present invention provides a kind of image processing apparatus, including：Acquisition module is used for
Obtain the characteristic pattern of image to be detected；Extraction module is based at least two kinds of different scales to the spy for passing through neural network
Sign figure carries out feature extraction, obtains at least two other characteristic patterns；Merging module, for merge the characteristic pattern and it is each it is described its
His characteristic pattern, obtains the fisrt feature figure of described image to be detected.
Optionally, it further includes：Detection module, for according to the fisrt feature figure to the target in described image to be detected
Object carries out critical point detection.
Optionally, the detection module includes：Subdivision is obtained, for obtaining the mesh respectively according to the fisrt feature figure
Mark the shot chart of an at least key point for object；Determination unit, for according to pixel included in each shot chart
Score determines the position of the corresponding key point of the target object.
Optionally, the neural network includes at least one feature pyramid sub-network, the feature pyramid sub-network
At least one second branching networks in parallel with first branching networks including the first branching networks and respectively；It is described other
Characteristic pattern includes second feature figure or third feature figure；First branching networks are used for the original scale based on the characteristic pattern
Feature extraction is carried out to the characteristic pattern, obtains the second feature figure；Each second branching networks are used to be based respectively on not
Other scales for being same as the original scale carry out feature extraction to the characteristic pattern, obtain the third feature figure.
Optionally, first branching networks include the second convolutional layer, third convolutional layer and Volume Four lamination；Described second
Convolutional layer for reducing the characteristic pattern dimension；The third convolutional layer is for the original scale based on the characteristic pattern to drop
Characteristic pattern after low dimensional carries out process of convolution；The Volume Four lamination is used to be promoted the dimension of the characteristic pattern Jing Guo process of convolution
Degree, obtains the second feature figure.
Optionally, at least 1 second branching networks include the 5th convolutional layer, down-sampled layer, the 6th convolutional layer, on adopt
Sample layer and the 7th convolutional layer；5th convolutional layer for reducing the characteristic pattern dimension；The down-sampled layer is used for basis
It is down-sampled to the characteristic pattern progress after reducing dimension to set down-sampled ratio, wherein, the scale of the characteristic pattern after down-sampled
Less than the original scale of the characteristic pattern；6th convolutional layer is used to carry out at convolution by down-sampled characteristic pattern to described
Reason；The up-sampling layer is used to, according to setting up-sampling ratio, up-sample the characteristic pattern Jing Guo convolution, wherein, pass through
The scale of characteristic pattern after up-sampling is equal to the original scale of the characteristic pattern；7th convolutional layer is above adopted for being promoted to pass through
The dimension of characteristic pattern after sample obtains the third feature figure.
Optionally, second branching networks have multiple；The down-sampled ratio of setting of at least two second branching networks
Example it is different and/or, the down-sampled ratio of setting of at least two second branching networks is identical.
Optionally, second branching networks have multiple；The 6th convolution of at least two second branching networks
Layer shared parameter.
Optionally, second branching networks include the 5th convolutional layer, expansion convolutional layer and the 7th convolutional layer；Described 5th
Convolutional layer for reducing the characteristic pattern dimension；The expansion convolutional layer is used to carry out the characteristic pattern after reduction dimension
Expand process of convolution；7th convolutional layer obtains the third for promoting the dimension of the characteristic pattern after expanding convolution
Characteristic pattern.
Optionally, second branching networks have multiple；At least two second branching networks share described volume five
Lamination and/or the 7th convolutional layer；And/or at least two second branching networks have respective 5th convolution
Layer and/or the 7th convolutional layer.
Optionally, the feature pyramid sub-network further includes the first output merging layer；First output merges layer and uses
In respective defeated before the 7th convolutional layer at least two second branching networks for sharing the 7th convolutional layer
Go out to merge and export amalgamation result to shared the 7th convolutional layer.
Optionally, the neural network includes at least two feature pyramid sub-networks；Each feature pyramid subnet
Network, the fisrt feature figure for being exported using the previous feature pyramid sub-network being connect with current signature pyramid sub-network is defeated
Enter, and according to the fisrt feature figure of input, the fisrt feature figure based on different scale extraction current signature pyramid sub-network.
Optionally, the neural network be hourglass HOURGLASS neural networks, the hourglass HOURGLASS neural networks
Including an at least hourglass module include at least one feature pyramid sub-network.
Optionally, the initialization network parameter of an at least network layer for the neural network, from according to the initialization net
It is obtained, and the mean value of the initialization network parameter is zero in the network parameter distribution that the mean value and variance of network parameter determine.
Optionally, if existing in neural network including the situation that at least two identical mappings are added, needing what is be added
Output adjustment module is set in an at least identical mapping branch, and the output adjustment module is defeated for adjusting the identical mapping branch
The fisrt feature figure gone out.
The third aspect according to embodiments of the present invention provides a kind of computer readable storage medium, is stored thereon with meter
Calculation machine program instruction, wherein, described program instructs the step of any one of aforementioned image processing method is realized when being executed by processor.
Fourth aspect according to embodiments of the present invention, provides a kind of electronic equipment, including：Processor, memory, communication
Element and communication bus, the processor, the memory and the communication device are completed mutual by the communication bus
Communication；For the memory for storing an at least executable instruction, it is aforementioned that the executable instruction performs the processor
The corresponding operation of image processing method of any one.
5th aspect according to embodiments of the present invention, provides a kind of computer program, including：At least one executable finger
It enables, it is described that any one of aforementioned image processing method corresponding behaviour is used to implement when at least an executable instruction is processed by the processor
Make.
Image procossing scheme according to embodiments of the present invention after the characteristic pattern for obtaining image to be detected, passes through nerve
Network carries out feature extraction to obtain other multiple characteristic patterns based on a variety of different scales to characteristic pattern, and by characteristic pattern with it is multiple
Other characteristic patterns merge to obtain the fisrt feature figure of image to be detected, utilize the spy of neural network learning and extraction different scale
Sign improves accuracy and robustness that neural network carries out feature extraction.
Description of the drawings
Fig. 1 is a kind of step flow chart of according to embodiments of the present invention one image processing method；
Fig. 2 is a kind of step flow chart of according to embodiments of the present invention two image processing method；
Fig. 3 is the first structure schematic diagram of according to embodiments of the present invention two feature pyramid sub-network；
Fig. 4 is the second structure diagram of according to embodiments of the present invention two feature pyramid sub-network；
Fig. 5 is the third structure diagram of according to embodiments of the present invention two feature pyramid sub-network；
Fig. 6 is a kind of structure diagram of according to embodiments of the present invention two neural network for image procossing；
Fig. 7 is a kind of structure diagram of according to embodiments of the present invention two HOURGLASS networks；
Fig. 8 is the shot chart of according to embodiments of the present invention two image processing method output；
Fig. 9 is a kind of structure diagram of according to embodiments of the present invention two identical mapping addition；
Figure 10 is a kind of structure diagram of according to embodiments of the present invention three image processing apparatus；
Figure 11 is the structure diagram of according to embodiments of the present invention four a kind of electronic equipment.
Specific embodiment
(identical label represents identical element in several attached drawings) and embodiment below in conjunction with the accompanying drawings, implement the present invention
The specific embodiment of example is described in further detail.Following embodiment is used to illustrate the present invention, but be not limited to the present invention
Range.
It will be understood by those skilled in the art that the terms such as " first ", " second " in the embodiment of the present invention are only used for distinguishing
Different step, equipment or module etc. neither represent any particular technology meaning, also do not indicate that the inevitable logic between them is suitable
Sequence.
Embodiment one
With reference to Fig. 1, a kind of step flow chart of according to embodiments of the present invention one image processing method is shown.
The image processing method of the present embodiment includes the following steps：
Step S102：Obtain the characteristic pattern of image to be detected.
In the present embodiment, arbitrary image analysis processing method may be used to carry out at feature extraction image to be detected
Reason, to obtain the characteristic pattern of image to be detected.Optionally, feature is carried out to image to be detected for example, by convolutional neural networks to carry
Extract operation obtains the characteristic pattern (Feature Map) for the characteristic information for including image to be detected.Wherein, image to be detected can be with
It is any one frame image in independent still image or video sequence.
Illustrate herein, the characteristic pattern of acquisition can be the global characteristics figure of image to be detected or the non-overall situation
Characteristic pattern, the present embodiment are not construed as limiting this.For example, it in practical applications, is used to carry out at image according to the characteristic pattern of acquisition
The different application scenarios such as reason or object identification, can obtain the global characteristics figure of image to be detected or including object respectively
The local feature figure of body.
Step S104：At least two kinds of different scales are based on by neural network, feature extraction is carried out to characteristic pattern, obtained extremely
Few two other characteristic patterns.
Wherein, at least two other characteristic patterns are characteristic pattern of the neural network to image to be detected, based at least two kinds not
Carry out the characteristic pattern of further feature extraction operation acquisition respectively with scale, each scale corresponds to other features
Figure.
Neural network carries out the scale that feature extraction operation is based on, and can limit the feature that feature extraction operation is extracted
Scale.In the embodiment of the present invention, neural network is based on different scale and carries out feature extraction to image to be detected, passes through nerve net
Network learns and the feature of extraction different scale, can stablize the feature for accurately extracting image to be detected.The embodiment of the present invention
The problem of characteristic dimension for occurring causing image to be detected the problems such as example blocking, have an X-rayed sends variation can be successfully managed, from
And improve the robustness of feature extraction.
In practical applications, feature extraction is based on scale is different, can be image physics size dimension it is different or
The size of the live part of person's image it is different (although for example, the physics size dimension of image is identical, the partial pixel of the image
Pixel value used but be not limited to the modes such as zero setting and handled, in addition to the portion of other pixels composition of these treated pixels
Divide and be equivalent to live part, the physical size of the size relative image of live part is smaller) etc., but not limited to this.
Optionally, at least two kinds of different scales can include the original scale of image to be detected with being different from original scale
At least one scale, alternatively, at least two kinds of different scales including being different from original scale.
Step S106：Merge characteristic pattern and other each characteristic patterns, obtain the fisrt feature figure of image to be detected.
Characteristic pattern and other each characteristic patterns are merged to obtain fisrt feature figure so that fisrt feature figure includes extracting
Different scale feature.Merge obtained fisrt feature figure to can be used for carrying out subsequent image procossing, example to image to be detected
Such as critical point detection, object detection, object identification, image segmentation, object cluster can improve the effect of subsequent image procossing
Fruit.
Image processing method according to embodiments of the present invention after the characteristic pattern for obtaining image to be detected, passes through nerve
Network carries out feature extraction to obtain other multiple characteristic patterns based on a variety of different scales to characteristic pattern, and by characteristic pattern with it is multiple
Other characteristic patterns merge to obtain the fisrt feature figure of image to be detected, utilize the spy of neural network learning and extraction different scale
Sign improves accuracy and robustness that neural network carries out feature extraction.
Any image processing method provided in an embodiment of the present invention can have data-handling capacity by any suitable
Equipment perform, including but not limited to：Terminal device and server etc..Alternatively, any image provided in an embodiment of the present invention
Processing method can be performed by processor, as processor by the command adapted thereto that memory is called to store performs implementation of the present invention
Any image processing method that example refers to.Hereafter repeat no more.
Embodiment two
With reference to Fig. 2, a kind of step flow chart of according to embodiments of the present invention two image processing method is shown.
The image processing method of the present embodiment includes the following steps：
Step S202：Obtain the characteristic pattern of image to be detected.
In the present embodiment, feature extraction operation is carried out to image to be detected by neural network to obtain characteristic pattern.For example,
Neural network includes the convolutional layer (Convolution, Conv) for carrying out feature extraction, to the to be detected of input neural network
Image carries out Preliminary detection and feature extraction operation, and acquisition includes the initial characteristic pattern of image to be detected.
Step S204：At least two kinds of different scales are based on by neural network, feature extraction is carried out to characteristic pattern, obtained extremely
Few two other characteristic patterns.
Optionally, neural network includes at least one feature pyramid sub-network, for being based at least two kinds of different scales
Feature extraction is carried out to characteristic pattern, obtains at least two other characteristic patterns.Feature pyramid includes the first branching networks and divides
At least one second branching networks not in parallel with the first branching networks.Original ruler of first branching networks based on image to be detected
Degree, to input feature vector, pyramidal characteristic pattern carries out further feature extraction, obtains second feature figure；Each second branching networks
Further feature extraction is carried out to characteristic pattern based on other scales different from the original scale, obtains third feature figure.
That is, at least two other characteristic patterns include second feature figure and third feature figure.
In a kind of optional embodiment, with reference to Fig. 3, the first branching networks include the second convolutional layer (Convolutio 2,
Conv 2), third convolutional layer (Conv 3) and Volume Four lamination (Conv 4).At least one second branching networks include the 5th convolution
Layer (Conv 5), down-sampled layer, the 6th convolutional layer (Conv 6), up-sampling layer and the 7th convolutional layer (Conv 7).
First branching networks are f0, each second branching networks are respectively f1To fc, wherein, f0Retain the original of input feature vector
Scale.The characteristic pattern of input feature vector pyramid sub-network is separately input to f0To fc。f0The second convolutional layer and f1To fc
1 × 1 convolutional network may be used in five convolutional layers, for reducing the dimension of input feature vector figure.f1To fcDown-sampled layer difference
According to the down-sampled ratio Ratio 1 to Ratio c of setting, respectively to the feature after the reduction dimension of each 5th convolutional layer output
Figure progress is down-sampled, obtains the characteristic pattern of different resolution.Wherein, the scale of the characteristic pattern after down-sampled is less than characteristic pattern
Original scale.f0Third convolutional layer and f1To fcThe 6th convolutional layer may be used 3 × 3 convolutional network, for point
Characteristic pattern and corresponding down-sampled layer after other the reductions dimension to the output of the second convolutional layer export by down-sampled spy
Sign figure carries out the feature of convolution, study and extraction different scale.f1To fcUp-sampling layer be based respectively on different up-sampling ratios
Example, to being up-sampled by the characteristic pattern of convolution for each 6th convolutional layer output, wherein, characteristic pattern after up-sampling
Scale is equal to the original scale of characteristic pattern.f0Volume Four lamination promote the feature by process of convolution of third convolutional layer output
The dimension of figure obtains second feature figure.f1To fcThe 7th convolutional layer promoted each corresponding up-sampling layer output by up-sampling
The dimension of characteristic pattern obtains third feature figure respectively.
Wherein, each second branching networks f1To fcIn, the down-sampled ratio of setting of at least two the second branching networks is different,
And/or the down-sampled ratio of setting of at least two the second branching networks is identical.That is, the drop that each second branching networks use is adopted
Sample ratio can be differed, can part it is identical, can also be all identical.For these three situations, and based on original scale
The first branching networks match, feature pyramid sub-network can extract different features based at least two kinds of different scales.
Further, since f0Retain the original scale of input feature vector, without changing the resolution ratio of feature, therefore, f0Do not use
Down-sampled layer and up-sampling layer, in practical applications, f0Down-sampled ratio and up-sampling ratio can also be used as 1 down-sampled layer
With up-sampling layer.
Optionally, the 6th convolutional layer shared parameter of at least two the second branching networks.For example, at least two the second branches
6th convolutional layer of network shares convolution kernel, that is, the convolution kernel of at least two the 6th convolutional layers has identical parameter, with logical
It crosses using inner parameter shared mechanism, to reduce number of parameters, while also is able to based on being obtained by data and tasking learning
The higher accuracy rate of gain of parameter.
In another optional embodiment, the structure type of the feature pyramid sub-network shown in Fig. 4 can also be used,
At least one second branching networks include the 5th convolutional layer, expansion convolutional layer and the 7th convolutional layer；5th convolutional layer reduces characteristic pattern
Dimension；Expansion convolutional layer carries out expansion process of convolution to the characteristic pattern after reducing dimension；7th convolutional layer is promoted by expansion
The dimension of characteristic pattern after convolution obtains third feature figure.That is, by at least down-sampled layer of one second branching networks, the 6th
(dilated convolution are expressed as dstride 1 to dstride in figure by expansion convolutional layer for convolutional layer and up-sampling layer
C) it replaces, simplifies the network structure inside feature pyramid sub-network, and the resolution ratio of input feature vector can be increased, utilize expansion
Convolutional layer completes the sampling operation of different resolution feature, and the extraction operation of different scale feature and similary resolution ratio is special
Sampling operation of sign etc., so as to obtain the feature of different scale.Wherein, expansion process of convolution can also realize it is down-sampled for example,
In a manner that the pixel value of the one part of pixel of characteristic pattern is set to 0, in the physical size situation of the same size for keeping image
Under, the part for having valid pixel value in characteristic pattern is become smaller, equally also achieves down-sampled effect.
Optionally, at least two the second branching networks share the 5th convolutional layer and/or the 7th convolutional layer；And/or at least two
A second branching networks have respective 5th convolutional layer and/or the 7th convolutional layer.
For example, in order to simplify the structure of feature pyramid sub-network, the feature pyramid sub-network shown in Fig. 5 may be used
Structure type, at least will share same 5th convolutional layer by two the second branching networks.For example, the 5th convolutional layer is 1 × 1
Convolutional network after the feature of input feature vector pyramid sub-network is carried out dimension-reduction treatment, is exported to shared 5th convolutional layer
Each second branching networks down-sampled layer.The number of parameters of the feature pyramid sub-network of the structure is less, computation complexity
It is relatively low.
Optionally, feature pyramid sub-network further includes the first output merging layer, and the first output merges layer to sharing the 7th
Respective output of at least two the second branching networks of convolutional layer before the 7th convolutional layer merges and amalgamation result is defeated
Go out to the 7th shared convolutional layer.
It is connected between shared 7th convolutional layer up-sampling layer and the 7th convolutional layer, is used for for example, the first output merges layer
It merges processing to the characteristic pattern that the up-sampling layers of each second branching networks exports, and the characteristic pattern after merging is exported to the
Seven convolutional layers.Here, merging treatment can include phase add operation or serial operation.For example, shown in figureRepresent output
Phase add operation, in figureIt may be replaced byTo represent output serial operation (Concatenation).Wherein, it is added behaviour
Work can be expressed as the point-to-point addition of multiple tensors, and serial operation can be expressed as string of multiple tensors in a dimension
Connection.If c the second branching networks f1To fcThe characteristic pattern of c 256 × 64 × 64 is exported, after phase add operation or 256 × 64 ×
64 characteristic pattern is concatenated then becoming after operating the characteristic pattern of (256 × c) × 64 × 64.
In addition, the 7th convolutional layer, which is additionally operable to the feature for exporting each second branching networks, carries out linear transformation, so as to the
The feature of the original scale of one branching networks output is added.It is grasped if the first output merges the series connection that is consolidated into that layer carries out
Make, the characteristic pattern that the 7th convolutional layer is additionally operable to merge the first output layer output carries out mapping transformation processing, and characteristic pattern is reflected
Penetrate the size of the characteristic pattern before being transformed to series connection.For example, it is 256 by the characteristic pattern mapping transformation of above-mentioned (256 × c) × 64 × 64
× 64 × 64 characteristic pattern.
Step S206：Merge characteristic pattern and other each characteristic patterns, obtain the fisrt feature figure of image to be detected.
Optionally, feature pyramid sub-network further includes the second output merging layer, the first branching networks and each second branch
The output terminal of network is connected to the second output and merges layer, and here, the output terminal of the second branching networks includes shared volume seven
The output terminal of the up-sampling layer of each second branching networks of the output terminal of lamination and not shared 7th convolutional layer.Second output
Merge layer to be used for characteristic pattern, the second feature figure of the first branching networks output and the third of each second branching networks output
Characteristic pattern merges processing, obtains fisrt feature figure.Here, it is consolidated into phase add operation.
In the present embodiment, neural network includes at least two feature pyramid sub-networks；Each feature pyramid sub-network, with
The fisrt feature figure of previous feature pyramid sub-network output being connect with current signature pyramid sub-network is input, and according to
The fisrt feature figure of input, the fisrt feature figure based on different scale extraction current signature pyramid sub-network.Wherein, first spy
The input of the golden word sub-network of sign is the characteristic pattern that step S202 is obtained, and performs step S204 to step S206 and obtains fisrt feature
Figure；Fisrt feature figure of the input of non-first feature gold word sub-network for the output of previous feature pyramid sub-network, and perform step
Rapid S204 to step S206 carries out feature extraction, by acquisition based at least two kinds of different scales to the fisrt feature figure of input
Other each characteristic patterns are merged with the fisrt feature figure inputted, obtain the fisrt feature figure of current signature pyramid sub-network.
In the present embodiment, each sub-neural network includes multiple feature pyramid sub-networks, previous feature pyramid subnet
The output of network can be the input of adjacent latter feature pyramid sub-network.If for example, x(l)And W(l)Represent l-th of feature
The input (characteristic pattern) of pyramid sub-network and the output of parameter, then this feature pyramid sub-network, that is, next feature is golden
The input of word tower sub-network can be expressed as：
x(l+1)=x(l)+p(x(l)+W(l)) (1)
Wherein, p (x(l)+W(l)) feature extraction operation performed by feature pyramid sub-network, and can be into one
Step is expressed as：
Wherein, c is the number of the second branching networks,Represent each second branching networks fcPerformed feature
Extraction operation,Represent the first branching networks f0Performed feature extraction operation,Represent the 7th convolutional layer
Performed processing.
In practical applications, neural network can be by using feature pyramid sub-network as basic comprising modules, utilizing feature
Pyramid study mechanism, to extract the feature of different scale.
In a kind of optional embodiment, HOURGLASS shown in Fig. 6 (hourglass) network structure can be used in neural network
As a kind of optional basic network topology, but not limited to this.Multiple HOURGLASS structures ends pair that neural network structure includes
End connection, forms HOURGLASS network structures, and each HOURGLASS structures include at least one feature pyramid sub-network.Before
Input of the output of one HOURGLASS structures for adjacent latter HOURGLASS structures, passes through this network structure so that from
Penetration model is analyzed and learnt in bottom always upward, top-downly so that neural network extraction feature it is more efficient and
Accurately, ensure the accuracy of the fisrt feature figure obtained.Wherein, since HOURGLASS networks use residual error module (Residual
Unit basic comprising modules) are used as, therefore, the feature pyramid sub-network of the present embodiment can be to be used to form HOURGLASS
The feature pyramid residual error module (Pyramids Residual Module, PRM) of network structure.Here, HOURGLASS structures
And the quantity of feature pyramid sub-network can be suitably set according to actual needs.
In HOURGLASS network structures shown in Fig. 7, each HOURGLASS structures can be by multiple feature pyramid subnets
Network forms, and with the feature for being learnt using feature pyramid sub-network and being extracted different scale, and exports fisrt feature figure.Its
In, the structure of any feature pyramid sub-network shown in above-mentioned Fig. 3 to Fig. 5 may be used in feature pyramid sub-network.Its
In, the neural network shown in Fig. 7 further includes the first convolutional layer (Conv1), and characteristic pattern is obtained available for performing abovementioned steps S202；
And pond layer (Pooling, Pool), the resolution ratio for constantly reducing characteristic pattern can be used, it, then will be global to obtain global characteristics
The position that resolution ratio is corresponded in feature interpolation amplification and characteristic pattern combines, that is, by carrying out global pool to characteristic pattern, obtains
Take the characteristic pattern of image to be detected.The characteristic pattern of acquisition can be with input feature vector pyramid sub-network so that feature pyramid subnet
Network carries out characteristic pattern deeper study and extraction, and then extracts fisrt feature figure based on different scale.Optionally, in pond
Feature pyramid sub-network or convolutional layer can also be set between layer and feature pyramid sub-network by changing, for adjusting characteristic pattern
The attributes such as resolution ratio.
Step S208：Critical point detection is carried out to the target object in image to be detected according to fisrt feature figure.
Optionally, the shot chart of an at least key point for target object is obtained respectively according to fisrt feature figure；According to each getting
The score of included pixel in component determines the position of the corresponding key point of target object.Pass through feature pyramid subnet
The fisrt feature figure of image to be detected that network obtains, based on different scale come the feature of Detection and Extraction image to be detected, Ke Yiwen
The fixed feature for accurately detecting different scale on this basis, carries out critical point detection, effectively according to fisrt feature figure
Improve the accuracy of critical point detection.
In a kind of optional embodiment, for some key point, the highest position of shot chart mid-score represents detection
The key point position arrived.As shown in figure 8, corresponding with the image to be detected for inputting neural network, the shot chart of output corresponds to
Each key point of target object in image to be detected.Wherein, target object is behaved in image to be detected, including 16 keys
Point, such as hand, knee etc..By the position of highest scoring in 16 shot charts, the position of corresponding key point is determined, you can completion pair
The detection and localization of 16 key points.
In practical application scene, the image processing method of the embodiment of the present invention can be used but be not limited in progress human body attitude
Estimation, video understand analysis, Activity recognition and human-computer interaction, image segmentation, object cluster etc..
For example, when carrying out human body attitude estimation, image to be detected is inputted into neural network, utilizes feature pyramid subnet
Network is based on different scale and carries out feature extraction, and carries out critical point detection to target object according to the feature of extraction, so as to foundation
The position of each key point detected carries out human body attitude estimation.For example, obtain the corresponding pass of 16 shot charts shown in Fig. 8
The position (for example, coordinate) of key point, human body attitude can be accurately estimated according to the position of 16 key points.Due to this implementation
The image processing method of example extracts feature using feature pyramid study mechanism, can detect the target object of different scale,
So as to ensure the robustness of human body Attitude estimation.
For another example for the video sequence comprising target object, the image processing method of the present embodiment may be used, utilize
Feature pyramid study mechanism extracts the characteristic pattern of video frame images to stablize, and then accurately carries out the key point of target object
Positioning helps to realize video and understands analysis.
Optionally, the initialization network parameter of an at least network layer for the neural network of the present embodiment, joins from according to network
It is obtained in the network parameter distribution that several mean values and variance determines.Wherein, network parameter distribution can be the Gauss of a setting
It is distributed or is uniformly distributed, the mean value and variance of network parameter distribution are determined by the number that outputs and inputs with parameter layer, just
Beginningization network parameter stochastical sampling can be obtained from network parameter distribution.The parameter initialization method can be to multiple-limb
The neural network of network structure is trained, which is not only applicable in what is proposed based on single branching networks, is also applicable to have
The problem of residual module of feature pyramid for having multiple-limb network is trained so that the training process of neural network is more stablized.
For example, in network parameter initialization procedure, for neural network propagated forward process, by the mean value of network parameter
0 is initialized as, to ensure that the variance output and input of each layer of neural network is basically identical.In the variance for obtaining network parameter
After σ, it is possible to from mean value be 0, variance is the Gaussian Profile of σ or initialization network parameter is adopted in being uniformly distributed
Sample, the initialization network parameter as propagated forward process.For neural network back-propagating process, by the mean value of network parameter
It is initialized as 0 so that the mean value of the gradient of network parameter is 0, and gradient is output and input so as to ensure each layer of neural network
Variance it is basically identical.Obtaining the variances sigma of gradient of network parameter ' after, it is possible to from mean value it is 0, the side of gradient
Difference is the Gaussian Profile of σ ' or initialization network parameter is sampled in being uniformly distributed, the initialization as back-propagating process
Network parameter.
Optionally, if in the presence of the feelings for including at least two identical mapping (Identity Mapping) additions in neural network
Shape then sets output adjustment module in at least identical mapping branch for needing to be added, should by the adjustment of output adjustment module
The fisrt feature figure of identical mapping branch output.
It for example, (might as well be with two shown in Fig. 9 if there is the situation that at least two identical mappings are added in neural network
Illustrated for a), then BN-ReLU-Conv (batch normalization- are set in some identical mapping branch
Rectified Linear Units-Convolution, batch standardization-activation primitive-convolution) module, it is identical to adjust this
Parameters, the so treated the output phase added-time in two identical mappings such as the range of variance of mapping branch output can avoid this
The variance that Liang Ge identical mappings branch leads to the problem of output response is multiplied, and is conducive to keep neural network learning process
Stability.It is illustrated by taking the situation that two identical mappings shown in Fig. 9 are added as an example, it can be in Liang Ge identical mappings branch
Output adjustment module is set in any one.
In another example the neural network referred in the corresponding embodiments of above-mentioned Fig. 3 to Fig. 5, also identical reflected there are multiple
The situation of branch's addition is penetrated, it can at least one of which identical mapping branch (such as f0、f1... or fc) increase setting BN-ReLU-
Conv layers, the output of the branch is thus adjusted, multiple identical mapping branches is avoided to be added out the problems such as corresponding variance of appearance is superimposed.
Image processing method according to embodiments of the present invention, by the feature pyramid sub-network of neural network, based on more
Kind different scale carries out feature extraction to the characteristic pattern of image to be detected, and will obtain other multiple characteristic patterns and be closed with characteristic pattern
And come the fisrt feature figure that obtains image to be detected, feature pyramid e-learning and the feature of extraction different scale are utilized, is protected
Accuracy and robustness that neural network carries out feature extraction are demonstrate,proved；On this basis, according to the fisrt feature figure of acquisition come into
Row critical point detection is effectively improved the accuracy of critical point detection.
Embodiment three
With reference to Figure 10, a kind of structure diagram of according to embodiments of the present invention three image processing apparatus is shown.
The image processing apparatus of the present embodiment, including：Acquisition module 1002, for obtaining the characteristic pattern of image to be detected；
Extraction module 1004 is based at least two kinds of different scales to characteristic pattern progress feature extraction for passing through neural network, obtains
Obtain at least two other characteristic patterns；Merging module 1006 for merging the characteristic pattern and other each described characteristic patterns, obtains institute
State the fisrt feature figure of image to be detected.
Optionally, it further includes：Detection module 1008, for according to the fisrt feature figure in described image to be detected
Target object carries out critical point detection.
Optionally, the detection module 1008 includes：Subdivision (not shown) is obtained, for according to the fisrt feature
Figure obtains the shot chart of an at least key point for the target object respectively；Determination unit (not shown), for according to each
The score of included pixel in the shot chart determines the position of the corresponding key point of the target object.
Optionally, the neural network includes at least one feature pyramid sub-network, the feature pyramid sub-network
At least one second branching networks in parallel with first branching networks including the first branching networks and respectively；It is described other
Characteristic pattern includes second feature figure or third feature figure；First branching networks are used for the original scale based on the characteristic pattern
Feature extraction is carried out to the characteristic pattern, obtains the second feature figure；Each second branching networks are used to be based respectively on not
Other scales for being same as the original scale carry out feature extraction to the characteristic pattern, obtain the third feature figure.
Optionally, first branching networks include the second convolutional layer, third convolutional layer and Volume Four lamination；Described second
Convolutional layer for reducing the characteristic pattern dimension；The third convolutional layer is for the original scale based on the characteristic pattern to drop
Characteristic pattern after low dimensional carries out process of convolution；The Volume Four lamination is used to be promoted the dimension of the characteristic pattern Jing Guo process of convolution
Degree, obtains the second feature figure.
Optionally, at least 1 second branching networks include the 5th convolutional layer, down-sampled layer, the 6th convolutional layer, on adopt
Sample layer and the 7th convolutional layer；5th convolutional layer for reducing the characteristic pattern dimension；The down-sampled layer is used for basis
It is down-sampled to the characteristic pattern progress after reducing dimension to set down-sampled ratio, wherein, the scale of the characteristic pattern after down-sampled
Less than the original scale of the characteristic pattern；6th convolutional layer is used to carry out at convolution by down-sampled characteristic pattern to described
Reason；The up-sampling layer is used to, according to setting up-sampling ratio, up-sample the characteristic pattern Jing Guo convolution, wherein, pass through
The scale of characteristic pattern after up-sampling is equal to the original scale of the characteristic pattern；7th convolutional layer is above adopted for being promoted to pass through
The dimension of characteristic pattern after sample obtains the third feature figure.
Optionally, second branching networks have multiple；The down-sampled ratio of setting of at least two second branching networks
Example it is different and/or, the down-sampled ratio of setting of at least two second branching networks is identical.
Optionally, second branching networks have multiple；The 6th convolution of at least two second branching networks
Layer shared parameter.
Optionally, second branching networks include the 5th convolutional layer, expansion convolutional layer and the 7th convolutional layer；Described 5th
Convolutional layer for reducing the characteristic pattern dimension；The expansion convolutional layer is used to carry out the characteristic pattern after reduction dimension
Expand process of convolution；7th convolutional layer obtains the third for promoting the dimension of the characteristic pattern after expanding convolution
Characteristic pattern.
Optionally, second branching networks have multiple；At least two second branching networks share described volume five
Lamination and/or the 7th convolutional layer；And/or at least two second branching networks have respective 5th convolution
Layer and/or the 7th convolutional layer.
Optionally, the feature pyramid sub-network further includes the first output merging layer；First output merges layer and uses
In respective defeated before the 7th convolutional layer at least two second branching networks for sharing the 7th convolutional layer
Go out to merge and export amalgamation result to shared the 7th convolutional layer.
Optionally, the neural network includes at least two feature pyramid sub-networks；Each feature pyramid subnet
Network, the fisrt feature figure for being exported using the previous feature pyramid sub-network being connect with current signature pyramid sub-network is defeated
Enter, and according to the fisrt feature figure of input, the fisrt feature figure based on different scale extraction current signature pyramid sub-network.
Optionally, the neural network be hourglass HOURGLASS neural networks, the hourglass HOURGLASS neural networks
Including an at least hourglass module include at least one feature pyramid sub-network.
Optionally, the initialization network parameter of an at least network layer for the neural network, from according to the initialization net
It is obtained, and the mean value of the initialization network parameter is zero in the network parameter distribution that the mean value and variance of network parameter determine.
Optionally, if existing in neural network including the situation that at least two identical mappings are added, needing what is be added
Output adjustment module is set in an at least identical mapping branch, and the output adjustment module is defeated for adjusting the identical mapping branch
The fisrt feature figure gone out.
The image processing apparatus of the present embodiment is used to implement corresponding image processing method in preceding method embodiment, and has
There is the advantageous effect of corresponding embodiment of the method, details are not described herein.
The present embodiment also provides a kind of computer readable storage medium, is stored thereon with computer program instructions, wherein, it should
The step of any image processing method provided in an embodiment of the present invention is realized when program instruction is executed by processor.
The present embodiment also provides a kind of computer program, including：An at least executable instruction, described at least one executable finger
The step of any image processing method provided in an embodiment of the present invention is used to implement when order is executed by processor.
Example IV
The embodiment of the present invention four provides a kind of electronic equipment, such as can be mobile terminal, personal computer (PC), put down
Plate computer, server etc..Below with reference to Figure 11, it illustrates suitable for being used for realizing the terminal device of the embodiment of the present invention or service
The structure diagram of the electronic equipment 1100 of device：As shown in figure 11, electronic equipment 1100 includes one or more processors, communication
Element etc., one or more of processors are for example：One or more central processing unit (CPU) 1101 and/or one or more
A image processor (GPU) 1113 etc., processor can be according to the executable instructions being stored in read-only memory (ROM) 1102
Or performed from the executable instruction that storage section 1108 is loaded into random access storage device (RAM) 1103 it is various appropriate
Action and processing.Communication device includes communication component 1112 and/or communication interface 1109.Wherein, communication component 1112 may include
But network interface card is not limited to, the network interface card may include but be not limited to IB (Infiniband) network interface card, and communication interface 1109 includes such as LAN
The communication interface of the network interface card of card, modem etc., communication interface 1109 perform logical via the network of such as internet
Letter processing.
Processor can communicate to perform executable finger with read-only memory 1102 and/or random access storage device 1103
It enables, is connected by communication bus 1104 with communication component 1112 and communicated through communication component 1112 with other target devices, so as to
The corresponding operation of image processing method any one of provided in an embodiment of the present invention is completed, for example, obtaining the feature of image to be detected
Figure；By neural network be based at least two kinds of different scales to the characteristic pattern carry out feature extraction, obtain at least two other
Characteristic pattern；Merge the characteristic pattern and other each described characteristic patterns, obtain the fisrt feature figure of described image to be detected.
In addition, in RAM 1103, it can also be stored with various programs and data needed for device operation.CPU1101 or
GPU1113, ROM1102 and RAM1103 are connected with each other by communication bus 1104.In the case where there is RAM1103,
ROM1102 is optional module.RAM1103 stores executable instruction or executable instruction is written into ROM1102 at runtime,
Executable instruction makes processor perform the corresponding operation of above-mentioned communication means.Input/output (I/O) interface 1105 is also connected to logical
Believe bus 1104.Communication component 1112 can be integrally disposed, may be set to be with multiple submodule (such as multiple IB nets
Card), and chained in communication bus.
I/O interfaces 1105 are connected to lower component：Importation 1106 including keyboard, mouse etc.；Including such as cathode
The output par, c 1107 of ray tube (CRT), liquid crystal display (LCD) etc. and loud speaker etc.；Storage section including hard disk etc.
1108；And the communication interface 1109 of the network interface card including LAN card, modem etc..The also root of driver 1110
According to needing to be connected to I/O interfaces 1105.Detachable media 1111, such as disk, CD, magneto-optic disk, semiconductor memory etc.,
It is mounted on driver 1110 as needed, in order to be mounted into storage part as needed from the computer program read thereon
Divide 1108.
Need what is illustrated, framework as shown in figure 11 is only a kind of optional realization method, can root during concrete practice
The component count amount and type of above-mentioned Figure 11 are selected, are deleted, increased or replaced according to actual needs；It is set in different function component
Put, can also be used it is separately positioned or integrally disposed and other implementations, such as GPU and CPU separate setting or can be by GPU collection
Into on CPU, communication device separates setting, can also be integrally disposed on CPU or GPU, etc..These interchangeable embodiment party
Formula each falls within protection scope of the present invention.
Particularly, according to embodiments of the present invention, it is soft to may be implemented as computer for the process above with reference to flow chart description
Part program.For example, the embodiment of the present invention includes a kind of computer program product, including being tangibly embodied in machine readable media
On computer program, computer program included for the program code of the method shown in execution flow chart, and program code can wrap
The corresponding instruction of corresponding execution method and step provided in an embodiment of the present invention is included, for example, obtaining the characteristic pattern of image to be detected；It is logical
It crosses neural network and is based at least two kinds of different scales to characteristic pattern progress feature extraction, obtain at least two other features
Figure；Merge the characteristic pattern and other each described characteristic patterns, obtain the fisrt feature figure of described image to be detected.In such reality
It applies in example, which can be downloaded and installed from network by communication device and/or from detachable media 1111
It is mounted.When the computer program is executed by processor, the above-mentioned function of being limited in the method for the embodiment of the present invention is performed.
It may be noted that according to the needs of implementation, all parts/step described in the embodiment of the present invention can be split as more
The part operation of two or more components/steps or components/steps can be also combined into new component/step by multi-part/step
Suddenly, to realize the purpose of the embodiment of the present invention.
It is above-mentioned to realize or be implemented as in hardware, firmware according to the method for the embodiment of the present invention to be storable in note
Software or computer code in recording medium (such as CD ROM, RAM, floppy disk, hard disk or magneto-optic disk) are implemented through net
The original storage that network is downloaded is in long-range recording medium or nonvolatile machine readable media and will be stored in local recording medium
In computer code, can be stored in using all-purpose computer, application specific processor or can compile so as to method described here
Such software processing in journey or the recording medium of specialized hardware (such as ASIC or FPGA).It is appreciated that computer, processing
Device, microprocessor controller or programmable hardware include can storing or receive software or computer code storage assembly (for example,
RAM, ROM, flash memory etc.), when the software or computer code are by computer, processor or hardware access and when performing, realize
Processing method described here.In addition, when all-purpose computer access is used to implement the code for the processing being shown here, code
It performs and is converted to all-purpose computer to perform the special purpose computer of processing being shown here.
Those of ordinary skill in the art may realize that each exemplary lists described with reference to the embodiments described herein
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 performed with hardware or software mode, specific application and design constraint depending on technical solution.Professional technician
Described function can be realized using distinct methods to each specific application, but this realization is it is not considered that exceed
The range of the embodiment of the present invention.
Embodiment of above is merely to illustrate the embodiment of the present invention, and is not the limitation to the embodiment of the present invention, related skill
The those of ordinary skill in art field in the case where not departing from the spirit and scope of the embodiment of the present invention, can also make various
Variation and modification, therefore all equivalent technical solutions also belong to the scope of the embodiment of the present invention, the patent of the embodiment of the present invention
Protection domain should be defined by the claims.
Claims (10)
1. a kind of image processing method, including：
Obtain the characteristic pattern of image to be detected；
By neural network be based at least two kinds of different scales to the characteristic pattern carry out feature extraction, obtain at least two other
Characteristic pattern；
Merge the characteristic pattern and other each described characteristic patterns, obtain the fisrt feature figure of described image to be detected.
2. it according to the method described in claim 1, wherein, further includes：
Critical point detection is carried out to the target object in described image to be detected according to the fisrt feature figure.
It is 3. described that the target object is closed according to the fisrt feature figure according to the method described in claim 2, wherein
Key point detects, including：
Obtain the shot chart of an at least key point for the target object respectively according to the fisrt feature figure；
According to the score of pixel included in each shot chart, the position of the corresponding key point of the target object is determined
It puts.
4. according to the method any in claims 1 to 3, wherein, the neural network includes at least one feature gold word
Tower sub-network, the feature pyramid sub-network include the first branching networks and in parallel with first branching networks respectively
At least one second branching networks；Other described characteristic patterns include second feature figure or third feature figure；
The original scale of first branching networks based on the characteristic pattern carries out feature extraction to the characteristic pattern, described in acquisition
Second feature figure；
Each second branching networks are based respectively on other scales different from the original scale to characteristic pattern progress spy
Sign extraction, obtains the third feature figure.
5. according to the method described in claim 4, wherein, first branching networks include the second convolutional layer, third convolutional layer
With Volume Four lamination；
Second convolutional layer reduces the dimension of the characteristic pattern；
The original scale of the third convolutional layer based on the characteristic pattern carries out process of convolution to the characteristic pattern after reducing dimension；
The Volume Four lamination promotes the dimension of the characteristic pattern Jing Guo process of convolution, obtains the second feature figure.
6. method according to claim 4 or 5, wherein, at least 1 second branching networks include the 5th convolutional layer, drop
Sample level, the 6th convolutional layer, up-sampling layer and the 7th convolutional layer；
5th convolutional layer reduces the dimension of the characteristic pattern；
The down-sampled layer is down-sampled to the characteristic pattern progress after reducing dimension according to down-sampled ratio is set, wherein, by drop
The scale of characteristic pattern after sampling is less than the original scale of the characteristic pattern；
6th convolutional layer carries out process of convolution to described by down-sampled characteristic pattern；
The up-sampling layer up-samples the characteristic pattern Jing Guo convolution according to up-sampling ratio is set, wherein, by above adopting
The scale of characteristic pattern after sample is equal to the original scale of the characteristic pattern；
7th convolutional layer promotes the dimension of the characteristic pattern after up-sampling, obtains the third feature figure.
7. a kind of image processing apparatus, including：
Acquisition module, for obtaining the characteristic pattern of image to be detected；
Extraction module is based at least two kinds of different scales to characteristic pattern progress feature extraction for passing through neural network, obtains
Obtain at least two other characteristic patterns；
Merging module, for merging the characteristic pattern and other each described characteristic patterns, obtain described image to be detected first is special
Sign figure.
8. a kind of computer readable storage medium, is stored thereon with computer program instructions, wherein, described program instruction is handled
The step of any one of claim 1 to 6 described image processing method is realized when device performs.
9. a kind of electronic equipment, including：Processor, memory, communication device and communication bus, the processor, the storage
Device and the communication device complete mutual communication by the communication bus；
For the memory for storing an at least executable instruction, the executable instruction makes the processor perform right such as will
Ask the corresponding operation of the image processing method described in any one of 1 to 6.
10. a kind of computer program, including：An at least executable instruction, an at least executable instruction are executed by processor
When be used to implement such as the corresponding operation of any one of claim 1 to 6 described image processing method.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710632941.0A CN108229497B (en) | 2017-07-28 | 2017-07-28 | Image processing method, image processing apparatus, storage medium, computer program, and electronic device |
| PCT/CN2018/097227 WO2019020075A1 (en) | 2017-07-28 | 2018-07-26 | Image processing method, device, storage medium, computer program, and electronic device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710632941.0A CN108229497B (en) | 2017-07-28 | 2017-07-28 | Image processing method, image processing apparatus, storage medium, computer program, and electronic device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108229497A true CN108229497A (en) | 2018-06-29 |
| CN108229497B CN108229497B (en) | 2021-01-05 |
Family
ID=62655195
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710632941.0A Active CN108229497B (en) | 2017-07-28 | 2017-07-28 | Image processing method, image processing apparatus, storage medium, computer program, and electronic device |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN108229497B (en) |
| WO (1) | WO2019020075A1 (en) |
Cited By (33)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108921225A (en) * | 2018-07-10 | 2018-11-30 | 深圳市商汤科技有限公司 | A kind of image processing method and device, computer equipment and storage medium |
| CN109117888A (en) * | 2018-08-20 | 2019-01-01 | 北京旷视科技有限公司 | Recongnition of objects method and its neural network generation method and device |
| WO2019020075A1 (en) * | 2017-07-28 | 2019-01-31 | 北京市商汤科技开发有限公司 | Image processing method, device, storage medium, computer program, and electronic device |
| CN109344840A (en) * | 2018-08-07 | 2019-02-15 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment, storage medium, program product |
| CN109359676A (en) * | 2018-10-08 | 2019-02-19 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating vehicle damage information |
| CN109360633A (en) * | 2018-09-04 | 2019-02-19 | 北京市商汤科技开发有限公司 | Medical imaging processing method and processing device, processing equipment and storage medium |
| CN109410218A (en) * | 2018-10-08 | 2019-03-01 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating vehicle damage information |
| CN109614876A (en) * | 2018-11-16 | 2019-04-12 | 北京市商汤科技开发有限公司 | Critical point detection method and device, electronic equipment and storage medium |
| CN109670516A (en) * | 2018-12-19 | 2019-04-23 | 广东工业大学 | A kind of image characteristic extracting method, device, equipment and readable storage medium storing program for executing |
| CN109784350A (en) * | 2018-12-29 | 2019-05-21 | 天津大学 | In conjunction with the dress ornament key independent positioning method of empty convolution and cascade pyramid network |
| CN109784194A (en) * | 2018-12-20 | 2019-05-21 | 上海图森未来人工智能科技有限公司 | Target detection network establishing method and training method, object detection method |
| CN109815770A (en) * | 2019-01-31 | 2019-05-28 | 北京旷视科技有限公司 | Two-dimentional code detection method, apparatus and system |
| CN109871890A (en) * | 2019-01-31 | 2019-06-11 | 北京字节跳动网络技术有限公司 | Image processing method and device |
| CN109902738A (en) * | 2019-02-25 | 2019-06-18 | 深圳市商汤科技有限公司 | Network module and distribution method and device, electronic equipment and storage medium |
| CN110046627A (en) * | 2018-10-16 | 2019-07-23 | 杭州依图医疗技术有限公司 | A kind of method and device of breast image identification |
| CN110163197A (en) * | 2018-08-24 | 2019-08-23 | 腾讯科技（深圳）有限公司 | Object detection method, device, computer readable storage medium and computer equipment |
| CN110390394A (en) * | 2019-07-19 | 2019-10-29 | 深圳市商汤科技有限公司 | Criticize processing method and processing device, electronic equipment and the storage medium of normalization data |
| CN110503063A (en) * | 2019-08-28 | 2019-11-26 | 东北大学秦皇岛分校 | Fall detection method based on hourglass convolution autocoding neural network |
| CN110619604A (en) * | 2019-09-17 | 2019-12-27 | 中国气象局公共气象服务中心(国家预警信息发布中心) | Three-dimensional downscaling method and device, electronic equipment and readable storage medium |
| WO2020020146A1 (en) * | 2018-07-25 | 2020-01-30 | 深圳市商汤科技有限公司 | Method and apparatus for processing laser radar sparse depth map, device, and medium |
| CN110956190A (en) * | 2018-09-27 | 2020-04-03 | 深圳云天励飞技术有限公司 | Image recognition method and device, computer device and computer readable storage medium |
| CN111091593A (en) * | 2018-10-24 | 2020-05-01 | 深圳云天励飞技术有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| WO2020093837A1 (en) * | 2018-11-07 | 2020-05-14 | 北京达佳互联信息技术有限公司 | Method for detecting key points in human skeleton, apparatus, electronic device, and storage medium |
| CN111191486A (en) * | 2018-11-14 | 2020-05-22 | 杭州海康威视数字技术股份有限公司 | Drowning behavior recognition method, monitoring camera and monitoring system |
| CN111291660A (en) * | 2020-01-21 | 2020-06-16 | 天津大学 | Anchor-free traffic sign identification method based on void convolution |
| CN111556337A (en) * | 2020-05-15 | 2020-08-18 | 腾讯科技（深圳）有限公司 | Media content implantation method, model training method and related device |
| CN111582206A (en) * | 2020-05-13 | 2020-08-25 | 北京字节跳动网络技术有限公司 | Method and device for generating organism posture key point information |
| CN112232361A (en) * | 2020-10-13 | 2021-01-15 | 国网电子商务有限公司 | Image processing method and device, electronic equipment and computer readable storage medium |
| CN112784629A (en) * | 2019-11-06 | 2021-05-11 | 株式会社理光 | Image processing method, apparatus and computer-readable storage medium |
| US11048935B2 (en) * | 2019-01-28 | 2021-06-29 | Adobe Inc. | Generating shift-invariant neural network outputs |
| CN113076914A (en) * | 2021-04-16 | 2021-07-06 | 咪咕文化科技有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN113420641A (en) * | 2021-06-21 | 2021-09-21 | 梅卡曼德（北京）机器人科技有限公司 | Image data processing method, image data processing device, electronic equipment and storage medium |
| WO2022022695A1 (en) * | 2020-07-31 | 2022-02-03 | 华为技术有限公司 | Image recognition method and apparatus |
Families Citing this family (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111860557A (en) * | 2019-04-30 | 2020-10-30 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and computer storage medium |
| CN110188776A (en) * | 2019-05-30 | 2019-08-30 | 京东方科技集团股份有限公司 | Image processing method and device, the training method of neural network, storage medium |
| CN112116060B (en) * | 2019-06-21 | 2023-07-25 | 杭州海康威视数字技术股份有限公司 | Network configuration implementation method and device |
| CN110348537B (en) * | 2019-07-18 | 2022-11-29 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN110807757B (en) * | 2019-08-14 | 2023-07-25 | 腾讯科技（深圳）有限公司 | Image quality evaluation method and device based on artificial intelligence and computer equipment |
| CN110472732B (en) * | 2019-08-19 | 2023-02-21 | 杭州凝眸智能科技有限公司 | Image feature extraction system based on optimized feature extraction device |
| CN111047630B (en) * | 2019-11-13 | 2023-06-13 | 芯启源(上海)半导体科技有限公司 | Neural network and target detection and depth prediction method based on neural network |
| CN112883981A (en) * | 2019-11-29 | 2021-06-01 | 阿里巴巴集团控股有限公司 | Image processing method, device and storage medium |
| CN111190952B (en) * | 2019-12-23 | 2023-10-03 | 中电海康集团有限公司 | Method for extracting and persistence of multi-scale features of city portrait based on image pyramid |
| CN111414990B (en) * | 2020-02-20 | 2024-03-19 | 北京迈格威科技有限公司 | Convolutional neural network processing method and device, electronic equipment and storage medium |
| CN111523377A (en) * | 2020-03-10 | 2020-08-11 | 浙江工业大学 | Multi-task human body posture estimation and behavior recognition method |
| CN111476740B (en) * | 2020-04-28 | 2023-10-31 | 北京大米未来科技有限公司 | Image processing method, device, storage medium and electronic equipment |
| CN111783934A (en) * | 2020-05-15 | 2020-10-16 | 北京迈格威科技有限公司 | Convolutional neural network construction method, device, equipment and medium |
| CN111914997B (en) * | 2020-06-30 | 2024-04-02 | 华为技术有限公司 | Method for training neural network, image processing method and device |
| CN111932530B (en) * | 2020-09-18 | 2024-02-23 | 北京百度网讯科技有限公司 | Three-dimensional object detection method, device, equipment and readable storage medium |
| CN112149558A (en) * | 2020-09-22 | 2020-12-29 | 驭势科技(南京)有限公司 | Image processing method, network and electronic equipment for key point detection |
| CN112184687B (en) * | 2020-10-10 | 2023-09-26 | 南京信息工程大学 | Road crack detection method based on capsule feature pyramid and storage medium |
| CN112528900B (en) * | 2020-12-17 | 2022-09-16 | 南开大学 | Image salient object detection method and system based on extreme down-sampling |
| CN112836804A (en) * | 2021-02-08 | 2021-05-25 | 北京迈格威科技有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| CN113344862B (en) * | 2021-05-20 | 2024-04-12 | 北京百度网讯科技有限公司 | Defect detection method, device, electronic equipment and storage medium |
| CN113591573A (en) * | 2021-06-28 | 2021-11-02 | 北京百度网讯科技有限公司 | Training and target detection method and device for multi-task learning deep network model |
| CN113837104B (en) * | 2021-09-26 | 2024-03-15 | 大连智慧渔业科技有限公司 | Underwater fish target detection method and device based on convolutional neural network and storage medium |
| CN116091486B (en) * | 2023-03-01 | 2024-02-06 | 合肥联宝信息技术有限公司 | Surface defect detection method, surface defect detection device, electronic equipment and storage medium |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016054802A1 (en) * | 2014-10-10 | 2016-04-14 | Beijing Kuangshi Technology Co., Ltd. | Hierarchical interlinked multi-scale convolutional network for image parsing |
| CN105956626A (en) * | 2016-05-12 | 2016-09-21 | 成都新舟锐视科技有限公司 | Deep learning based vehicle license plate position insensitive vehicle license plate recognition method |
| CN106529447A (en) * | 2016-11-03 | 2017-03-22 | 河北工业大学 | Small-sample face recognition method |
| CN106651877A (en) * | 2016-12-20 | 2017-05-10 | 北京旷视科技有限公司 | Example segmenting method and device |
| CN106650786A (en) * | 2016-11-14 | 2017-05-10 | 沈阳工业大学 | Image recognition method based on multi-column convolutional neural network fuzzy evaluation |
| CN106650913A (en) * | 2016-12-31 | 2017-05-10 | 中国科学技术大学 | Deep convolution neural network-based traffic flow density estimation method |
| CN106909905A (en) * | 2017-03-02 | 2017-06-30 | 中科视拓（北京）科技有限公司 | A kind of multi-modal face identification method based on deep learning |
| US20170200067A1 (en) * | 2016-01-08 | 2017-07-13 | Siemens Healthcare Gmbh | Deep Image-to-Image Network Learning for Medical Image Analysis |
| CN106951867A (en) * | 2017-03-22 | 2017-07-14 | 成都擎天树科技有限公司 | Face identification method, device, system and equipment based on convolutional neural networks |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9665802B2 (en) * | 2014-11-13 | 2017-05-30 | Nec Corporation | Object-centric fine-grained image classification |
| CN106611169B (en) * | 2016-12-31 | 2018-10-23 | 中国科学技术大学 | A kind of dangerous driving behavior real-time detection method based on deep learning |
| CN108229497B (en) * | 2017-07-28 | 2021-01-05 | 北京市商汤科技开发有限公司 | Image processing method, image processing apparatus, storage medium, computer program, and electronic device |
-
2017
- 2017-07-28 CN CN201710632941.0A patent/CN108229497B/en active Active
-
2018
- 2018-07-26 WO PCT/CN2018/097227 patent/WO2019020075A1/en active Application Filing
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016054802A1 (en) * | 2014-10-10 | 2016-04-14 | Beijing Kuangshi Technology Co., Ltd. | Hierarchical interlinked multi-scale convolutional network for image parsing |
| US20170200067A1 (en) * | 2016-01-08 | 2017-07-13 | Siemens Healthcare Gmbh | Deep Image-to-Image Network Learning for Medical Image Analysis |
| CN105956626A (en) * | 2016-05-12 | 2016-09-21 | 成都新舟锐视科技有限公司 | Deep learning based vehicle license plate position insensitive vehicle license plate recognition method |
| CN106529447A (en) * | 2016-11-03 | 2017-03-22 | 河北工业大学 | Small-sample face recognition method |
| CN106650786A (en) * | 2016-11-14 | 2017-05-10 | 沈阳工业大学 | Image recognition method based on multi-column convolutional neural network fuzzy evaluation |
| CN106651877A (en) * | 2016-12-20 | 2017-05-10 | 北京旷视科技有限公司 | Example segmenting method and device |
| CN106650913A (en) * | 2016-12-31 | 2017-05-10 | 中国科学技术大学 | Deep convolution neural network-based traffic flow density estimation method |
| CN106909905A (en) * | 2017-03-02 | 2017-06-30 | 中科视拓（北京）科技有限公司 | A kind of multi-modal face identification method based on deep learning |
| CN106951867A (en) * | 2017-03-22 | 2017-07-14 | 成都擎天树科技有限公司 | Face identification method, device, system and equipment based on convolutional neural networks |
Non-Patent Citations (6)
| Title |
|---|
| FISHER YU等: "Multi Scale Context Aggregation by Dilated Convolutions", 《ARXIV:1511.07122V3》 * |
| HYEONSEOB NAM等: "Learning Multi-domain Convolutional Neural Networks for Visual Tracking", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
| XIAO CHU等: "Multi-context Attention for Human Pose Estimation", 《ARXIV：1702.07432V1》 * |
| 张文达等: "基于多尺度分块卷积神经网络的图像目标识别算法", 《计算机应用》 * |
| 谭贞刚: "基于多分辨率卷积网络的交通标志检测与识别", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
| 费春: "基于智能优化和视觉显著性的图像融合研究", 《中国博士学位论文全文数据库 信息科技辑（月刊）》 * |
Cited By (61)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019020075A1 (en) * | 2017-07-28 | 2019-01-31 | 北京市商汤科技开发有限公司 | Image processing method, device, storage medium, computer program, and electronic device |
| CN108921225A (en) * | 2018-07-10 | 2018-11-30 | 深圳市商汤科技有限公司 | A kind of image processing method and device, computer equipment and storage medium |
| CN108921225B (en) * | 2018-07-10 | 2022-06-24 | 深圳市商汤科技有限公司 | Image processing method and device, computer equipment and storage medium |
| WO2020020146A1 (en) * | 2018-07-25 | 2020-01-30 | 深圳市商汤科技有限公司 | Method and apparatus for processing laser radar sparse depth map, device, and medium |
| CN109344840A (en) * | 2018-08-07 | 2019-02-15 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment, storage medium, program product |
| CN109344840B (en) * | 2018-08-07 | 2022-04-01 | 深圳市商汤科技有限公司 | Image processing method and apparatus, electronic device, storage medium, and program product |
| CN109117888A (en) * | 2018-08-20 | 2019-01-01 | 北京旷视科技有限公司 | Recongnition of objects method and its neural network generation method and device |
| CN110163197A (en) * | 2018-08-24 | 2019-08-23 | 腾讯科技（深圳）有限公司 | Object detection method, device, computer readable storage medium and computer equipment |
| US11710293B2 (en) | 2018-08-24 | 2023-07-25 | Tencent Technology (Shenzhen) Company Limited | Target detection method and apparatus, computer-readable storage medium, and computer device |
| CN109360633A (en) * | 2018-09-04 | 2019-02-19 | 北京市商汤科技开发有限公司 | Medical imaging processing method and processing device, processing equipment and storage medium |
| CN109360633B (en) * | 2018-09-04 | 2022-08-30 | 北京市商汤科技开发有限公司 | Medical image processing method and device, processing equipment and storage medium |
| CN110956190A (en) * | 2018-09-27 | 2020-04-03 | 深圳云天励飞技术有限公司 | Image recognition method and device, computer device and computer readable storage medium |
| CN109359676A (en) * | 2018-10-08 | 2019-02-19 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating vehicle damage information |
| US11244435B2 (en) | 2018-10-08 | 2022-02-08 | Baidu Online Network Technology (Beijing) Co., Ltd. | Method and apparatus for generating vehicle damage information |
| US11392792B2 (en) | 2018-10-08 | 2022-07-19 | Baidu Online Network Technology (Beijing) Co., Ltd. | Method and apparatus for generating vehicle damage information |
| CN109410218A (en) * | 2018-10-08 | 2019-03-01 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating vehicle damage information |
| CN110046627A (en) * | 2018-10-16 | 2019-07-23 | 杭州依图医疗技术有限公司 | A kind of method and device of breast image identification |
| CN110046627B (en) * | 2018-10-16 | 2021-09-10 | 杭州依图医疗技术有限公司 | Method and device for identifying mammary gland image |
| CN111091593B (en) * | 2018-10-24 | 2024-03-22 | 深圳云天励飞技术有限公司 | Image processing method, device, electronic equipment and storage medium |
| CN111091593A (en) * | 2018-10-24 | 2020-05-01 | 深圳云天励飞技术有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| WO2020093837A1 (en) * | 2018-11-07 | 2020-05-14 | 北京达佳互联信息技术有限公司 | Method for detecting key points in human skeleton, apparatus, electronic device, and storage medium |
| US11373426B2 (en) | 2018-11-07 | 2022-06-28 | Beijing Dajia Internet Information Technology Co., Ltd. | Method for detecting key points in skeleton, apparatus, electronic device and storage medium |
| CN111191486A (en) * | 2018-11-14 | 2020-05-22 | 杭州海康威视数字技术股份有限公司 | Drowning behavior recognition method, monitoring camera and monitoring system |
| CN111191486B (en) * | 2018-11-14 | 2023-09-05 | 杭州海康威视数字技术股份有限公司 | Drowning behavior recognition method, monitoring camera and monitoring system |
| CN109614876B (en) * | 2018-11-16 | 2021-07-27 | 北京市商汤科技开发有限公司 | Key point detection method and device, electronic equipment and storage medium |
| CN113591754B (en) * | 2018-11-16 | 2022-08-02 | 北京市商汤科技开发有限公司 | Key point detection method and device, electronic equipment and storage medium |
| CN109614876A (en) * | 2018-11-16 | 2019-04-12 | 北京市商汤科技开发有限公司 | Critical point detection method and device, electronic equipment and storage medium |
| CN113591750A (en) * | 2018-11-16 | 2021-11-02 | 北京市商汤科技开发有限公司 | Key point detection method and device, electronic equipment and storage medium |
| CN113591754A (en) * | 2018-11-16 | 2021-11-02 | 北京市商汤科技开发有限公司 | Key point detection method and device, electronic equipment and storage medium |
| CN113569796A (en) * | 2018-11-16 | 2021-10-29 | 北京市商汤科技开发有限公司 | Key point detection method and device, electronic equipment and storage medium |
| CN109670516A (en) * | 2018-12-19 | 2019-04-23 | 广东工业大学 | A kind of image characteristic extracting method, device, equipment and readable storage medium storing program for executing |
| CN109784194A (en) * | 2018-12-20 | 2019-05-21 | 上海图森未来人工智能科技有限公司 | Target detection network establishing method and training method, object detection method |
| CN109784350A (en) * | 2018-12-29 | 2019-05-21 | 天津大学 | In conjunction with the dress ornament key independent positioning method of empty convolution and cascade pyramid network |
| US11562166B2 (en) | 2019-01-28 | 2023-01-24 | Adobe Inc. | Generating shift-invariant neural network feature maps and outputs |
| US11048935B2 (en) * | 2019-01-28 | 2021-06-29 | Adobe Inc. | Generating shift-invariant neural network outputs |
| CN109815770A (en) * | 2019-01-31 | 2019-05-28 | 北京旷视科技有限公司 | Two-dimentional code detection method, apparatus and system |
| CN109871890A (en) * | 2019-01-31 | 2019-06-11 | 北京字节跳动网络技术有限公司 | Image processing method and device |
| US11443438B2 (en) | 2019-02-25 | 2022-09-13 | Shenzhen Sensetime Technology Co., Ltd. | Network module and distribution method and apparatus, electronic device, and storage medium |
| JP2021517282A (en) * | 2019-02-25 | 2021-07-15 | シェンチェン センスタイム テクノロジー カンパニー リミテッドＳｈｅｎｚｈｅｎ Ｓｅｎｓｅｔｉｍｅ Ｔｅｃｈｎｏｌｏｇｙ Ｃｏ．，Ｌｔｄ | Network modules, allocation methods and devices, electronic devices and storage media |
| TWI766228B (en) * | 2019-02-25 | 2022-06-01 | 大陸商深圳市商湯科技有限公司 | Network module, distribution method and apparatus, electronic device, and computer-readable storage medium |
| JP7096888B2 (en) | 2019-02-25 | 2022-07-06 | シェンチェン センスタイム テクノロジー カンパニー リミテッド | Network modules, allocation methods and devices, electronic devices and storage media |
| CN109902738A (en) * | 2019-02-25 | 2019-06-18 | 深圳市商汤科技有限公司 | Network module and distribution method and device, electronic equipment and storage medium |
| CN109902738B (en) * | 2019-02-25 | 2021-07-20 | 深圳市商汤科技有限公司 | Network module, distribution method and device, electronic equipment and storage medium |
| CN110390394A (en) * | 2019-07-19 | 2019-10-29 | 深圳市商汤科技有限公司 | Criticize processing method and processing device, electronic equipment and the storage medium of normalization data |
| CN110390394B (en) * | 2019-07-19 | 2021-11-05 | 深圳市商汤科技有限公司 | Batch normalization data processing method and device, electronic equipment and storage medium |
| CN110503063B (en) * | 2019-08-28 | 2021-12-17 | 东北大学秦皇岛分校 | Falling detection method based on hourglass convolution automatic coding neural network |
| CN110503063A (en) * | 2019-08-28 | 2019-11-26 | 东北大学秦皇岛分校 | Fall detection method based on hourglass convolution autocoding neural network |
| CN110619604A (en) * | 2019-09-17 | 2019-12-27 | 中国气象局公共气象服务中心(国家预警信息发布中心) | Three-dimensional downscaling method and device, electronic equipment and readable storage medium |
| CN112784629A (en) * | 2019-11-06 | 2021-05-11 | 株式会社理光 | Image processing method, apparatus and computer-readable storage medium |
| CN111291660B (en) * | 2020-01-21 | 2022-08-12 | 天津大学 | Anchor-free traffic sign identification method based on void convolution |
| CN111291660A (en) * | 2020-01-21 | 2020-06-16 | 天津大学 | Anchor-free traffic sign identification method based on void convolution |
| CN111582206A (en) * | 2020-05-13 | 2020-08-25 | 北京字节跳动网络技术有限公司 | Method and device for generating organism posture key point information |
| CN111582206B (en) * | 2020-05-13 | 2023-08-22 | 抖音视界有限公司 | Method and device for generating organism posture key point information |
| CN111556337A (en) * | 2020-05-15 | 2020-08-18 | 腾讯科技（深圳）有限公司 | Media content implantation method, model training method and related device |
| CN111556337B (en) * | 2020-05-15 | 2021-09-21 | 腾讯科技（深圳）有限公司 | Media content implantation method, model training method and related device |
| WO2022022695A1 (en) * | 2020-07-31 | 2022-02-03 | 华为技术有限公司 | Image recognition method and apparatus |
| CN112232361A (en) * | 2020-10-13 | 2021-01-15 | 国网电子商务有限公司 | Image processing method and device, electronic equipment and computer readable storage medium |
| CN112232361B (en) * | 2020-10-13 | 2021-09-21 | 国网电子商务有限公司 | Image processing method and device, electronic equipment and computer readable storage medium |
| CN113076914A (en) * | 2021-04-16 | 2021-07-06 | 咪咕文化科技有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN113076914B (en) * | 2021-04-16 | 2024-04-12 | 咪咕文化科技有限公司 | Image processing method, device, electronic equipment and storage medium |
| CN113420641A (en) * | 2021-06-21 | 2021-09-21 | 梅卡曼德（北京）机器人科技有限公司 | Image data processing method, image data processing device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108229497B (en) | 2021-01-05 |
| WO2019020075A1 (en) | 2019-01-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108229497A (en) | Image processing method, device, storage medium, computer program and electronic equipment | |
| Wu et al. | Fast end-to-end trainable guided filter | |
| CN109191382B (en) | Image processing method, device, electronic equipment and computer readable storage medium | |
| WO2018153322A1 (en) | Key point detection method, neural network training method, apparatus and electronic device | |
| CN105320965B (en) | Sky based on depth convolutional neural networks composes united hyperspectral image classification method | |
| CN107798381B (en) | Image identification method based on convolutional neural network | |
| CN110378381A (en) | Object detecting method, device and computer storage medium | |
| CN108475415A (en) | Method and system for image procossing | |
| CN110222700A (en) | SAR image recognition methods and device based on Analysis On Multi-scale Features and width study | |
| CN109166130A (en) | A kind of image processing method and image processing apparatus | |
| CN108717569A (en) | It is a kind of to expand full convolutional neural networks and its construction method | |
| CN111860398B (en) | Remote sensing image target detection method and system and terminal equipment | |
| CN108229591A (en) | Neural network adaptive training method and apparatus, equipment, program and storage medium | |
| CN108921789A (en) | Super-resolution image reconstruction method based on recurrence residual error network | |
| CN113657397B (en) | Training method for circularly generating network model, method and device for establishing word stock | |
| CN107220971A (en) | A kind of Lung neoplasm feature extracting method based on convolutional neural networks and PCA | |
| CN110232318A (en) | Acupuncture point recognition methods, device, electronic equipment and storage medium | |
| CN113228061A (en) | Electronic device and control method thereof | |
| Shu et al. | LVC-Net: Medical image segmentation with noisy label based on local visual cues | |
| CN108876716A (en) | Super resolution ratio reconstruction method and device | |
| CN111507184B (en) | Human body posture detection method based on parallel cavity convolution and body structure constraint | |
| CN108154153A (en) | Scene analysis method and system, electronic equipment | |
| CN112613541A (en) | Target detection method and device, storage medium and electronic equipment | |
| CN108229680A (en) | Nerve network system, remote sensing images recognition methods, device, equipment and medium | |
| CN117495876B (en) | Coronary artery image segmentation method and system based on deep learning |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |