CN110222829A - Feature extracting method, device, equipment and medium based on convolutional neural networks - Google Patents
Feature extracting method, device, equipment and medium based on convolutional neural networks Download PDFInfo
- Publication number
- CN110222829A CN110222829A CN201910507522.3A CN201910507522A CN110222829A CN 110222829 A CN110222829 A CN 110222829A CN 201910507522 A CN201910507522 A CN 201910507522A CN 110222829 A CN110222829 A CN 110222829A
- Authority
- CN
- China
- Prior art keywords
- image processing
- processing stage
- output
- result
- input
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—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]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present disclosure provides a kind of feature extracting method based on convolutional neural networks, device, equipment and medium.This method comprises: target image is input to scheduled convolutional neural networks, convolutional neural networks include S image processing stage, each image processing stage includes multiple convolutional layers, since first image processing stage, execute step A: the output result to the multiple convolutional layers for including in present image processing stage, respective handling is carried out according to scheduled convolution interlayer information interchange processing mode, obtain the output result of present image processing stage and the input as next image processing stage, and using next image processing stage as present image processing stage, circulation step A, until obtaining the output result of the S image processing stage, feature extraction result as target image.The embodiment of the present disclosure improves the accuracy of present image processing stage output result, and then improves the accuracy of convolutional neural networks processing result.
Description
Technical field
This disclosure relates to technical field of image processing, specifically, the present invention relates to a kind of based on convolutional neural networks
Feature extracting method, device, equipment and medium.
Background technique
Image procossing is to be analyzed using computer image, to reach the technology of required result.Usually utilize work
The equipment such as industry camera, video camera, scanner obtain a big two-dimensional array by shooting, and the element of the array is known as pixel,
Its value is known as gray value.
Image procossing can be carried out by neural network at present, it is generally the case that carry out image procossing using neural network
When information transmission direction be one direction transmission, input information of the output result of a convolutional layer as next convolutional layer,
It is carry out information interchange unidirectional between two convolutional layers.It, can shadow if the accuracy of the output information of previous convolutional layer is lower
Ring the accuracy of the output information of subsequent all convolutional layers.Therefore, existing image procossing mode, the accuracy of processing result compared with
It is low.
Summary of the invention
The purpose of the disclosure is intended at least can solve above-mentioned one of technological deficiency, especially because the unidirectional letter of convolutional layer
Breath exchange way leads to the lower technological deficiency of processing result accuracy.
In a first aspect, a kind of feature extracting method based on convolutional neural networks is provided, this method comprises:
Target image is input to scheduled convolutional neural networks, convolutional neural networks include S image processing stage, often
One image processing stage includes multiple convolutional layers, and S is the positive integer greater than 1；
Since present image processing stage is first image processing stage, step A is executed: rank is handled to present image
The output for the multiple convolutional layers for including in section is as a result, carry out corresponding position according to scheduled convolution interlayer information interchange processing mode
Reason, obtains the output result of present image processing stage and the input as next image processing stage, and by next figure
As processing stage is as present image processing stage；
Circulation step A, until obtaining the exporting as a result, feature extraction as target image of the S image processing stage
As a result.
Output knot in a possible implementation, to the multiple convolutional layers for including in present image processing stage
Fruit carries out respective handling according to scheduled convolution interlayer information interchange processing mode, obtains the output of present image processing stage
And the input as next image processing stage as a result, comprising:
1st to T output result of the corresponding output of the 1st to T convolutional layer in present image processing stage is input to
In default non-local network's block module, and the first meter is carried out to the 1st to T output result in default non-local network block module
Calculation processing obtains intermediate output as a result, T is the number of the convolutional layer in present image processing stage, and T is positive integer；
The second calculation processing is carried out according to the T output result and intermediate output result, obtains present image processing stage
Output result and input as next image processing stage.
In a possible implementation, in present image processing stage, the corresponding output of any one convolutional layer is determined
Either one or two of output result mode, including it is any one of following:
By the output of the upper image processing stage of present image processing stage as a result, being input to t-th of convolutional layer, obtain
The output of t-th of convolutional layer output is as a result, t=1；
In present image processing stage, the output result that the t-1 convolutional layer exports is input to t-th of convolutional layer,
The output of t-th of convolutional layer output is obtained as a result, 2≤t≤T, T are positive integer.
In a possible implementation, by the corresponding output of the 1st to T convolutional layer in present image processing stage
1st to T output result is input in default non-local network's block module, and to the 1st in default non-local network block module
The first calculation processing is carried out to T output result, obtains intermediate output result, comprising:
The 1st to T convolutional layer in present image processing stage is corresponded to the 1st of output using default connection array function
It links together, and is input in default non-local network's block module to T output result；
1st to T output result carries out Gaussian Computation processing in default non-local network block module, obtains intermediate output
As a result.
In a possible implementation, any image processing stage and its including any convolutional layer output result
It is characteristic image, characteristic image is indicated by formula F eature maps=W*H*C,
Feature maps indicates that characteristic image, W indicate that the width of characteristic image, H indicate that the height of characteristic image, C indicate
The dimension of characteristic image.
Second aspect, provides a kind of feature deriving means based on convolutional neural networks, which includes:
Input module, for target image to be input to scheduled convolutional neural networks, convolutional neural networks include S figure
As processing stage, each image processing stage includes multiple convolutional layers, and S is the positive integer greater than 1；
Processing module, for executing step A: right since present image processing stage is first image processing stage
The output for the multiple convolutional layers for including in present image processing stage is as a result, according to scheduled convolution interlayer information interchange processing side
Formula carries out respective handling, obtains the output result of present image processing stage and the input as next image processing stage,
And using next image processing stage as present image processing stage；
Loop module is used for circulation step A, until obtaining the output of the S image processing stage as a result, as target figure
The feature extraction result of picture.
In a possible implementation, processing module, comprising:
First processing module, for since present image processing stage is first image processing stage, execution will to work as
1st to T output result of the corresponding output of the 1st to T convolutional layer in preceding image processing stage, which is input to, presets non local net
In network block module, and the first calculation processing is carried out to the 1st to T output result in default non-local network block module, obtained
Centre output is as a result, T is the number of the convolutional layer in present image processing stage, and T is positive integer；
Second processing module is obtained for carrying out the second calculation processing according to the T output result and intermediate output result
The output result of present image processing stage and input as next image processing stage, and by next image procossing rank
Duan Zuowei present image processing stage.
In a possible implementation, first processing module, comprising:
Input module is connected, for executing utilization since present image processing stage is first image processing stage
Default connection array function ties the 1st to T output of the corresponding output of the 1st to T convolutional layer in present image processing stage
Fruit links together, and is input in default non-local network's block module；
Calculation processing module carries out Gaussian Computation in default non-local network block module for the 1st to T output result
Processing obtains intermediate output result.
The third aspect provides a kind of electronic equipment, which includes:
One or more processors；
Memory；
One or more application program, wherein one or more application programs be stored in memory and be configured as by
One or more processors execute, and one or more programs are configured to: executing feature extracting method described in any of the above embodiments.
For example, the third aspect of the disclosure, provides a kind of calculating equipment, comprising: processor, memory, communication interface
And communication bus, processor, memory and communication interface complete mutual communication by communication bus；
Memory makes processor execute the first party such as the disclosure for storing an at least executable instruction, executable instruction
The corresponding operation of feature extracting method shown in face.
Fourth aspect provides a kind of computer readable storage medium, is stored thereon with computer program, which is located
Reason device realizes feature extracting method described in any of the above embodiments when executing.
For example, the fourth aspect of the embodiment of the present disclosure, provides a kind of computer readable storage medium, it is computer-readable to deposit
It is stored with computer program on storage media, feature extraction shown in disclosure first aspect is realized when which is executed by processor
Method.
The technical solution that the embodiment of the present disclosure provides has the benefit that
By the output to the multiple convolutional layers for including in present image processing stage as a result, convolution inter-layer information is taken to hand over
Stream process mode carries out respective handling, so that information interchange mode is mutually carried out between multiple convolutional layers, rather than two convolutional layers
Between carry out unidirectional information interchange, improve present image processing stage output result accuracy, and then improve convolution mind
Accuracy through network processes result.
Detailed description of the invention
It, below will be to institute in embodiment of the present disclosure description in order to illustrate more clearly of the technical solution in the embodiment of the present disclosure
Attached drawing to be used is needed to be briefly described.
Fig. 1 is a kind of process signal for feature extracting method based on convolutional neural networks that the embodiment of the present disclosure provides
Figure；
Fig. 2 is a kind of structural representation for feature deriving means based on convolutional neural networks that the embodiment of the present disclosure provides
Figure；
Fig. 3 is a kind of structure of the electronic equipment for feature extraction based on convolutional neural networks that the embodiment of the present disclosure provides
Schematic diagram.
Specific embodiment
Embodiment of the disclosure is described below in detail, examples of the embodiments are shown in the accompanying drawings, wherein from beginning to end
Same or similar label indicates same or similar element or element with the same or similar functions.Below with reference to attached
The embodiment of figure description is exemplary, and is only used for explaining the disclosure, and is not construed as limiting the claims.
Those skilled in the art of the present technique are appreciated that unless expressly stated, singular " one " used herein, " one
It is a ", " described " and "the" may also comprise plural form.It is to be further understood that being arranged used in the specification of the disclosure
Diction " comprising " refer to that there are the feature, integer, step, operation, element and/or component, but it is not excluded that in the presence of or addition
Other one or more features, integer, step, operation, element, component and/or their group.It should be understood that when we claim member
Part is " connected " or when " coupled " to another element, it can be directly connected or coupled to other elements, or there may also be
Intermediary element.In addition, " connection " used herein or " coupling " may include being wirelessly connected or wirelessly coupling.It is used herein to arrange
Diction "and/or" includes one or more associated wholes for listing item or any cell and all combinations.
To keep the purposes, technical schemes and advantages of the disclosure clearer, below in conjunction with attached drawing to disclosure embodiment party
Formula is described in further detail.
First pair this disclosure relates to several nouns be introduced and explain: convolutional neural networks are a kind of comprising convolution meter
The feedforward neural network for calculating and having depth structure, is one of representative algorithm of deep learning.Convolutional neural networks include input
Layer, hidden layer and output layer.Input layer is for handling Multidimensional numerical, and generally, the input layer of one-dimensional convolutional neural networks receives
One-dimensional perhaps two-dimensional array one-dimensional data is usually time or spectral sample, and two-dimensional array may include multiple channels；Two dimension
The input layer of convolutional neural networks receives two dimension or three-dimensional array, and the input layer of Three dimensional convolution neural network receives thinking number
Group.Hidden layer includes convolutional layer, pond layer and 3 class of full articulamentum is common constructs and some complexity are constructed.Wherein, convolutional layer is used for
Feature extraction is carried out to input data, internal includes multiple convolution kernels, forms the corresponding power of each element of convolution kernel
Weight coefficient and a departure；Pond layer is to carry out feature selecting and letter to output result after convolutional layer carries out feature extraction
Breath filtering；Full articulamentum usually builds the decline in convolutional neural networks hidden layer, and only transmits to other full articulamentums
Signal.The upstream of output layer is usually full articulamentum, and for image classification problem, output layer is referred to using logical function or normalization
Number function output category label, in object identification problem, output layer may be designed as exporting the centre coordinate of object, size and divide
Class, in image, semantic segmentation, output layer directly exports the classification results of each image.
When carrying out image procossing by neural network at present, it is generally the case that when carrying out image procossing using neural network
The transmission direction of information is one direction transmission, i.e., the output information of each convolutional layer of neural network is as the defeated of next convolutional layer
Enter information, e.g., believe as the input of second convolutional layer by totally three convolutional layers, the output information of first convolutional layer for neural network
Breath, input information of the output information of second convolutional layer as third convolutional layer, the output information of third convolutional layer are
Final result.The image procossing mode is carry out information interchange unidirectional between two neighboring convolutional layer, if previous convolutional layer
Output information accuracy it is lower, will affect the accuracy of the output information of subsequent all convolutional layers.Therefore, existing image
The accuracy of processing mode, processing result is lower.
The feature extracting method based on convolutional neural networks, device, equipment and the medium that the disclosure provides, it is intended to solve existing
There is the technical problem as above of technology.
How the technical solution of the disclosure and the technical solution of the disclosure are solved with specifically embodiment below above-mentioned
Technical problem is described in detail.These specific embodiments can be combined with each other below, for the same or similar concept
Or process may repeat no more in certain embodiments.Below in conjunction with attached drawing, embodiment of the disclosure is described.
Embodiment one
The embodiment of the present disclosure provides a kind of feature extracting method based on convolutional neural networks, as shown in Figure 1, this method
Include:
Target image is input to scheduled convolutional neural networks by S101, and convolutional neural networks include S image procossing rank
Section, each image processing stage include multiple convolutional layers, and S is the positive integer greater than 1；
Convolutional neural networks include multiple images processing stage, and each image processing stage includes multiple convolutional layers.With volume
For product neural network is ResNet50, ResNet50 includes five image processing stages, i.e. 5 stage, 5 stage are successively
It is named as C1, C2, C3, C4 and C5.Different stage includes different convolutional layer, and C1 includes 1 convolutional layer, and C2 includes 3 volumes
Lamination, C3 include 4 convolutional layers, and C4 includes 6 convolutional layers, and C5 includes 3 convolutional layers.
S102 executes step A: to present image since present image processing stage is first image processing stage
The output for the multiple convolutional layers for including in processing stage is as a result, carry out phase according to scheduled convolution interlayer information interchange processing mode
It should handle, obtain the output result of present image processing stage and the input as next image processing stage, and will be next
A image processing stage is as present image processing stage；
Step A is the output to the multiple convolutional layers for including in present image processing stage as a result, according to scheduled convolution
Inter-layer information exchanges processing mode and carries out respective handling, obtains the output result of present image processing stage and as next figure
As the input of processing stage, and using next image processing stage as present image processing stage.I.e. at present image
The output for the multiple convolutional layers for including in the reason stage is as a result, handle multiple output results with scheduled convolution interlayer information interchange
Mode carries out respective handling, so that multiple output results mutually carry out information interchange, the unidirectional information with traditional convolution interlayer
Exchange way is had any different, and is not only the interaction in same channel between pixel and surrounding pixel, is also made channel and channel
Between there are the interaction of pixel, be conducive to the accuracy for improving present image processing stage output result.
If present image processing stage is first image processing stage, target image is input to convolutional neural networks
First image processing stage, if the image processing stage that present image processing stage is second or second or more,
The output result of a upper image processing stage is input to present image processing stage.
S103, circulation step A, until obtaining the exporting as a result, feature as target image of the S image processing stage
Extract result.
Step A is executed by circulation, so that the output result accuracy of each image processing stage improves, and then is improved special
Sign extracts the accuracy of result.
In the embodiments of the present disclosure, by the output to the multiple convolutional layers for including in present image processing stage as a result,
Convolution interlayer information interchange processing mode is taken to carry out respective handling, so that mutually carrying out information interchange side between multiple convolutional layers
Formula, rather than unidirectional information interchange is carried out between two convolutional layers, improve the accurate of present image processing stage output result
Degree, and then improve the accuracy of convolutional neural networks processing result.
Embodiment two
The embodiment of the present disclosure provides alternatively possible implementation, further includes implementing on the basis of example 1
Method shown in example two, wherein
Output to the multiple convolutional layers for including in present image processing stage is as a result, according to scheduled convolution inter-layer information
It exchanges processing mode and carries out respective handling, obtain the output result of present image processing stage and as next image procossing rank
The input of section, including S1021 (not marked in figure) and S1022 (not marked in figure):
S1021, by the 1st to T output result of the corresponding output of the 1st to T convolutional layer in present image processing stage
It is input in default non-local network's block module, and the 1st to T output result is carried out in default non-local network block module
First calculation processing obtains intermediate output as a result, T is the number of the convolutional layer in present image processing stage, and T be positive it is whole
Number；
S1022 carries out the second calculation processing according to the T output result and intermediate output result, obtains at present image
The output result in reason stage and input as next image processing stage.
The English name of non-local network's block module is Non-local Block, is a kind of non local connection (Non-
Local operations) calculation block (building block) be for handling the Long-range dependence relationship of video and image
Capture the general module of long-distance dependence.1st to T output result of the corresponding output of the 1st to T convolutional layer is input to default
In non-local network's block module, information interchange can be mutually carried out between the 1st to T output result, to realize channel and channel
Between information interchange, obtain the intermediate output after information interchange as a result, feature i.e. after information interchange.Centre output knot
After fruit and the T output result carry out the second calculation processing, the output result of present image processing stage is obtained and as under
The input of one image processing stage.
First calculation processing is mainly that the 1st to T output result mutually carries out information friendship in non-local network's block module
Stream, the predominantly intermediate output result of the second calculation processing carry out information with the T output result and are added processing.
The embodiment of the present disclosure provides another possible implementation, wherein
In present image processing stage, the mode of either one or two of the corresponding output of any one convolutional layer output result is determined,
Including any one of following:
S1023 (not marked in figure) or S1024 (not marked in figure),
S1023, by the output of the upper image processing stage of present image processing stage as a result, being input to t-th of convolution
Layer obtains the output of t-th of convolutional layer output as a result, t=1；
The output result that the t-1 convolutional layer exports is input to t-th volume in present image processing stage by S1024
Lamination obtains the output of t-th of convolutional layer output as a result, 2≤t≤T, T are positive integer.
As t=1, the input information of the 1st convolutional layer of present image processing stage is a upper image processing stage
Export result；As 2≤t≤T, the input information of t-th of convolutional layer of present image processing stage is that present image handles rank
The output result of the t-1 convolutional layer output of section.Wherein, if present image processing stage is the 1st image processing stage, and
When t=1, the input information of the 1st convolutional layer of the 1st image processing stage is target image.
It is input to t-th of convolutional layer (2≤t≤T), is realized same by the output result for exporting the t-1 convolutional layer
In channel, the information interchange of pixel and surrounding pixel.
The embodiment of the present disclosure provides another possible implementation, wherein
1st to T output result of the corresponding output of the 1st to T convolutional layer in present image processing stage is input to
In default non-local network's block module, and the first meter is carried out to the 1st to T output result in default non-local network block module
Calculation processing obtains intermediate output as a result, including S10211 (not marking in figure) and S10212 (not marking in figure):
S10211 is corresponded to the 1st to T convolutional layer in present image processing stage using default connection array function defeated
The 1st to T output result out links together, and is input in default non-local network's block module；
S10212, the 1st to T output result carry out Gaussian Computation processing in default non-local network block module, obtain
Centre output result.
Preferably, preset connection array function is concat function.Concat function be it is a kind of for connect two or
The method of more than two arrays, will not change existing array, only be returned only to a copy of connected number group.To carry out
The parameter of concat operation is array, then what is added is the element in array, rather than array itself.Such as, three input difference
For X1:B*C1*H*W, X2:B*C2*H*W and X3:B*C3*H*W, X1, X2 and X3 are connected using default connection array function
Become X:B* (C1+C2+C3) * H*W after connecing.
The 1st to T convolutional layer in present image processing stage is corresponded to the 1st of output using default connection array function
It links together to T output result, after becoming an input information, by this input information input to presetting non local net
In network block module, after this presets the progress Gaussian Computation processing of non-local network's block module, intermediate output result is obtained.It needs
Illustrate, after input information enters default non-local network's block module, preferably implementation port number, which first halves, is restored
Principle defers to the design philosophy of bottleneck network (bottleneck), reduces the calculation amount of half, while former data can be adopted by drop
Data after sample replace, and further decrease calculation amount.
The embodiment of the present disclosure provides another possible implementation, wherein
Any image processing stage and its including the output result of any convolutional layer be characteristic image, characteristic image is logical
Formula F eature maps=W*H*C is crossed to indicate,
Feature maps indicates that characteristic image, W indicate that the width of characteristic image, H indicate that the height of characteristic image, C indicate
The dimension of characteristic image.
In the embodiments of the present disclosure, by the output to the multiple convolutional layers for including in present image processing stage as a result,
Convolution interlayer information interchange processing mode is taken to carry out respective handling, so that mutually carrying out information interchange side between multiple convolutional layers
Formula, rather than unidirectional information interchange is carried out between two convolutional layers, improve the accurate of present image processing stage output result
Degree, and then improve the accuracy of convolutional neural networks processing result.
Embodiment three
The embodiment of the present disclosure provides a kind of feature deriving means based on convolutional neural networks, as shown in Fig. 2, this feature
Extraction element 20 may include: input module 201, processing module 202 and loop module 203, wherein
Input module 201, for target image to be input to scheduled convolutional neural networks, convolutional neural networks include S
A image processing stage, each image processing stage include multiple convolutional layers, and S is the positive integer greater than 1；
Processing module 202, for executing step A since present image processing stage is first image processing stage:
Output to the multiple convolutional layers for including in present image processing stage according to scheduled convolution interlayer information interchange as a result, handle
Mode carries out respective handling, obtains the output result of present image processing stage and as the defeated of next image processing stage
Enter, and using next image processing stage as present image processing stage；
Loop module 203 is used for circulation step A, until obtaining the output of the S image processing stage as a result, as mesh
The feature extraction result of logo image.
Further, processing module 202, including first processing module (not marking in figure) and Second processing module are (in figure
Do not mark):
First processing module, for since present image processing stage is first image processing stage, execution will to work as
1st to T output result of the corresponding output of the 1st to T convolutional layer in preceding image processing stage, which is input to, presets non local net
In network block module, and the first calculation processing is carried out to the 1st to T output result in default non-local network block module, obtained
Centre output is as a result, T is the number of the convolutional layer in present image processing stage, and T is positive integer；
Second processing module is obtained for carrying out the second calculation processing according to the T output result and intermediate output result
The output result of present image processing stage and input as next image processing stage, and by next image procossing rank
Duan Zuowei present image processing stage.
Further, first processing module was also used in present image processing stage, determined that any one convolutional layer is corresponding
Either one or two of output output result.It includes any one of following:
First processing module is also used to the output by the upper image processing stage of present image processing stage as a result, defeated
Enter to obtain the output of t-th of convolutional layer output to t-th of convolutional layer as a result, t=1；
First processing module was also used in present image processing stage, the output result that the t-1 convolutional layer is exported
It is input to t-th of convolutional layer, obtains the output of t-th of convolutional layer output as a result, 2≤t≤T, T are positive integer.
Further, first processing module, including connection input module (not marking in figure) and calculation processing module are (in figure
Do not mark):
Input module is connected, for executing utilization since present image processing stage is first image processing stage
Default connection array function ties the 1st to T output of the corresponding output of the 1st to T convolutional layer in present image processing stage
Fruit links together, and is input in default non-local network's block module；
Calculation processing module carries out Gaussian Computation in default non-local network block module for the 1st to T output result
Processing obtains intermediate output result.
Further, any image processing stage and its including the output result of any convolutional layer be characteristic image,
Characteristic image indicated by formula F eature maps=W*H*C,
Feature maps indicates that characteristic image, W indicate that the width of characteristic image, H indicate that the height of characteristic image, C indicate
The dimension of characteristic image.
Feature extraction side shown in the embodiment of the present disclosure one or embodiment two can be performed in the feature deriving means of the present embodiment
Method, realization principle is similar, and details are not described herein again.
In the embodiments of the present disclosure, by the output to the multiple convolutional layers for including in present image processing stage as a result,
Convolution interlayer information interchange processing mode is taken to carry out respective handling, so that mutually carrying out information interchange side between multiple convolutional layers
Formula, rather than unidirectional information interchange is carried out between two convolutional layers, improve the accurate of present image processing stage output result
Degree, and then improve the accuracy of convolutional neural networks processing result.
Example IV
The embodiment of the present disclosure additionally provides a kind of electronic equipment comprising:
One or more processors；
Memory；
One or more application program, wherein one or more application programs be stored in memory and be configured as by
One or more processors execute, and one or more programs are configured to: executing the feature according to any of the above-described embodiment
Extracting method.
The embodiment of the present disclosure additionally provides computer readable storage medium, is stored thereon with computer program, the program quilt
Processor realizes the feature extracting method that any of the above-described embodiment provides when executing.
Scheme in embodiment of the disclosure, turns next to Fig. 3, and it illustrates be suitable for being used to realizing the embodiment of the present disclosure
Electronic equipment 300 structural schematic diagram.Electronic equipment in the embodiment of the present disclosure can include but is not limited to such as move electricity
Words, laptop, digit broadcasting receiver, PDA (personal digital assistant), PAD (tablet computer), PMP (portable multimedia
Player), the mobile terminal and such as number TV, desktop computer etc. of car-mounted terminal (such as vehicle mounted guidance terminal) etc.
Fixed terminal.Electronic equipment shown in Fig. 3 is only an example, should not function and use scope to the embodiment of the present disclosure
Bring any restrictions.
As shown in figure 3, electronic equipment 300 may include processing unit (such as central processing unit, graphics processor etc.)
301, random access can be loaded into according to the program being stored in read-only memory (ROM) 302 or from storage device 308
Program in memory (RAM) 303 and execute various movements appropriate and processing.In RAM 303, it is also stored with electronic equipment
Various programs and data needed for 300 operations.Processing unit 301, ROM 302 and RAM 303 pass through the phase each other of bus 304
Even.Input/output (I/O) interface 305 is also connected to bus 304.
In general, following device can connect to I/O interface 305: including such as touch screen, touch tablet, keyboard, mouse, taking the photograph
As the input unit 306 of head, microphone, accelerometer, gyroscope etc.；Including such as liquid crystal display (LCD), loudspeaker, vibration
The output device 307 of dynamic device etc.；Storage device 308 including such as tape, hard disk etc.；And communication device 309.Communication device
309, which can permit electronic equipment 300, is wirelessly or non-wirelessly communicated with other equipment to exchange data.Although Fig. 3 shows tool
There is the electronic equipment 300 of various devices, it should be understood that being not required for implementing or having all devices shown.It can be with
Alternatively implement or have more or fewer devices.
Particularly, in accordance with an embodiment of the present disclosure, it may be implemented as computer above with reference to the process of flow chart description
Software program.For example, embodiment of the disclosure includes a kind of computer program product comprising be carried on computer-readable medium
On computer program, which includes the program code for method shown in execution flow chart.In such reality
It applies in example, which can be downloaded and installed from network by communication device 309, or from storage device 308
It is mounted, or is mounted from ROM 302.When the computer program is executed by processing unit 301, the embodiment of the present disclosure is executed
Method in the above-mentioned function that limits.
It should be noted that the above-mentioned computer-readable medium of the disclosure can be computer-readable signal media or meter
Calculation machine readable storage medium storing program for executing either the two any combination.Computer readable storage medium for example can be --- but not
Be limited to --- electricity, magnetic, optical, electromagnetic, infrared ray or semiconductor system, device or device, or any above combination.Meter
The more specific example of calculation machine readable storage medium storing program for executing can include but is not limited to: have the electrical connection, just of one or more conducting wires
Taking formula computer disk, hard disk, random access storage device (RAM), read-only memory (ROM), erasable type may be programmed read-only storage
Device (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), light storage device, magnetic memory device,
Or above-mentioned any appropriate combination.In the disclosure, computer readable storage medium can be it is any include or storage journey
The tangible medium of sequence, the program can be commanded execution system, device or device use or in connection.And at this
In open, computer-readable signal media may include in a base band or as the data-signal that carrier wave a part is propagated,
In carry computer-readable program code.The data-signal of this propagation can take various forms, including but not limited to
Electromagnetic signal, optical signal or above-mentioned any appropriate combination.Computer-readable signal media can also be computer-readable and deposit
Any computer-readable medium other than storage media, the computer-readable signal media can send, propagate or transmit and be used for
By the use of instruction execution system, device or device or program in connection.Include on computer-readable medium
Program code can transmit with any suitable medium, including but not limited to: electric wire, optical cable, RF (radio frequency) etc. are above-mentioned
Any appropriate combination.
Above-mentioned computer-readable medium can be included in above-mentioned electronic equipment；It is also possible to individualism, and not
It is fitted into the electronic equipment.
Above-mentioned computer-readable medium carries one or more program, when said one or multiple programs are by the electricity
When sub- equipment executes, so that the electronic equipment: obtaining at least two internet protocol addresses；Send to Node evaluation equipment includes institute
State the Node evaluation request of at least two internet protocol addresses, wherein the Node evaluation equipment is internet from described at least two
In protocol address, chooses internet protocol address and return；Receive the internet protocol address that the Node evaluation equipment returns；Its
In, the fringe node in acquired internet protocol address instruction content distributing network.
Alternatively, above-mentioned computer-readable medium carries one or more program, when said one or multiple programs
When being executed by the electronic equipment, so that the electronic equipment: receiving the Node evaluation including at least two internet protocol addresses and request；
From at least two internet protocol address, internet protocol address is chosen；Return to the internet protocol address selected；Wherein,
The fringe node in internet protocol address instruction content distributing network received.
The calculating of the operation for executing the disclosure can be write with one or more programming languages or combinations thereof
Machine program code, above procedure design language include object oriented program language-such as Java, Smalltalk, C+
+, it further include conventional procedural programming language-such as " C " language or similar programming language.Program code can
Fully to execute, partly execute on the user computer on the user computer, be executed as an independent software package,
Part executes on the remote computer or executes on a remote computer or server completely on the user computer for part.
In situations involving remote computers, remote computer can pass through the network of any kind --- including local area network (LAN)
Or wide area network (WAN)-is connected to subscriber computer, or, it may be connected to outer computer (such as utilize Internet service
Provider is connected by internet).
Flow chart and block diagram in attached drawing are illustrated according to the system of the various embodiments of the disclosure, method and computer journey
The architecture, function and operation in the cards of sequence product.In this regard, each box in flowchart or block diagram can generation
A part of one module, program segment or code of table, a part of the module, program segment or code include one or more use
The executable instruction of the logic function as defined in realizing.It should also be noted that in some implementations as replacements, being marked in box
The function of note can also occur in a different order than that indicated in the drawings.For example, two boxes succeedingly indicated are actually
It can be basically executed in parallel, they can also be executed in the opposite order sometimes, and this depends on the function involved.Also it to infuse
Meaning, the combination of each box in block diagram and or flow chart and the box in block diagram and or flow chart can be with holding
The dedicated hardware based system of functions or operations as defined in row is realized, or can use specialized hardware and computer instruction
Combination realize.
Being described in unit involved in the embodiment of the present disclosure can be realized by way of software, can also be by hard
The mode of part is realized.Wherein, the title of unit does not constitute the restriction to the unit itself under certain conditions, for example, the
One acquiring unit is also described as " obtaining the unit of at least two internet protocol addresses ".
Above description is only the preferred embodiment of the disclosure and the explanation to institute's application technology principle.Those skilled in the art
Member is it should be appreciated that the open scope involved in the disclosure, however it is not limited to technology made of the specific combination of above-mentioned technical characteristic
Scheme, while should also cover in the case where not departing from design disclosed above, it is carried out by above-mentioned technical characteristic or its equivalent feature
Any combination and the other technical solutions formed.Such as features described above has similar function with (but being not limited to) disclosed in the disclosure
Can technical characteristic replaced mutually and the technical solution that is formed.
Claims (10)
1. a kind of feature extracting method based on convolutional neural networks characterized by comprising
Target image is input to scheduled convolutional neural networks, the convolutional neural networks include S image processing stage, often
One image processing stage includes multiple convolutional layers, and the S is the positive integer greater than 1；
Since present image processing stage is first image processing stage, step A is executed: in present image processing stage
Including multiple convolutional layers output as a result, according to scheduled convolution interlayer information interchange processing mode carry out respective handling, obtain
Output result to present image processing stage and input as next image processing stage, and by next image procossing
Stage is as present image processing stage；
The step A is recycled, until obtaining the exporting as a result, feature as the target image of the S image processing stage
Extract result.
2. feature extracting method according to claim 1, which is characterized in that described to including in present image processing stage
Multiple convolutional layers output as a result, according to scheduled convolution interlayer information interchange processing mode carry out respective handling, worked as
The output result of preceding image processing stage and input as next image processing stage, comprising:
1st to T output result of the corresponding output of the 1st to T convolutional layer in present image processing stage is input to default
In non-local network's block module, and the is carried out to the 1st to the T output result in default non-local network's block module
One calculation processing obtains intermediate output as a result, the T is the number of the convolutional layer in the present image processing stage, and institute
Stating T is positive integer；
The second calculation processing is carried out according to the T output result and the intermediate output result, obtains present image processing stage
Output result and input as next image processing stage.
3. feature extracting method according to claim 1, which is characterized in that in present image processing stage, determine and appoint
The mode of either one or two of the corresponding output of one convolutional layer output result, including any one of following:
By the output of the upper image processing stage of present image processing stage as a result, being input to t-th of convolutional layer, t is obtained
The output of a convolutional layer output is as a result, t=1；
In present image processing stage, the output result that the t-1 convolutional layer exports is input to t-th of convolutional layer, is obtained
The output of t-th of convolutional layer output is as a result, 2≤t≤T, T are positive integer.
4. feature extracting method according to claim 2, which is characterized in that by present image processing stage
1st to T output result of the corresponding output of 1 to T convolutional layer is input in default non-local network's block module, and described pre-
If carrying out the first calculation processing to the 1st to the T output result in non-local network's block module, obtain intermediate output as a result,
Include:
The 1st to T convolutional layer in present image processing stage is corresponded to the 1st to T of output using default connection array function
A output result links together, and is input in default non-local network's block module；
1st to the T output result carries out Gaussian Computation processing in default non-local network's block module, obtains centre
Export result.
5. feature extracting method according to any one of claims 1 to 4, which is characterized in that any image processing stage
And its including the output result of any convolutional layer be characteristic image, the characteristic image passes through formula F eature maps=
W*H*C indicates,
Feature maps indicates that characteristic image, W indicate that the width of the characteristic image, H indicate the height of the characteristic image, C
Indicate the dimension of the characteristic image.
6. a kind of feature deriving means based on convolutional neural networks characterized by comprising
Input module, for target image to be input to scheduled convolutional neural networks, the convolutional neural networks include S figure
As processing stage, each image processing stage includes multiple convolutional layers, and the S is the positive integer greater than 1；
Processing module, for executing step A: to current since present image processing stage is first image processing stage
The output of the multiple convolutional layers for including in image processing stage as a result, according to scheduled convolution interlayer information interchange processing mode into
Row respective handling obtains the output result of present image processing stage and the input as next image processing stage, and will
Next image processing stage is as present image processing stage；
Loop module, for recycling the step A, until obtaining the output of the S image processing stage as a result, as the mesh
The feature extraction result of logo image.
7. feature deriving means according to claim 6, which is characterized in that the processing module, comprising:
First processing module, for since present image processing stage is first image processing stage, execution will currently to be schemed
As the 1st to T output result of the corresponding output of the 1st to T convolutional layer in processing stage is input to default non-local network's block
In module, and the first calculation processing is carried out to the 1st to the T output result in default non-local network's block module,
Intermediate output is obtained as a result, the T is the number of the convolutional layer in the present image processing stage, and the T is positive integer；
Second processing module is obtained for carrying out the second calculation processing according to the T output result and the intermediate output result
The output result of present image processing stage and input as next image processing stage, and by next image procossing rank
Duan Zuowei present image processing stage.
8. feature deriving means according to claim 7, which is characterized in that the first processing module, comprising:
Input module is connected, for since present image processing stage is first image processing stage, executing using default
Array function is connected to connect the 1st to T output result of the corresponding output of the 1st to T convolutional layer in present image processing stage
It is connected together, and is input in default non-local network's block module；
Calculation processing module carries out Gauss in default non-local network's block module for the 1st to the T output result
Calculation processing obtains intermediate output result.
9. a kind of electronic equipment, characterized in that it comprises:
One or more processors；
Memory；
One or more application program, wherein one or more of application programs are stored in the memory and are configured
To be executed by one or more of processors, one or more of programs are configured to: being executed according to claim 1~5
Described in any item feature extracting methods.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is by processor
Claim 1-5 described in any item feature extracting methods are realized when execution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910507522.3A CN110222829A (en) | 2019-06-12 | 2019-06-12 | Feature extracting method, device, equipment and medium based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910507522.3A CN110222829A (en) | 2019-06-12 | 2019-06-12 | Feature extracting method, device, equipment and medium based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110222829A true CN110222829A (en) | 2019-09-10 |
Family
ID=67816818
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910507522.3A Pending CN110222829A (en) | 2019-06-12 | 2019-06-12 | Feature extracting method, device, equipment and medium based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110222829A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110826694A (en) * | 2019-10-30 | 2020-02-21 | 瀚博半导体(上海)有限公司 | Image processing method and device based on convolutional neural network |
| CN110929780A (en) * | 2019-11-19 | 2020-03-27 | 腾讯科技（深圳）有限公司 | Video classification model construction method, video classification device, video classification equipment and media |
| CN111428664A (en) * | 2020-03-30 | 2020-07-17 | 厦门瑞为信息技术有限公司 | Real-time multi-person posture estimation method based on artificial intelligence deep learning technology for computer vision |
| CN113095106A (en) * | 2019-12-23 | 2021-07-09 | 华为数字技术（苏州）有限公司 | Human body posture estimation method and device |
| US20220214421A1 (en) * | 2020-02-27 | 2022-07-07 | Panasonic Intellectual Property Management Co., Ltd. | Estimation device, estimation method, and recording medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106097353A (en) * | 2016-06-15 | 2016-11-09 | 北京市商汤科技开发有限公司 | The method for segmenting objects merged based on multi-level regional area and device, calculating equipment |
| CN107728142A (en) * | 2017-09-18 | 2018-02-23 | 西安电子科技大学 | Radar High Range Resolution target identification method based on two-dimensional convolution network |
| CN108229418A (en) * | 2018-01-19 | 2018-06-29 | 北京市商汤科技开发有限公司 | Human body critical point detection method and apparatus, electronic equipment, storage medium and program |
| CN108269275A (en) * | 2018-02-23 | 2018-07-10 | 深圳市唯特视科技有限公司 | A kind of non local modeling method based on non local neural network |
| CN109815964A (en) * | 2019-01-31 | 2019-05-28 | 北京字节跳动网络技术有限公司 | The method and apparatus for extracting the characteristic pattern of image |
| CN109816037A (en) * | 2019-01-31 | 2019-05-28 | 北京字节跳动网络技术有限公司 | The method and apparatus for extracting the characteristic pattern of image |
| CN109840528A (en) * | 2019-01-31 | 2019-06-04 | 北京字节跳动网络技术有限公司 | The method and apparatus for extracting the characteristic pattern of image |
-
2019
- 2019-06-12 CN CN201910507522.3A patent/CN110222829A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106097353A (en) * | 2016-06-15 | 2016-11-09 | 北京市商汤科技开发有限公司 | The method for segmenting objects merged based on multi-level regional area and device, calculating equipment |
| CN107728142A (en) * | 2017-09-18 | 2018-02-23 | 西安电子科技大学 | Radar High Range Resolution target identification method based on two-dimensional convolution network |
| CN108229418A (en) * | 2018-01-19 | 2018-06-29 | 北京市商汤科技开发有限公司 | Human body critical point detection method and apparatus, electronic equipment, storage medium and program |
| CN108269275A (en) * | 2018-02-23 | 2018-07-10 | 深圳市唯特视科技有限公司 | A kind of non local modeling method based on non local neural network |
| CN109815964A (en) * | 2019-01-31 | 2019-05-28 | 北京字节跳动网络技术有限公司 | The method and apparatus for extracting the characteristic pattern of image |
| CN109816037A (en) * | 2019-01-31 | 2019-05-28 | 北京字节跳动网络技术有限公司 | The method and apparatus for extracting the characteristic pattern of image |
| CN109840528A (en) * | 2019-01-31 | 2019-06-04 | 北京字节跳动网络技术有限公司 | The method and apparatus for extracting the characteristic pattern of image |
Non-Patent Citations (1)
| Title |
|---|
| XIAOLONG WANG ET AL: "Non-local Neural Networks", 《ARXIV》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110826694A (en) * | 2019-10-30 | 2020-02-21 | 瀚博半导体(上海)有限公司 | Image processing method and device based on convolutional neural network |
| CN110929780A (en) * | 2019-11-19 | 2020-03-27 | 腾讯科技（深圳）有限公司 | Video classification model construction method, video classification device, video classification equipment and media |
| CN110929780B (en) * | 2019-11-19 | 2023-07-11 | 腾讯科技（深圳）有限公司 | Video classification model construction method, video classification device, video classification equipment and medium |
| CN113095106A (en) * | 2019-12-23 | 2021-07-09 | 华为数字技术（苏州）有限公司 | Human body posture estimation method and device |
| US20220214421A1 (en) * | 2020-02-27 | 2022-07-07 | Panasonic Intellectual Property Management Co., Ltd. | Estimation device, estimation method, and recording medium |
| CN111428664A (en) * | 2020-03-30 | 2020-07-17 | 厦门瑞为信息技术有限公司 | Real-time multi-person posture estimation method based on artificial intelligence deep learning technology for computer vision |
| CN111428664B (en) * | 2020-03-30 | 2023-08-25 | 厦门瑞为信息技术有限公司 | Computer vision real-time multi-person gesture estimation method based on deep learning technology |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110222829A (en) | Feature extracting method, device, equipment and medium based on convolutional neural networks | |
| CN111476309B (en) | Image processing method, model training method, device, equipment and readable medium | |
| CN111755078B (en) | Drug molecule attribute determination method, device and storage medium | |
| CN110321958A (en) | Training method, the video similarity of neural network model determine method | |
| CN107688823A (en) | A kind of characteristics of image acquisition methods and device, electronic equipment | |
| CN109726806A (en) | Information processing method and terminal device | |
| Strong et al. | Self-sorting map: An efficient algorithm for presenting multimedia data in structured layouts | |
| CN109360028A (en) | Method and apparatus for pushed information | |
| CN110222726A (en) | Image processing method, device and electronic equipment | |
| CN108764319A (en) | A kind of sample classification method and apparatus | |
| CN111414953A (en) | Point cloud classification method and device | |
| CN109902763A (en) | Method and apparatus for generating characteristic pattern | |
| CN113033580B (en) | Image processing method, device, storage medium and electronic equipment | |
| CN109948699A (en) | Method and apparatus for generating characteristic pattern | |
| CN109948762A (en) | Method and apparatus for generating two dimensional code | |
| CN110210501A (en) | Virtual objects generation method, electronic equipment and computer readable storage medium | |
| CN107688783A (en) | 3D rendering detection method, device, electronic equipment and computer-readable medium | |
| CN110288037A (en) | Image processing method, device and electronic equipment | |
| CN109325480A (en) | The input method and terminal device of identity information | |
| Yang et al. | A real-time image forensics scheme based on multi-domain learning | |
| CN110489955A (en) | Applied to the image procossing of electronic equipment, device, calculate equipment, medium | |
| CN110717555A (en) | Picture generation system and device based on natural language and generation countermeasure network | |
| CN110198473A (en) | Method for processing video frequency, device, electronic equipment and computer readable storage medium | |
| CN110362698A (en) | A kind of pictorial information generation method, device, mobile terminal and storage medium | |
| CN110717405A (en) | Face feature point positioning method, device, medium and electronic equipment |
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 |