CN109299722A - Characteristic pattern processing method, device and system and storage medium for neural network - Google Patents
Characteristic pattern processing method, device and system and storage medium for neural network Download PDFInfo
- Publication number
- CN109299722A CN109299722A CN201810932282.7A CN201810932282A CN109299722A CN 109299722 A CN109299722 A CN 109299722A CN 201810932282 A CN201810932282 A CN 201810932282A CN 109299722 A CN109299722 A CN 109299722A
- Authority
- CN
- China
- Prior art keywords
- characteristic pattern
- channel
- network
- network block
- layer
- 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
- 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
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
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 invention provides a kind of characteristic pattern processing method, device and system and storage medium for neural network.Method includes: that each network block at least one network block is cut layer via channel and receives the input feature vector figure of the network block, and the channel of input feature vector figure is cut to two parts, to obtain first group of characteristic pattern and second group of characteristic pattern；First group of characteristic pattern is inputted into the first access, and second group of characteristic pattern is inputted into alternate path；The output characteristic pattern of the output characteristic pattern of the first access and alternate path is carried out channel via channel articulamentum to connect, with the characteristic pattern after being connected；Layer is reset via channel, channel rearrangement is carried out to the characteristic pattern after connection, to obtain the output characteristic pattern of the network block.This working method can reduce the calculation amount of neural network, and improve the speed and accuracy rate of neural network.
Description
Technical field
The present invention relates to machine learning field, relate more specifically to a kind of characteristic pattern processing method for neural network,
Device and system and storage medium.
Background technique
In field of image recognition, the network layer of deep neural network can achieve layers up to a hundred, and port number can achieve
Thousands of.The recognition accuracy of deep neural network can increase with the depth of network and the increase of port number, but simultaneously
Need a large amount of calculation amount and parameter amount.In general, it is (per second to can achieve more than one hundred million FLOPs for the calculation amount of deep neural network
Flops).However, the mobile devices such as mobile phone, robot only have tens or several hundred MFLOPs, (per second million floating
Point processing number) calculation amount budget, this makes deep neural network be difficult to run on devices.
Summary of the invention
The present invention is proposed in view of the above problem.The present invention provides a kind of characteristic pattern processing for neural network
Methods, devices and systems and storage medium.
According to an aspect of the present invention, a kind of characteristic pattern processing method for neural network, the neural network are provided
Including at least one network block, each network block at least one described network block includes that channel cuts layer, the first access, the
Layer is reset in two accesses, channel articulamentum and channel, and the alternate path includes at least one convolutional layer, and method includes: for institute
Each network block at least one network block is stated, cuts the input feature vector figure that layer receives the network block via the channel, and
The channel of the input feature vector figure is cut to two parts, to obtain first group of characteristic pattern and second group of characteristic pattern；For described
First group of characteristic pattern is inputted first access by each network block at least one network block, and by described second
Group characteristic pattern inputs the alternate path；For each network block at least one described network block, connect via the channel
It connects layer and connects the output characteristic pattern of first access with the output characteristic pattern of alternate path progress channel, to be connected
Characteristic pattern after connecing；For each network block at least one described network block, layer is reset to the company via the channel
Characteristic pattern after connecing carries out channel rearrangement, to obtain the output characteristic pattern of the network block.
Illustratively, at least one described network block includes at least one first network block, at least one described first net
The alternate path of each first network block in network block includes the first convolutional layer, the second convolutional layer and third convolutional layer, institute
State method further include: for each first network block at least one described first network block, via first convolutional layer
Dimensionality reduction convolution is carried out to second group of characteristic pattern；For each first network block at least one described first network block,
Depth separation volume is carried out to the output characteristic pattern of first convolutional layer via second convolutional layer and the third convolutional layer
Product, to obtain the output characteristic pattern of the alternate path.
Illustratively, second convolutional layer is used to the output characteristic pattern to first convolutional layer and executes step-length be 1
Convolution, first access are the jump connecting paths between the channel cutting layer and the channel articulamentum.
Illustratively, second convolutional layer is used to the output characteristic pattern to first convolutional layer and executes step-length be 2
Convolution, first access include pond layer, the method also includes: first group of characteristic pattern is held via the pond layer
The pond that row step-length is 2, to obtain the output characteristic pattern of first access.
Illustratively, at least one described network block includes at least one second network block, at least one described second net
The alternate path of the second network block of each of network block includes residual error structure, and the residual error structure includes residual error branch and jump
Jump connection branch, the residual error branch include at least one described convolutional layer, the method also includes: for it is described at least one
The second network block of each of second network block carries out convolution to second group of characteristic pattern via the residual error branch；For
The second network block of each of at least one second network block, by the output characteristic pattern of the residual error branch and the jump
The second group of characteristic pattern for connecting branch output carries out element addition, to obtain the output characteristic pattern of the alternate path.
Illustratively, the residual error branch includes Volume Four lamination, the 5th convolutional layer and the 6th convolutional layer, wherein described
Method further include: for the second network block of each of at least one second network block, via the Volume Four lamination pair
Second group of characteristic pattern carries out dimensionality reduction convolution；For the second network block of each of at least one second network block, warp
Depth is carried out to the output characteristic pattern of the Volume Four lamination by the 5th convolutional layer and the 6th convolutional layer and separates convolution,
To obtain the output characteristic pattern of the residual error branch；Wherein, the port number and described the of the output characteristic pattern of the residual error branch
The port number of two groups of characteristic patterns is equal.
Illustratively, the 5th convolutional layer is used to the output characteristic pattern to the Volume Four lamination and executes step-length be 1
Convolution, first access are jump connecting path of the channel cutting layer to the channel articulamentum.
Illustratively, the neural network include sequentially connected initial convolutional layer, maximum pond layer, first network section,
Second network segment, third network segment, global pool layer and full articulamentum, wherein the first network section includes described at least one
Four network blocks in a network block, second network segment include eight network blocks at least one described network block, institute
Stating third network segment includes four network blocks at least one described network block.
Illustratively, the port number of first group of characteristic pattern and the port number of second group of characteristic pattern are equal.
According to a further aspect of the invention, a kind of characteristic pattern processing unit for neural network, the nerve net are provided
Network includes at least one network block, each network block at least one described network block include channel cut layer, the first access,
Layer is reset in alternate path, channel articulamentum and channel, and the alternate path includes at least one convolutional layer, and described device includes:
Channel cuts module, for cutting layer via the channel and receiving for each network block at least one described network block
The input feature vector figure of the network block, and the channel of the input feature vector figure is cut to two parts, to obtain first group of characteristic pattern
With second group of characteristic pattern；Input module, for for each network block at least one described network block, by described first group
Characteristic pattern inputs first access, and second group of characteristic pattern is inputted the alternate path；Channel link block, is used for
It is via the channel articulamentum that the output of first access is special for each network block at least one described network block
The output characteristic pattern progress channel of sign figure and the alternate path connects, with the characteristic pattern after being connected；Channel reordering module,
For resetting layer to the feature after the connection via the channel for each network block at least one described network block
Figure carries out channel rearrangement, to obtain the output characteristic pattern of the network block.
According to a further aspect of the invention, a kind of characteristic pattern processing system for neural network, including processor are provided
And memory, wherein computer program instructions are stored in the memory, the computer program instructions are by the processor
For executing the above-mentioned characteristic pattern processing method for neural network when operation.
According to a further aspect of the invention, a kind of storage medium is provided, stores program instruction on said storage,
Described program instruction is at runtime for executing the above-mentioned characteristic pattern processing method for neural network.
It is according to an embodiment of the present invention to be situated between for the characteristic pattern processing method of neural network, device and system and storage
Matter is cut by channel the channel of input feature vector figure being divided into two parts, enables input feature vector figure complete by part
Portion's convolution, can then be connected by channel and channel resets and two parts characteristic pattern is attached and is reset.This work side
Formula can reduce the calculation amount of neural network, and improve the speed and accuracy rate of neural network.
Detailed description of the invention
The embodiment of the present invention is described in more detail in conjunction with the accompanying drawings, the above and other purposes of the present invention,
Feature and advantage will be apparent.Attached drawing is used to provide to further understand the embodiment of the present invention, and constitutes explanation
A part of book, is used to explain the present invention together with the embodiment of the present invention, is not construed as limiting the invention.In the accompanying drawings,
Identical reference label typically represents same parts or step.
Fig. 1 shows for realizing the characteristic pattern treating method and apparatus according to an embodiment of the present invention for neural network
The schematic block diagram of exemplary electronic device；
Fig. 2 shows the schematic flows of the characteristic pattern processing method according to an embodiment of the invention for neural network
Figure；
Fig. 3 a shows the structural schematic diagram of the network block of neural network according to an embodiment of the invention；
Fig. 3 b shows the structural schematic diagram of the network block of neural network in accordance with another embodiment of the present invention；
Fig. 4 shows the structural schematic diagram of the network block of neural network in accordance with another embodiment of the present invention；
Fig. 5 shows the schematic frame of the characteristic pattern processing unit according to an embodiment of the invention for neural network
Figure；And
Fig. 6 shows the schematic frame of the characteristic pattern processing system according to an embodiment of the invention for neural network
Figure.
Specific embodiment
In order to enable the object, technical solutions and advantages of the present invention become apparent, root is described in detail below with reference to accompanying drawings
According to example embodiments of the present invention.Obviously, described embodiment is only a part of the embodiments of the present invention, rather than this hair
Bright whole embodiments, it should be appreciated that the present invention is not limited by example embodiment described herein.
To solve the above-mentioned problems, the embodiment of the present invention provides a kind of characteristic pattern processing processing method for neural network
And device, this method and device are cut by channel the channel of input feature vector figure being divided into two parts, so that input feature vector figure
Can by part not all convolution, can then be connected by channel and channel reset two parts characteristic pattern is attached and
It resets.This working method can reduce the calculation amount of neural network, and improve the speed and accuracy rate of neural network.According to this
The characteristic pattern treating method and apparatus for neural network of inventive embodiments can be applied to various deep learning fields, such as scheme
As identification etc..
Firstly, the characteristic pattern to describe for realizing according to an embodiment of the present invention for neural network is handled referring to Fig.1
The exemplary electronic device 100 of method and apparatus.
As shown in Figure 1, electronic equipment 100 includes one or more processors 102, one or more storage devices 104.It can
Selection of land, electronic equipment 100 can also include input unit 106, output device 108 and image collecting device 110, these groups
Part passes through the interconnection of bindiny mechanism's (not shown) of bus system 112 and/or other forms.It should be noted that electronics shown in FIG. 1 is set
Standby 100 component and structure be it is illustrative, and not restrictive, as needed, the electronic equipment also can have it
His component and structure.
The processor 102 can use microprocessor, digital signal processor (DSP), field programmable gate array
(FPGA), at least one of programmable logic array (PLA) example, in hardware realizes that the processor 102 can be center
Processing unit (CPU), image processor (GPU), dedicated integrated circuit (ASIC) have data-handling capacity and/or refer to
The combination of one or more of processing unit of other forms of executive capability is enabled, and can control the electronic equipment
Other components in 100 are to execute desired function.
The storage device 104 may include one or more computer program products, and the computer program product can
To include various forms of computer readable storage mediums, such as volatile memory and/or nonvolatile memory.It is described easy
The property lost memory for example may include random access memory (RAM) and/or cache memory (cache) etc..It is described non-
Volatile memory for example may include read-only memory (ROM), hard disk, flash memory etc..In the computer readable storage medium
On can store one or more computer program instructions, processor 102 can run described program instruction, to realize hereafter institute
The client functionality (realized by processor) in the embodiment of the present invention stated and/or other desired functions.In the meter
Can also store various application programs and various data in calculation machine readable storage medium storing program for executing, for example, the application program use and/or
The various data etc. generated.
The input unit 106 can be the device that user is used to input instruction, and may include keyboard, mouse, wheat
One or more of gram wind and touch screen etc..
The output device 108 can export various information (such as image and/or sound) to external (such as user), and
It and may include one or more of display, loudspeaker etc..Optionally, the input unit 106 and the output device
108 can integrate together, be realized using same interactive device (such as touch screen).
Described image acquisition device 110 can acquire image (including still image and video frame), and will be collected
Image is stored in the storage device 104 for the use of other components.Image collecting device 110 can be individual camera,
The imaging sensor in camera or capture machine in mobile terminal.It should be appreciated that image collecting device 110 is only example, electricity
Sub- equipment 100 can not include image collecting device 110.In such a case, it is possible to using other with Image Acquisition ability
Device acquire image, and the image of acquisition is sent to electronic equipment 100.
Illustratively, for realizing the characteristic pattern treating method and apparatus according to an embodiment of the present invention for neural network
Exemplary electronic device can be realized in the equipment of personal computer or remote server etc..
In the following, reference Fig. 2 is described into the characteristic pattern processing method according to an embodiment of the present invention for being used for neural network, it is described
Neural network includes at least one network block, and each network block at least one described network block includes that channel cuts layer, the
Layer is reset in one access, alternate path, channel articulamentum and channel, and the alternate path includes at least one convolutional layer.Fig. 2 shows
The schematic flow chart of characteristic pattern processing method 200 according to an embodiment of the invention for neural network.Such as Fig. 2 institute
Show, the characteristic pattern processing method 200 for neural network includes the following steps S210, S220, S230 and S240.
Is cut by layer via channel and receives the network for each network block at least one network block in step S210
The input feature vector figure of block, and the channel of input feature vector figure is cut to two parts, it is special to obtain first group of characteristic pattern and second group
Sign figure.
Fig. 3 a shows the structural schematic diagram of the network block of neural network according to an embodiment of the invention.Such as Fig. 3 a institute
Show, block structure includes that channel cuts layer, channel articulamentum and channel rearrangement layer, cuts between layer and channel articulamentum and has in channel
There are two accesses, i.e. the first access, alternate path.
It is cut by channel, the channel for the input feature vector figure for inputting network block can be cut to two parts.Assuming that input
The port number of characteristic pattern is caA, the port number of the first group of characteristic pattern obtained after cutting is c1A, being obtained after cutting second group
The port number of characteristic pattern is c2It is a, then ca=c1+c2.That is, the port number of first group of characteristic pattern and second group of characteristic pattern it
It is equal with the port number of input feature vector figure.The port number of first group of characteristic pattern and the port number of second group of characteristic pattern can bases
It needs to set.The port number of first group of characteristic pattern and the port number of second group of characteristic pattern can be equal or different.Compare desirable
It is that the port number of first group of characteristic pattern and the port number of second group of characteristic pattern are equal, compared with port number not equal situation, channel
The equal speed that neural network can be improved of number.
Each network block at least one network block is led to first group of characteristic pattern input first in step S220
Road, and second group of characteristic pattern is inputted into alternate path.
Channel is cut into the two groups of characteristic patterns obtained and inputs the first access and alternate path respectively.For the first access and
For two accesses, in the case where access is jump connection (shortcut) access, the characteristic pattern for inputting the access is not carried out
Especially processing, directly exports characteristic pattern.In the case where including some network layers (such as convolutional layer) in access, to defeated
The characteristic pattern for entering the access carries out respective handling, obtains treated characteristic pattern and will treated characteristic pattern output.
In step S230, for each network block at least one network block, via channel articulamentum by the first access
The output characteristic pattern of output characteristic pattern and alternate path carry out channel and connect, with the characteristic pattern after being connected.
(channel concatenate) is connected by channel, the output characteristic pattern of the first access and second can be led to
The output characteristic pattern on road connects together, the characteristic pattern after being connected.Assuming that the port number of the output characteristic pattern of the first access
For c3A, the port number of the output characteristic pattern of alternate path is c4A, the port number of the characteristic pattern after connection is cbIt is a, then cb=c3
+c4。
In step S240, for each network block at least one network block, after resetting layer to connection via channel
Characteristic pattern carries out channel rearrangement, to obtain the output characteristic pattern of the network block.
(channel shuffle) is reset by channel, the characteristic pattern in different channels can be upset and be resequenced,
Obtain new characteristic pattern.When channel is reset, the mode that characteristic pattern is resequenced can be set as needed.At one
In example, the output characteristic pattern of the first access can be divided into m group, and the output characteristic pattern of alternate path is also divided into m
Group resets in layer in channel, i-th group of characteristic pattern in the output characteristic pattern of alternate path can be come to the output of alternate path
After i-th group of characteristic pattern in characteristic pattern and before i+1 group characteristic pattern.For example, it is assumed that the output characteristic pattern packet of the first access
Four groups of characteristic patterns that number is 1,2,3,4 are included, the output characteristic pattern of alternate path includes four groups of features that number is 5,6,7,8
Figure, after resetting, the sequence of the characteristic pattern of acquisition can be 1,5,2,6,3,7,4,8.
Channel, which is reset, can make infeasible channel is cut, channel connects originally to become feasible, if without channel weight
Row, then network can be divided into two parts by being cut using channel, a part carries out full convolution, and another part is without operation, in this way
Meeting is so that neural network can not train well.Channel reordering operations can make two parts characteristic pattern for cutting acquisition mutually pass on from one to another
Information is passed, avoids passage portion from passing through convolution always or does not pass through convolution always, so that neural network can be trained normally.
It only needs transposition to operate (dimension shuffle) in addition, channel is reset, does not need calculation amount, therefore will not additionally increase
The calculation amount of neural network.
Characteristic pattern processing method according to an embodiment of the present invention for neural network, has the following characteristics that
1) it cuts to obtain passage portion by channel and carries out convolutional calculation, the port number of convolution can be reduced in this way, and adopt
It is compared with the neural network of grouping convolution, the processing speed of neural network can be improved, because the speed of grouping convolution is slow；
2) channel of input feature vector figure is cut, the output that can neatly control the first access and alternate path is logical
Road number, one of the first access and alternate path can be used to implement convolution, so that convolutional layer no longer need to maintain with it is entire
The consistent port number of input feature vector figure of block structure, so that the design of convolutional layer is more flexible；In addition, also due to above-mentioned reason,
So that the port number of input feature vector figure can increase considerably, in the case where calculation amount is constant so as to effectively improve mind
Information representation ability through network；
3) after splicing two parts channel, then channel rearrangement is carried out, convolution or beginning can be passed through always to avoid passage portion
Eventually not by convolution, while channel resets and only transposition is needed to operate, and not will increase calculation amount.
Due to above-mentioned characteristic, under identical abbreviation multiple, opposite known models, the embodiment of the present invention provides neural network
Possess apparent accuracy benefits and speed advantage.Compared to other light weight models, neural network provided in an embodiment of the present invention
Allow there are more port numbers, i.e., stronger code capacity has on especially small network and more preferably shows.Inventor exists
It is tested in ImageNet classification task, experiment display, neural network provided in an embodiment of the present invention is than other models in speed
It is all more superior on degree and in accuracy rate.For example, accuracy rate ratio MobileNet improves 8.7% under 40MFLOPs magnitude, than
ShuffleNet (g=1) improves 4.4%.In addition, inventor is also tested for actual image processor (GPU) hardware-accelerated feelings
Condition, the results show that performance of the neural network provided in an embodiment of the present invention on GTX 1080Ti is better than
ShuffleNet。
Illustratively, the characteristic pattern processing method according to an embodiment of the present invention for neural network can have storage
It is realized in the unit or system of device and processor.
Characteristic pattern processing method according to an embodiment of the present invention for neural network can be deployed at personal terminal, all
Such as smart phone, tablet computer, personal computer.
Alternatively, the characteristic pattern processing method according to an embodiment of the present invention for neural network can also dispose with being distributed
At server end and client.For example, can client obtain video flowing (such as Image Acquisition end acquisition user people
Face image), the image that client will acquire sends server end (or cloud) to, carries out characteristic pattern by server end (or cloud)
Processing.
According to embodiments of the present invention, at least one network block may include at least one first network block, at least one
The alternate path of each first network block in one network block includes the first convolutional layer, the second convolutional layer and third convolutional layer, side
Method 200 can also include: for each first network block at least one first network block, via the first convolutional layer to second
Group characteristic pattern carries out dimensionality reduction convolution；For each first network block at least one first network block, via the second convolutional layer
Depth is carried out with output characteristic pattern of the third convolutional layer to the first convolutional layer and separates convolution, to obtain the output feature of alternate path
Figure.
With continued reference to Fig. 3 a, cutting between layer and channel articulamentum in channel includes two paths, and left side is the first access,
Right side is alternate path.Alternate path includes three convolutional layers, is followed successively by the first convolutional layer, the second convolutional layer and third convolution
Layer.Illustratively, the second convolutional layer and third convolutional layer can be used to implement depth separation convolution.Depth separation convolution refer to by
3x3 convolution operation splits into two parts: the convolution (i.e. the second convolutional layer) and 1x1 convolution (i.e. third convolution of 3x3 channel dimension
Layer), it is therefore an objective to reduce calculation amount.In addition, can be separated in depth to reduce port number in 3x3 channel dimension conventional part
1x1 convolution (i.e. the first convolutional layer) is used to reduce channel dimension before convolution.
Compared with the port number of the characteristic pattern (i.e. second group of characteristic pattern) of input alternate path, the output feature of alternate path
The port number of figure can remain unchanged, increases or reduce, this is determined by the convolutional layer that alternate path includes.Alternate path includes
Convolutional layer can be set as needed.
According to embodiments of the present invention, the second convolutional layer is used to the output characteristic pattern to the first convolutional layer and executes step-length be 1
Convolution, the first access are the jump connecting paths between channel cutting layer and channel articulamentum.
The structure of network block shown in Fig. 3 a is the structure that step-length (stride) is 1.As shown in Figure 3a, the first of left side is logical
Road is jump connecting path, i.e., channel cutting layer is directly connected to channel articulamentum, channel is enabled to cut layer for first
Group characteristic pattern is input to channel articulamentum via the first access.In the first access, extra process is not executed.Those skilled in the art
In the case that member is appreciated that the convolutional layer in alternate path executes the convolution that step-length is 1, the size of characteristic pattern before and after convolution
(i.e. width and height) is constant.
According to embodiments of the present invention, the second convolutional layer is used to the output characteristic pattern to the first convolutional layer and executes step-length be 2
Convolution, the first access include pond layer, and it is 2 to first group of characteristic pattern execution step-length that method 200, which can also include: via pond layer,
Pond, with obtain the first access output characteristic pattern.
Fig. 3 b shows the structural schematic diagram of the network block of neural network in accordance with another embodiment of the present invention.With Fig. 3 a class
As, in fig 3b, cutting between layer and channel articulamentum in channel includes two paths, and left side is the first access, and right side is
Alternate path.The alternate path of Fig. 3 b and the alternate path of Fig. 3 a are substantially similar, and what difference was the execution of the second convolutional layer is step
A length of 2 convolution.In this case, the first access can no longer be jump connecting path, and may include some network layers.
As shown in Figure 3b, the first access includes a pond layer, and the convolution kernel size of the pond layer is 3 × 3, step-length 2.This field skill
In the case that art personnel are appreciated that the convolutional layer in alternate path executes the convolution that step-length is 2, characteristic pattern after convolution
Size (i.e. width and height) can reduce half.First access is operated by pondization also reduces one for the size of first group of characteristic pattern
Half, the size of the output characteristic pattern of the output characteristic pattern and alternate path of the first access subsequent in this way still maintains consistent, convenient
Carry out channel connection.
According to embodiments of the present invention, at least one network block includes at least one second network block, at least one second net
The alternate path of the second network block of each of network block includes residual error structure, and residual error structure includes residual error branch and jump connection branch
Road, residual error branch include at least one convolutional layer, and method 200 can also include: for every at least one second network block
A second network block carries out convolution to second group of characteristic pattern via residual error branch；For every at least one second network block
Second group of characteristic pattern of the output characteristic pattern of residual error branch and jump connection branch output is carried out element phase by a second network block
Add, to obtain the output characteristic pattern of alternate path.
Optionally, block structure may include residual error structure, can not also include residual error structure.In order to distinguish, will not include
The network block of residual error structure is known as first network block (as shown in Figure 3a and Figure 3b shows), will include that the network block of residual error structure is known as the
Two network blocks.Herein, the first, second etc term is only used for distinguishing purpose, is not offered as sequence or other special contain
Justice.
Fig. 4 shows the structural schematic diagram of the network block of neural network in accordance with another embodiment of the present invention.In Fig. 4,
First access is jump connecting path, and alternate path is divided into two branches, i.e. residual error branch and jump connection branch, the two branch
Road forms residual error structure.
The net in the alternate path in operation and Fig. 3 a that the network layer for including in residual error branch and each network layer are implemented
Network layers can be similar, and details are not described herein again.Jump connection branch is to be directly connected to branch, and second group of characteristic pattern is transmitted
To at the output of residual error branch, enables the output characteristic pattern of residual error branch to carry out element with second group of characteristic pattern and be added.It is residual
The port number of the port number and second group of characteristic pattern of the output characteristic pattern of poor branch is consistent, and the feature of residual error branch output
The size of figure is in the same size with second group of characteristic pattern, so that the two is able to carry out element addition.It is added in alternate path residual
Poor structure can enable neural network normally to restrain in training, avoid the gradient disappearance problem that may occur.
According to embodiments of the present invention, residual error branch may include Volume Four lamination, the 5th convolutional layer and the 6th convolutional layer,
In, method 200 can also include: for the second network block of each of at least one second network block, via Volume Four lamination
Dimensionality reduction convolution is carried out to second group of characteristic pattern；For the second network block of each of at least one second network block, via the 5th
Convolutional layer and the 6th convolutional layer carry out depth to the output characteristic pattern of Volume Four lamination and separate convolution, to obtain the defeated of residual error branch
Characteristic pattern out；Wherein, the port number of the output characteristic pattern of residual error branch is equal with the port number of second group of characteristic pattern.
As shown in figure 4, alternate path includes three convolutional layers, it is followed successively by Volume Four lamination, the 5th convolutional layer and volume six
Lamination.Illustratively, the 5th convolutional layer and the 6th convolutional layer can be used to implement depth separation convolution comprising the channel 3x3 dimension
The convolution (i.e. the 5th convolutional layer) and 1x1 convolution (i.e. the 6th convolutional layer) of degree.As described above, depth separation convolution can reduce meter
Calculation amount.In addition, in order to reduce port number in 3x3 channel dimension conventional part 1x1 volumes can be used before depth separates convolution
Product (i.e. Volume Four lamination) reduces channel dimension.
According to embodiments of the present invention, the 5th convolutional layer is used to the output characteristic pattern to Volume Four lamination and executes step-length be 1
Convolution, the first access are jump connecting path of the channel cutting layer to channel articulamentum.
The structure of network block shown in Fig. 4 is the structure that step-length (stride) is 1.As shown in figure 4, first access in left side
It is jump connecting path, i.e., channel cutting layer is directly connected to channel articulamentum, channel is enabled to cut layer for first group
Characteristic pattern is input to channel articulamentum via the first access.In the first access, extra process is not executed.Those skilled in the art
It is appreciated that in the case that the convolutional layer in alternate path executes the convolution that step-length is 1, the size of characteristic pattern before and after convolution
(i.e. width and height) is constant.
According to embodiments of the present invention, neural network includes sequentially connected initial convolutional layer, maximum pond layer, first network
Section, the second network segment, third network segment, global pool layer and full articulamentum, wherein first network section includes at least one network
Four network blocks in block, the second network segment include eight network blocks at least one network block, and third network segment includes extremely
Four network blocks in a few network block.
Illustratively, the whole network structure of neural network can use three phases (network segment), each network segment point
Not Chong Fu 4,8,4 above-mentioned network blocks, a kind of schematic structure of neural network is as shown in table 1 below.The nerve net being shown in Table 1
In network structure, each network segment executes the down-sampling that a step-length is 2 first with block structure as shown in Figure 3b, then utilizes
Block structure as shown in Figure 3a executes the convolution that subsequent step-length is 1.For example, first network segment (network segment 2) is first with such as
Block structure shown in Fig. 3 b executes the down-sampling that a step-length is 2, then executes a hyposynchronization using block structure as shown in Figure 3a
A length of 1 convolution.
The schematic structure of 1. neural network of table
In light weight model design, existing neural network model, neural network provided in an embodiment of the present invention are compared
Performance be obviously improved.As described above, inventor has done comparative test on ImageNet data set.The results show that this
The neural network that inventive embodiments provide promotes 8.7% than MobileNet in accuracy rate, at the same speed, than
ShuffleNet promotes 4.4%.Referring to table 2, existing neural network and neural network provided in an embodiment of the present invention are shown
Performance.
Above-mentioned the results show neural network provided in an embodiment of the present invention can save calculation amount significantly, simultaneously
Accuracy can be effectively improved, current best level has been reached in the design of light weight model.
The performance of table 2. existing neural network and neural network provided in an embodiment of the present invention
According to a further aspect of the invention, a kind of characteristic pattern processing unit for neural network, the neural network are provided
Including at least one network block, each network block at least one described network block includes that channel cuts layer, the first access, the
Layer is reset in two accesses, channel articulamentum and channel, and the alternate path includes at least one convolutional layer.Fig. 5 is shown according to this
The schematic block diagram of the characteristic pattern processing unit 500 for neural network of invention one embodiment.
As shown in figure 5, the characteristic pattern processing unit 500 according to an embodiment of the present invention for neural network is cut out including channel
Cut-off-die block 510, input module 520, channel link block 530 and channel reordering module 540.The modules can execute respectively
Above in conjunction with each step/function of Fig. 2-4 characteristic pattern processing method for neural network described.Below only to the use
It is described in the major function of each component of the characteristic pattern processing unit 500 of neural network, and omits and had been described above
Detail content.
Channel cuts module 510 and is used for for each network block at least one described network block, via the channel
It cuts layer and receives the input feature vector figure of the network block, and the channel of the input feature vector figure is cut to two parts, to obtain the
One group of characteristic pattern and second group of characteristic pattern.Channel cuts module 510 can processor 102 in electronic equipment as shown in Figure 1
The program instruction that stores in Running storage device 103 is realized.
Input module 520 is used for for each network block at least one described network block, by first group of feature
Figure inputs first access, and second group of characteristic pattern is inputted the alternate path.Input module 520 can be by Fig. 1
Shown in the program instruction that stores in 102 Running storage device 103 of processor in electronic equipment realize.
Channel link block 530 is used for for each network block at least one described network block, via the channel
The output characteristic pattern of first access is carried out channel with the output characteristic pattern of the alternate path and connected by articulamentum, to obtain
Characteristic pattern after connection.Channel link block 530 can the operation storage dress of processor 102 in electronic equipment as shown in Figure 1
The program instruction that stores in 103 is set to realize.
Channel reordering module 540 is used for for each network block at least one described network block, via the channel
It resets layer and channel rearrangement is carried out to the characteristic pattern after the connection, to obtain the output characteristic pattern of the network block.Reset mould in channel
The program instruction that block 540 can store in 102 Running storage device 103 of processor in electronic equipment as shown in Figure 1 comes real
It is existing.
Those of ordinary skill in the art may be aware that list described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and algorithm steps can be realized with the combination of electronic hardware or computer software and electronic hardware.These functions are actually
It is implemented in hardware or software, the specific application and design constraint depending on technical solution.Professional technician
Each specific application can be used different methods to achieve the described function, but this realization is it is not considered that exceed
The scope of the present invention.
Fig. 6 shows the signal of the characteristic pattern processing system 600 according to an embodiment of the invention for neural network
Property block diagram.The neural network includes at least one network block, and each network block at least one described network block includes logical
Road cuts layer, the first access, alternate path, channel articulamentum and channel and resets layer, and the alternate path includes at least one volume
Lamination.Characteristic pattern processing system 600 for neural network includes storage device (i.e. memory) 610 and processor 620.
The storage of storage device 610 is handled for realizing the characteristic pattern according to an embodiment of the present invention for neural network
The computer program instructions of corresponding steps in method.
The processor 620 is for running the computer program instructions stored in the storage device 610, to execute basis
The corresponding steps of the characteristic pattern processing method for neural network of the embodiment of the present invention.
In one embodiment, for executing following step when the computer program instructions are run by the processor 620
It is rapid: for each network block at least one described network block, to cut the input that layer receives the network block via the channel
Characteristic pattern, and the channel of the input feature vector figure is cut to two parts, to obtain first group of characteristic pattern and second group of characteristic pattern；
First group of characteristic pattern is inputted into first access, and second group of characteristic pattern is inputted into the alternate path；Via
The output characteristic pattern of the output characteristic pattern of first access and the alternate path is carried out channel company by the channel articulamentum
It connects, with the characteristic pattern after being connected；Layer is reset via the channel, and channel rearrangement is carried out to the characteristic pattern after the connection, with
Obtain the output characteristic pattern of the network block.
In addition, according to embodiments of the present invention, additionally providing a kind of storage medium, storing program on said storage
Instruction, when described program instruction is run by computer or processor for execute the embodiment of the present invention for neural network
The corresponding steps of characteristic pattern processing method, and for realizing the characteristic pattern according to an embodiment of the present invention for neural network at
Manage the corresponding module in device.The storage medium for example may include the storage unit of the storage card of smart phone, tablet computer
It is part, the hard disk of personal computer, read-only memory (ROM), Erasable Programmable Read Only Memory EPROM (EPROM), portable compact
Disk read-only memory (CD-ROM), any combination of USB storage or above-mentioned storage medium.
In one embodiment, described program instruction can make computer or place when being run by computer or processor
Reason device realizes each functional module of the characteristic pattern processing unit according to an embodiment of the present invention for neural network, and and/or
Person can execute the characteristic pattern processing method according to an embodiment of the present invention for neural network.
In one embodiment, described program instruction is at runtime for executing following steps: for it is described at least one
Each network block in network block cuts layer via the channel and receives the input feature vector figure of the network block, and by the input
The channel of characteristic pattern is cut to two parts, to obtain first group of characteristic pattern and second group of characteristic pattern；By first group of characteristic pattern
First access is inputted, and second group of characteristic pattern is inputted into the alternate path；Via the channel articulamentum by institute
The output characteristic pattern of the output characteristic pattern and the alternate path of stating the first access carries out channel and connects, with the spy after being connected
Sign figure；Layer is reset via the channel, channel rearrangement is carried out to the characteristic pattern after the connection, to obtain the output of the network block
Characteristic pattern.
Each module in characteristic pattern processing system according to an embodiment of the present invention for neural network can pass through basis
The processor operation for the electronic equipment that the implementation of the embodiment of the present invention is handled for the characteristic pattern of neural network is deposited in memory
The computer program instructions of storage realize, or can in the computer of computer program product according to an embodiment of the present invention can
The realization when computer instruction for reading to store in storage medium is run by computer.
Although describing example embodiment by reference to attached drawing here, it should be understood that above example embodiment are only exemplary
, and be not intended to limit the scope of the invention to this.Those of ordinary skill in the art can carry out various changes wherein
And modification, it is made without departing from the scope of the present invention and spiritual.All such changes and modifications are intended to be included in appended claims
Within required the scope of the present invention.
Those of ordinary skill in the art may be aware that list described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and algorithm steps can be realized with the combination of electronic hardware or computer software and electronic hardware.These functions are actually
It is implemented in hardware or software, the specific application and design constraint depending on technical solution.Professional technician
Each specific application can be used different methods to achieve the described function, but this realization is it is not considered that exceed
The scope of the present invention.
In several embodiments provided herein, it should be understood that disclosed device and method can pass through it
Its mode is realized.For example, apparatus embodiments described above are merely indicative, for example, the division of the unit, only
Only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple units or components can be tied
Another equipment is closed or is desirably integrated into, or some features can be ignored or not executed.
In the instructions provided here, numerous specific details are set forth.It is to be appreciated, however, that implementation of the invention
Example can be practiced without these specific details.In some instances, well known method, structure is not been shown in detail
And technology, so as not to obscure the understanding of this specification.
Similarly, it should be understood that in order to simplify the present invention and help to understand one or more of the various inventive aspects,
To in the description of exemplary embodiment of the present invention, each feature of the invention be grouped together into sometimes single embodiment, figure,
Or in descriptions thereof.However, the method for the invention should not be construed to reflect an intention that i.e. claimed
The present invention claims features more more than feature expressly recited in each claim.More precisely, such as corresponding power
As sharp claim reflects, inventive point is that the spy of all features less than some disclosed single embodiment can be used
Sign is to solve corresponding technical problem.Therefore, it then follows thus claims of specific embodiment are expressly incorporated in this specific
Embodiment, wherein each, the claims themselves are regarded as separate embodiments of the invention.
It will be understood to those skilled in the art that any combination pair can be used other than mutually exclusive between feature
All features disclosed in this specification (including adjoint claim, abstract and attached drawing) and so disclosed any method
Or all process or units of equipment are combined.Unless expressly stated otherwise, this specification (is wanted including adjoint right
Ask, make a summary and attached drawing) disclosed in each feature can be replaced with an alternative feature that provides the same, equivalent, or similar purpose.
In addition, it will be appreciated by those of skill in the art that although some embodiments described herein include other embodiments
In included certain features rather than other feature, but the combination of the feature of different embodiments mean it is of the invention
Within the scope of and form different embodiments.For example, in detail in the claims, embodiment claimed it is one of any
Can in any combination mode come using.
Various component embodiments of the invention can be implemented in hardware, or to run on one or more processors
Software module realize, or be implemented in a combination thereof.It will be understood by those of skill in the art that can be used in practice
Microprocessor or digital signal processor (DSP) are realized at the characteristic pattern according to an embodiment of the present invention for neural network
Manage some or all functions of some modules in device.The present invention is also implemented as executing side as described herein
Some or all program of device (for example, computer program and computer program product) of method.Such this hair of realization
Bright program can store on a computer-readable medium, or may be in the form of one or more signals.It is such
Signal can be downloaded from an internet website to obtain, and is perhaps provided on the carrier signal or is provided in any other form.
It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and ability
Field technique personnel can be designed alternative embodiment without departing from the scope of the appended claims.In the claims,
Any reference symbol between parentheses should not be configured to limitations on claims.Word "comprising" does not exclude the presence of not
Element or step listed in the claims.Word "a" or "an" located in front of the element does not exclude the presence of multiple such
Element.The present invention can be by means of including the hardware of several different elements and being come by means of properly programmed computer real
It is existing.In the unit claims listing several devices, several in these devices can be through the same hardware branch
To embody.The use of word first, second, and third does not indicate any sequence.These words can be explained and be run after fame
Claim.
The above description is merely a specific embodiment or to the explanation of specific embodiment, protection of the invention
Range is not limited thereto, and anyone skilled in the art in the technical scope disclosed by the present invention, can be easily
Expect change or replacement, should be covered by the protection scope of the present invention.Protection scope of the present invention should be with claim
Subject to protection scope.
Claims (12)
1. a kind of characteristic pattern processing method for neural network, the neural network include at least one network block, it is described extremely
Each network block in a few network block includes that channel cuts layer, the first access, alternate path, channel articulamentum and channel weight
Layer is arranged, the alternate path includes at least one convolutional layer, which comprises
For each network block at least one described network block,
Layer is cut via the channel and receives the input feature vector figure of the network block, and the channel of the input feature vector figure is cut to
Two parts, to obtain first group of characteristic pattern and second group of characteristic pattern；
First group of characteristic pattern is inputted into first access, and second group of characteristic pattern is inputted into the alternate path；
Via the channel articulamentum by first access output characteristic pattern and the alternate path output characteristic pattern into
Row of channels connection, with the characteristic pattern after being connected；
Layer is reset via the channel, and channel rearrangement is carried out to the characteristic pattern after the connection, it is special with the output for obtaining the network block
Sign figure.
2. the method for claim 1, wherein at least one described network block includes at least one first network block, institute
The alternate path for stating each first network block at least one first network block includes the first convolutional layer, the second convolutional layer
With third convolutional layer, the method also includes:
For each first network block at least one described first network block,
Dimensionality reduction convolution is carried out to second group of characteristic pattern via first convolutional layer；
Depth point is carried out to the output characteristic pattern of first convolutional layer via second convolutional layer and the third convolutional layer
From convolution, to obtain the output characteristic pattern of the alternate path.
3. method according to claim 2, wherein second convolutional layer is used for the output feature to first convolutional layer
Figure executes the convolution that step-length is 1, and first access is that the jump that the channel is cut between layer and the channel articulamentum connects
Connect road.
4. method according to claim 2, wherein second convolutional layer is used for the output feature to first convolutional layer
Figure executes the convolution that step-length is 2, and first access includes pond layer, the method also includes:
The pond that step-length is 2 is executed to first group of characteristic pattern via the pond layer, to obtain the defeated of first access
Characteristic pattern out.
5. such as the described in any item methods of Claims 1-4, wherein at least one described network block include at least one second
Network block, the alternate path of the second network block of each of at least one described second network block include residual error structure, institute
Stating residual error structure includes residual error branch and jump connection branch, and the residual error branch includes at least one described convolutional layer, described
Method further include:
For the second network block of each of at least one second network block,
Convolution is carried out to second group of characteristic pattern via the residual error branch；
Second group of characteristic pattern of the output characteristic pattern of the residual error branch and the jump connection branch output is subjected to member
Element is added, to obtain the output characteristic pattern of the alternate path.
6. method as claimed in claim 5, wherein the residual error branch includes Volume Four lamination, the 5th convolutional layer and the 6th
Convolutional layer, wherein the method also includes:
For the second network block of each of at least one second network block,
Dimensionality reduction convolution is carried out to second group of characteristic pattern via the Volume Four lamination；
Depth point is carried out to the output characteristic pattern of the Volume Four lamination via the 5th convolutional layer and the 6th convolutional layer
From convolution, to obtain the output characteristic pattern of the residual error branch；
Wherein, the port number of the output characteristic pattern of the residual error branch is equal with the port number of second group of characteristic pattern.
7. method as claimed in claim 6, wherein the 5th convolutional layer is used for the output feature to the Volume Four lamination
Figure executes the convolution that step-length is 1, and first access is that the jump connection that the channel cuts layer to the channel articulamentum is logical
Road.
8. such as the described in any item methods of Claims 1-4, wherein the neural network includes sequentially connected initial convolution
Layer, maximum pond layer, first network section, the second network segment, third network segment, global pool layer and full articulamentum,
Wherein, the first network section includes four network blocks at least one described network block, the second network segment packet
Eight network blocks at least one described network block are included, the third network segment includes four at least one described network block
A network block.
9. such as the described in any item methods of Claims 1-4, wherein the port number and described second of first group of characteristic pattern
The port number of group characteristic pattern is equal.
10. a kind of characteristic pattern processing unit for neural network, the neural network include at least one network block, it is described extremely
Each network block in a few network block includes that channel cuts layer, the first access, alternate path, channel articulamentum and channel weight
Layer is arranged, the alternate path includes at least one convolutional layer, and described device includes:
Channel cuts module, for cutting layer via the channel for each network block at least one described network block
The input feature vector figure of the network block is received, and the channel of the input feature vector figure is cut to two parts, it is special to obtain first group
Sign figure and second group of characteristic pattern；
Input module, for for each network block at least one described network block, first group of characteristic pattern to be inputted
First access, and second group of characteristic pattern is inputted into the alternate path；
Channel link block, for for each network block at least one described network block, via the channel articulamentum
The output characteristic pattern of first access is carried out channel with the output characteristic pattern of the alternate path to connect, after being connected
Characteristic pattern；
Channel reordering module, for resetting layer via the channel for each network block at least one described network block
Channel rearrangement is carried out to the characteristic pattern after the connection, to obtain the output characteristic pattern of the network block.
11. a kind of characteristic pattern processing system for neural network, including processor and memory, wherein in the memory
Computer program instructions are stored with, for executing such as claim 1 when the computer program instructions are run by the processor
To 9 described in any item characteristic pattern processing methods for neural network.
12. a kind of storage medium stores program instruction on said storage, described program instruction is at runtime for holding
The row characteristic pattern processing method as described in any one of claim 1 to 9 for neural network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810932282.7A CN109299722A (en) | 2018-08-16 | 2018-08-16 | Characteristic pattern processing method, device and system and storage medium for neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810932282.7A CN109299722A (en) | 2018-08-16 | 2018-08-16 | Characteristic pattern processing method, device and system and storage medium for neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109299722A true CN109299722A (en) | 2019-02-01 |
Family
ID=65165125
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810932282.7A Pending CN109299722A (en) | 2018-08-16 | 2018-08-16 | Characteristic pattern processing method, device and system and storage medium for neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109299722A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110533161A (en) * | 2019-07-24 | 2019-12-03 | 特斯联（北京）科技有限公司 | A kind of characteristic pattern processing method based on layering group convolutional neural networks |
| CN110543900A (en) * | 2019-08-21 | 2019-12-06 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN110910434A (en) * | 2019-11-05 | 2020-03-24 | 东南大学 | Method for realizing deep learning parallax estimation algorithm based on FPGA (field programmable Gate array) high energy efficiency |
| CN111784555A (en) * | 2020-06-16 | 2020-10-16 | 杭州海康威视数字技术股份有限公司 | Image processing method, device and equipment |
| CN112116032A (en) * | 2019-06-21 | 2020-12-22 | 富士通株式会社 | Object detection device and method and terminal equipment |
| CN113011384A (en) * | 2021-04-12 | 2021-06-22 | 重庆邮电大学 | Anchor-frame-free target detection method based on lightweight convolution |
| WO2022183345A1 (en) * | 2021-03-01 | 2022-09-09 | 浙江大学 | Encoding method, decoding method, encoder, decoder, and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107247949A (en) * | 2017-08-02 | 2017-10-13 | 北京智慧眼科技股份有限公司 | Face identification method, device and electronic equipment based on deep learning |
| CN108154192A (en) * | 2018-01-12 | 2018-06-12 | 西安电子科技大学 | High Resolution SAR terrain classification method based on multiple dimensioned convolution and Fusion Features |
| CN108288075A (en) * | 2018-02-02 | 2018-07-17 | 沈阳工业大学 | A kind of lightweight small target detecting method improving SSD |
| CN108401154A (en) * | 2018-05-25 | 2018-08-14 | 同济大学 | A kind of image exposure degree reference-free quality evaluation method |
-
2018
- 2018-08-16 CN CN201810932282.7A patent/CN109299722A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107247949A (en) * | 2017-08-02 | 2017-10-13 | 北京智慧眼科技股份有限公司 | Face identification method, device and electronic equipment based on deep learning |
| CN108154192A (en) * | 2018-01-12 | 2018-06-12 | 西安电子科技大学 | High Resolution SAR terrain classification method based on multiple dimensioned convolution and Fusion Features |
| CN108288075A (en) * | 2018-02-02 | 2018-07-17 | 沈阳工业大学 | A kind of lightweight small target detecting method improving SSD |
| CN108401154A (en) * | 2018-05-25 | 2018-08-14 | 同济大学 | A kind of image exposure degree reference-free quality evaluation method |
Non-Patent Citations (4)
| Title |
|---|
| NINGNING MA 等: "ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design", 《ARXIV:1807.11164V1》 * |
| XIANGYU ZHANG 等: "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices", 《ARXIV:1707.01083V2》 * |
| 吴天舒 等: "基于改进 SSD 的轻量化小目标检测算法", 《红外与激光工程》 * |
| 王磊 等: "面向嵌入式应用的深度神经网络模型压缩技术综述", 《北京交通大学学报》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112116032A (en) * | 2019-06-21 | 2020-12-22 | 富士通株式会社 | Object detection device and method and terminal equipment |
| JP2021002333A (en) * | 2019-06-21 | 2021-01-07 | 富士通株式会社 | Object detection device, object detection method, and terminal equipment |
| JP7428075B2 (en) | 2019-06-21 | 2024-02-06 | 富士通株式会社 | Object detection device, object detection method and terminal equipment |
| CN110533161A (en) * | 2019-07-24 | 2019-12-03 | 特斯联（北京）科技有限公司 | A kind of characteristic pattern processing method based on layering group convolutional neural networks |
| CN110533161B (en) * | 2019-07-24 | 2022-05-20 | 特斯联（北京）科技有限公司 | Feature map processing method based on hierarchical group convolution neural network |
| CN110543900A (en) * | 2019-08-21 | 2019-12-06 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN110910434A (en) * | 2019-11-05 | 2020-03-24 | 东南大学 | Method for realizing deep learning parallax estimation algorithm based on FPGA (field programmable Gate array) high energy efficiency |
| CN110910434B (en) * | 2019-11-05 | 2023-05-12 | 东南大学 | Method for realizing deep learning parallax estimation algorithm based on FPGA (field programmable Gate array) high energy efficiency |
| CN111784555B (en) * | 2020-06-16 | 2023-08-25 | 杭州海康威视数字技术股份有限公司 | Image processing method, device and equipment |
| CN111784555A (en) * | 2020-06-16 | 2020-10-16 | 杭州海康威视数字技术股份有限公司 | Image processing method, device and equipment |
| WO2022183345A1 (en) * | 2021-03-01 | 2022-09-09 | 浙江大学 | Encoding method, decoding method, encoder, decoder, and storage medium |
| CN113011384A (en) * | 2021-04-12 | 2021-06-22 | 重庆邮电大学 | Anchor-frame-free target detection method based on lightweight convolution |
| CN113011384B (en) * | 2021-04-12 | 2022-11-25 | 重庆邮电大学 | Anchor-frame-free target detection method based on lightweight convolution |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109299722A (en) | Characteristic pattern processing method, device and system and storage medium for neural network | |
| CN108595585B (en) | Sample data classification method, model training method, electronic equipment and storage medium | |
| KR102224510B1 (en) | Systems and methods for data management | |
| CN108875486A (en) | Recongnition of objects method, apparatus, system and computer-readable medium | |
| CN108875523A (en) | Human synovial point detecting method, device, system and storage medium | |
| Rybakken et al. | Decoding of neural data using cohomological feature extraction | |
| CN108875722A (en) | Character recognition and identification model training method, device and system and storage medium | |
| CN108875932A (en) | Image-recognizing method, device and system and storage medium | |
| McClure et al. | Representational distance learning for deep neural networks | |
| CN108875521A (en) | Method for detecting human face, device, system and storage medium | |
| CN108875732A (en) | Model training and example dividing method, device and system and storage medium | |
| CN108780519A (en) | Structure learning in convolutional neural networks | |
| CN108764133A (en) | Image-recognizing method, apparatus and system | |
| CN108875487A (en) | Pedestrian is identified the training of network again and is identified again based on its pedestrian | |
| CN108664897A (en) | Bank slip recognition method, apparatus and storage medium | |
| CN109155006A (en) | The audio analysis based on frequency is carried out using neural network | |
| CN106529511A (en) | Image structuring method and device | |
| CN109117953A (en) | Network parameter training method and system, server, client and storage medium | |
| CN106447721A (en) | Image shadow detection method and device | |
| CN108876804A (en) | It scratches as model training and image are scratched as methods, devices and systems and storage medium | |
| CN108876793A (en) | Semantic segmentation methods, devices and systems and storage medium | |
| CN108734052A (en) | character detecting method, device and system | |
| CN108108662A (en) | Deep neural network identification model and recognition methods | |
| CN108875492A (en) | Face datection and crucial independent positioning method, device, system and storage medium | |
| CN108009466A (en) | Pedestrian detection method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190201 |
|
| RJ01 | Rejection of invention patent application after publication |