CN108665475A - Neural metwork training, image processing method, device, storage medium and electronic equipment - Google Patents
Neural metwork training, image processing method, device, storage medium and electronic equipment Download PDFInfo
- Publication number
- CN108665475A CN108665475A CN201810463674.3A CN201810463674A CN108665475A CN 108665475 A CN108665475 A CN 108665475A CN 201810463674 A CN201810463674 A CN 201810463674A CN 108665475 A CN108665475 A CN 108665475A
- Authority
- CN
- China
- Prior art keywords
- segmentation result
- convolution
- image
- training
- result
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
A kind of neural metwork training of offer of the embodiment of the present invention, image processing method, device, computer readable storage medium and electronic equipment.Neural network training method includes：The first process of convolution is carried out to training sample image by nerve network system, obtains the first segmentation result of the preceding background segment of the training sample image；And the second process of convolution is carried out to first segmentation result, obtain the second segmentation result of the preceding background segment of the training sample image；According to the difference between the difference marked between segmentation result and first segmentation result and the mark segmentation result and second segmentation result of the preceding background segment of the training sample image, the training nerve network system.Using the technical solution of the embodiment of the present invention, background segment accuracy before image can be effectively improved, and ensure higher processing speed, so as to effectively promote the effect for carrying out monocular virtualization processing.
Description
Technical field
The present embodiments relate to technical field of computer vision more particularly to a kind of neural network training method, device,
Computer readable storage medium and electronic equipment and a kind of image processing method, device, computer readable storage medium and electricity
Sub- equipment.
Background technology
The background blurring of image can be such that shooting main body clearly shows, deep to be liked by shutterbugs.Due to image
It blurs effect and mainly utilizes optical imaging concept, realized using big lens aperture on hardware, therefore, the virtualization of image
Function is mainly integrated on the video camera of the profession such as slr camera, for smart mobile phone, tablet computer equal thickness it is limited and can only
The mobile terminal device of small aperture camera lens is installed, for user when taking pictures using mobile terminal device, can be only generated does not have void
Change effect or only with faint virtualization effect image.
Currently, enabling to the mobile end with single camera lens using single-lens background blurring technology (monocular virtualization technology)
End equipment simulates the background blurring effect of taking pictures of slr camera.In monocular blurs technology application process, need to carry out image
Preceding background segment processing, extracts target prospect region or background area.But before existing preceding background segment treatment technology exists
The problem that background segment accuracy is relatively low or processing speed is relatively low, in the case where processing speed is limited, preceding background segment is accurate
Exactness can be also limited, and cause final monocular to blur effect poor.
Invention content
The purpose of the embodiment of the present invention is, provides a kind of neural metwork training technology and a kind of image processing techniques.
According to a first aspect of the embodiments of the present invention, a kind of neural network training method is provided, including：Pass through neural network
System carries out the first process of convolution to training sample image, obtains the first segmentation of the preceding background segment of the training sample image
As a result；And the second process of convolution is carried out to first segmentation result, obtain the preceding background segment of the training sample image
The second segmentation result；It is tied according to the mark segmentation result of the preceding background segment of the training sample image and first segmentation
The difference between difference and the mark segmentation result and second segmentation result between fruit, the training nerve net
Network system.
Optionally, the mark segmentation result of the preceding background segment according to the training sample image with described first point
Cut the difference between difference and the mark segmentation result and second segmentation result between result, the training god
Through network system, including：The first variance data between the mark segmentation result and first segmentation result is obtained, and
The second variance data between the mark segmentation result and second segmentation result；According to first variance data and institute
State the second variance data and the adjustment nerve network system network parameter.
Optionally, the neural network includes system the first convolution sub-network and the second convolution sub-network；Pass through described
One convolution sub-network carries out first process of convolution to the training sample image, obtains first segmentation result；Pass through
The second convolution sub-network carries out second process of convolution to first segmentation result, obtains the second segmentation knot
Fruit.
Optionally, the first convolution sub-network includes multiple convolutional layers positioned at different depth；It is described to described first
Segmentation result carries out the second process of convolution, including：Obtain the shallow-layer of the convolutional layer in the multiple convolutional layer positioned at the first depth
The deep layer for exporting result and/or the convolutional layer positioned at the second depth exports result；Wherein, first depth is less than described second
Depth；First segmentation result and shallow-layer output result and/or deep layer output result are merged, obtain the
One amalgamation result；Second process of convolution is carried out to first amalgamation result.
Optionally, the nerve network system includes at least two the second convolution sub-networks positioned at different depth；Pass through
In at least two second convolution sub-networks positioned at minimum-depth the second convolution sub-network to first segmentation result into
The second process of convolution of row obtains second segmentation result；By the second convolution sub-network positioned at third depth to being located at
The second segmentation result that second convolution sub-network of four depth obtains carries out the second new process of convolution, obtains the second new segmentation
As a result；Wherein, the third depth is more than the 4th depth；The mark of the foreground segmentation according to the training sample image
Difference and the mark segmentation result between note segmentation result and first segmentation result and second segmentation result
Between difference, the training nerve network system, including：Divided according to the mark of the foreground segmentation of the training sample image
As a result the difference between first segmentation result, the difference between the mark segmentation result and second segmentation result
Difference between the different and described mark segmentation result and the second new segmentation result, the training nerve network system.
Optionally, the second convolution sub-network by positioned at third depth is to the second convolution positioned at the 4th depth
The second segmentation result that network obtains carries out the second new process of convolution, including：It will be positioned at the second convolution subnet of the 4th depth
The second segmentation result that network obtains exports result with the shallow-layer and/or deep layer output result merges, and obtains second
Amalgamation result；The new volume Two is carried out to second amalgamation result by the second convolution sub-network positioned at third depth
Product processing.
Optionally, the mark segmentation result of the foreground segmentation according to the training sample image is divided with described first
As a result the difference between, difference and the mark between the mark segmentation result and second segmentation result are divided
As a result the difference between the second new segmentation result, the training nerve network system, including：Obtain the mark point
Cut the first variance data between result and first segmentation result, the mark segmentation result and second segmentation result
Between the second variance data and the mark segmentation result and the second new segmentation result between third difference number
According to；According to first variance data, second variance data and the third variance data and the adjustment nerve
The network parameter of network system.
According to a second aspect of the embodiments of the present invention, a kind of image processing method is provided, including：Pending image is inputted
Nerve network system, background segment processing before being carried out to the pending image by the nerve network system；It obtains and carries out
The background parts of pending image after the preceding background segment processing carry out segmentation virtualization processing to the background parts.
Optionally, described that preceding background segment processing, packet are carried out to the pending image by the nerve network system
It includes：The first process of convolution is carried out to pending image by nerve network system, obtains the preceding background point of the pending image
The first segmentation result cut；Second process of convolution is carried out to first segmentation result, obtains the preceding back of the body of the pending image
Second segmentation result of scape segmentation；Background segment processing before being carried out to the pending image according to second segmentation result,
To carry out handling relevant processing with preceding background segment to the pending image.
Optionally, the nerve network system is using the neural metwork training as described in any one of claim 1 to 7
The nerve network system that method training obtains.
Optionally, it is handled to the pending preceding background segment of image progress by the nerve network system described
Afterwards, the method further includes：To carry out the pending image after the preceding background segment processing carry out Steerable filter processing and/
Or alpha matting processing.
Optionally, described that segmentation virtualization processing is carried out to the background parts, including：
According to the distance of designated edge of each pixel to the pending image in the background parts, to described every
A pixel carries out Fuzzy Processing.
According to a third aspect of the embodiments of the present invention, a kind of neural metwork training device is provided, including：First segmentation mould
Block obtains the training sample image for carrying out the first process of convolution to training sample image by nerve network system
First segmentation result of preceding background segment；And the second process of convolution is carried out to first segmentation result, obtain the training
Second segmentation result of the preceding background segment of sample image；Training module, for the preceding background according to the training sample image
Difference between the mark segmentation result of segmentation and first segmentation result and the mark segmentation result and described second
Difference between segmentation result, the training nerve network system.
Optionally, the training module is used to obtain the between the mark segmentation result and first segmentation result
The second variance data between one variance data and the mark segmentation result and second segmentation result；According to described
First variance data and second variance data and the adjustment nerve network system network parameter.
Optionally, the neural network includes system the first convolution sub-network and the second convolution sub-network；The segmentation mould
Block includes：First cutting unit, for carrying out described first to the training sample image by the first convolution sub-network
Process of convolution obtains first segmentation result；Second cutting unit, for by the second convolution sub-network to described the
One segmentation result carries out second process of convolution, obtains second segmentation result.
Optionally, the first convolution sub-network includes multiple convolutional layers positioned at different depth；The segmentation module is also
Including：First combining unit, the shallow-layer for obtaining the convolutional layer in the multiple convolutional layer positioned at the first depth export result
And/or the deep layer of the convolutional layer positioned at the second depth exports result；Wherein, first depth is less than second depth；With
And merge first segmentation result and shallow-layer output result and/or deep layer output result, obtain first
Amalgamation result；Second cutting unit is used to carry out the second process of convolution to first amalgamation result.
Optionally, the nerve network system includes at least two the second convolution sub-networks positioned at different depth；It is described
Second cutting unit is used for the second convolution sub-network by being located at minimum-depth at least two second convolution sub-networks
Second process of convolution is carried out to first segmentation result, obtains second segmentation result；The first segmentation module is also wrapped
Third cutting unit is included, for sub to the second convolution positioned at the 4th depth by the second convolution sub-network positioned at third depth
The second segmentation result that network obtains carries out the second new process of convolution, obtains the second new segmentation result；Wherein, the third
Depth is more than the 4th depth；The training module is used to be divided according to the mark of the foreground segmentation of the training sample image
As a result the difference between first segmentation result, the difference between the mark segmentation result and second segmentation result
Difference between the different and described mark segmentation result and the second new segmentation result, the training nerve network system.
Optionally, the first segmentation module further includes the second combining unit, for that will be located at the volume Two of the 4th depth
The second segmentation result that product sub-network obtains exports result with the shallow-layer and/or deep layer output result merges, and obtains
Obtain the second amalgamation result；The third cutting unit is used for through the second convolution sub-network positioned at third depth to stating the second conjunction
And result carries out the second new process of convolution.
Optionally, the training module is used to obtain the between the mark segmentation result and first segmentation result
One variance data, the second variance data between the mark segmentation result and second segmentation result and the mark
Third variance data between segmentation result and the second new segmentation result；According to first variance data, described
The network parameter with the adjustment nerve network system of two variance datas and the third variance data.
According to a fourth aspect of the embodiments of the present invention, a kind of image processing apparatus is provided, including：Second segmentation module,
For pending image to be inputted nerve network system, the back of the body before being carried out to the pending image by the nerve network system
Scape dividing processing；Blurring module, the background parts for obtaining the pending image after carrying out the preceding background segment processing,
Segmentation virtualization processing is carried out to the background parts.
Optionally, the second segmentation module includes：Third cutting unit, for passing through nerve network system to pending
Image carries out the first process of convolution, obtains the first segmentation result of the preceding background segment of the pending image；4th segmentation is single
Member obtains the of the preceding background segment of the pending image for carrying out the second process of convolution to first segmentation result
Two segmentation results；5th cutting unit, for background point before being carried out to the pending image according to second segmentation result
Processing is cut, to carry out handling relevant processing with preceding background segment to the pending image.
Optionally, the nerve network system is using any neural metwork training device instruction provided in an embodiment of the present invention
Practice the nerve network system obtained.
Optionally, described device further includes：Processing module, for carrying out waiting locating after the preceding background segment is handled
It manages image and carries out Steerable filter processing and/or alpha matting processing.
Optionally, the blurring module is used for according to each pixel in the background parts to the pending image
The distance of designated edge carries out Fuzzy Processing to each pixel.
According to a fifth aspect of the embodiments of the present invention, a kind of computer program is provided comprising have computer program to refer to
It enables, for realizing any neural network training method pair provided in an embodiment of the present invention when described program instruction is executed by processor
The step of answering.
According to a sixth aspect of the embodiments of the present invention, a kind of computer program is provided comprising have computer program to refer to
It enables, described program instruction is corresponding for realizing any image processing method provided in an embodiment of the present invention when being executed by processor
Step.
According to a seventh aspect of the embodiments of the present invention, a kind of computer readable storage medium is provided, meter is stored thereon with
Calculation machine program instruction, for realizing any neural network provided in an embodiment of the present invention when described program instruction is executed by processor
Step corresponding to training method.
According to a eighth aspect of the embodiments of the present invention, a kind of computer readable storage medium is provided, meter is stored thereon with
Calculation machine program instruction, for realizing any image provided in an embodiment of the present invention processing when described program instruction is executed by processor
Step corresponding to method.
According to a ninth aspect of the embodiments of the present invention, a kind of electronic equipment is provided, including：Processor, memory, communication
Element and communication bus, the processor, the memory and the communication device are completed mutual by the communication bus
Communication；The memory makes the processor execute this hair for storing an at least executable instruction, the executable instruction
Step corresponding to any neural network training method that bright embodiment provides.
According to a tenth aspect of the embodiments of the present invention, a kind of electronic equipment is provided, including：Processor, memory, communication
Element and communication bus, the processor, the memory and the communication device are completed mutual by the communication bus
Communication；The memory makes the processor execute this hair for storing an at least executable instruction, the executable instruction
Step corresponding to any image processing method that bright embodiment provides.
Image procossing scheme according to the ... of the embodiment of the present invention, by being carried out to training sample image in training neural network
Preliminary process of convolution, and the primary segmentation result to obtaining carries out further process of convolution, obtains training sample image
The Optimized Segmentation of preceding background segment is as a result, to according between training sample image and primary segmentation result and Optimized Segmentation result
Difference train nerve network system, to carry out secondary process of convolution to pending image by nerve network system, preliminary
Segmentation result is advanced optimized on the basis of segmentation result, background segment accuracy before can effectively improving, and ensure higher
Processing speed；Monocular virtualization processing is carried out on this basis, can effectively promote virtualization effect.
Description of the drawings
Fig. 1 is the flow chart for the neural network training method for showing according to embodiments of the present invention one；
Fig. 2 is the flow chart for the neural network training method for showing according to embodiments of the present invention two；
Fig. 3 is the schematic diagram for the neural metwork training system for showing that according to embodiments of the present invention two provide；
Fig. 4 shows the flow chart of according to embodiments of the present invention three image processing method；
Fig. 5 is the first structure block diagram for the neural metwork training device for showing according to embodiments of the present invention four；
Fig. 6 is the second structure diagram of the neural metwork training device for showing according to embodiments of the present invention four；
Fig. 7 is the structure diagram for the image processing apparatus for showing according to embodiments of the present invention five；
Fig. 8 is the structural schematic diagram of according to embodiments of the present invention six a kind of electronic equipment；
Fig. 9 is the structural schematic diagram of according to embodiments of the present invention seven a kind of electronic equipment.
Specific implementation mode
(identical label indicates identical element in several attached drawings) and embodiment below in conjunction with the accompanying drawings, implement the present invention
The specific implementation mode of example is described in further detail.Following embodiment is not limited to the present invention for illustrating the present invention
Range.
It will be understood by those skilled in the art that the terms such as " first ", " second " in the embodiment of the present invention are only used for distinguishing
Different step, equipment or module etc. neither represent any particular technology meaning, also do not indicate that the inevitable logic between them is suitable
Sequence.
Embodiment one
Fig. 1 is the flow chart for the neural network training method for showing according to embodiments of the present invention one.
Referring to Fig.1, in step S110, the first process of convolution is carried out to training sample image by nerve network system, is obtained
Take the first segmentation result of the preceding background segment of training sample image；And the second process of convolution is carried out to the first segmentation result,
Obtain the second segmentation result of the preceding background segment of training sample image.
In embodiments of the present invention, the nerve net of background segment processing before training sample image carries out image for training
Network system.Training sample image can be the image that arbitrary camera is shot under arbitrary scene.Optionally, neural network
System is deep neural network system.The setting of concrete structure can be by those skilled in the art according to reality in nerve network system
Border demand is suitably set, such as the number of plies of convolutional layer, the size of convolution kernel in depth convolutional neural networks, the embodiment of the present invention
This is not restricted.
According to an illustrative embodiment of the invention, after training sample image being inputted nerve network system, pass through nerve
Network system carries out the first process of convolution to training sample image, obtains the first segmentation result.Here, the first process of convolution can be with
Including multiple convolution operations, for example, multiple concatenated convolutional layers carry out convolution operation respectively forms the first process of convolution, for pair
Training sample image carries out feature extraction, feature learning etc..First segmentation result is used to indicate foreground portion in training sample image
(foreground area) and/or background parts (background area) is divided to be used for for example, the first segmentation result is specifically as follows a matrix
Each pixel belongs to the probability of foreground part or background parts in mark training sample image.It can according to the first segmentation result
The dividing processing (preceding background segment) of foreground part and background area is carried out to training sample image.
After obtaining the first segmentation result, carried out at the second convolution by the first segmentation result of nerve network system pair
Reason.Here, the second process of convolution can include identical operation with the first process of convolution, and the second segmentation result can be with first point
Cut result expression form having the same.Second process of convolution is used for the process of convolution to the progress of the first segmentation result again, right
Pending image carries out deeper feature learning and extraction, further to obtain higher second segmentation result of accuracy.
That is, the first process of convolution is to handle the preliminary preceding background segment that training sample image carries out, first point
Cut the primary segmentation result that result is；Second process of convolution is to advanced optimize processing to what primary segmentation result carried out,
The Optimized Segmentation result that second segmentation result is.
In step S120, according to the mark segmentation result of the preceding background segment of training sample image and the first segmentation result it
Between difference, and mark segmentation result and the second segmentation result between difference, training nerve network system.
Optionally, obtain mark segmentation result and the first segmentation result between the first variance data, and with second point
The second variance data between result is cut, to adjust the net of nerve network system according to the first variance data and the second variance data
Network parameter.For example, according to preset loss function or departure function, calculate between mark segmentation result and the first segmentation result
Penalty values either deviation and the penalty values between the second segmentation result or deviation, and according to being calculated
The network parameter of penalty values or deviation adjustment nerve network system.
In practical applications, above-mentioned neural network can be repeated based on multiple identical or different training sample images
Training method is adjusted the network parameter of nerve network system, until nerve network system is preferably restrained.
The nerve network system that training is completed can be used for the processing of background segment before image carries out, and pass through nerve network system
First process of convolution is carried out to pending image and obtains primary segmentation as a result, and carrying out further second to primary segmentation result
Process of convolution, obtained Optimized Segmentation result have the higher accuracy of separation；Also, the back of the body before being carried out by nerve network system
Scape dividing processing, the concrete operation step that can also include by controlling the second process of convolution, to ensure higher processing speed；
To realize the preceding background segment processing of pinpoint accuracy in the case where processing speed is limited.
According to embodiments of the present invention one neural network training method, by carrying out preliminary convolution to training sample image
Processing, obtains the primary segmentation result of the preceding background segment of training sample image；And by primary segmentation result into advance one
The process of convolution of step obtains the Optimized Segmentation result of the preceding background segment of training sample image；To according to training sample image
Difference between primary segmentation result and Optimized Segmentation result trains nerve network system.Using the neural network of the present embodiment
The nerve network system that training method is trained, by carrying out secondary process of convolution to training sample image, in primary segmentation
As a result segmentation result is advanced optimized on the basis of, background segment accuracy before can effectively improving, and ensure higher processing
Speed.
The neural network training method of the present embodiment can have corresponding image and data processing energy by any suitable
The equipment of power executes, including but not limited to：Such as the terminal devices such as computer, and integrated computer journey on the terminal device
Sequence, processor etc..
Embodiment two
Fig. 2 is the flow chart for the neural network training method for showing according to embodiments of the present invention two.
Training sample image is carried out by the first convolution sub-network of nerve network system in step S210 with reference to Fig. 2
First process of convolution obtains the first segmentation result of the preceding background segment of training sample image.
In the embodiment of the present invention, nerve network system is used to carry out preceding background segment to image, and nerve network system includes
First convolution sub-network obtains the first segmentation result for carrying out the first process of convolution to image.Here, the first convolution subnet
Network may include that multiple convolutional layers, the first process of convolution include the convolution operation that multiple convolutional layers execute respectively.In the first convolution
In network, the design parameters such as the number of convolutional layer, the size of convolution kernel (filter), port number and down-sampling position are set
It sets, can be suitably set according to actual demand by those skilled in the art, the embodiment of the present invention is not restricted this.
In a kind of optional embodiment, multiple convolutional layers that the first convolution sub-network includes are located at nerve network system
Different depth, as shown in figure 3, being provided with multiple concatenated Conv (Convolution, convolutional layer) in nerve network system.It will
After training sample image inputs nervous system, convolution operation is executed respectively by multiple convolutional layers, wherein multiple convolutional layers are held
Capable convolution operation is used to extract the feature of different levels of abstraction in training sample image.
In step S220, the shallow-layer output result and/or deep layer that obtain in the first convolution sub-network export as a result, by first
Segmentation result exports result with shallow-layer and/or deep layer output result merges, and obtains the first amalgamation result.
Wherein, shallow-layer output result is the shallow-layer convolution positioned at the first depth in multiple convolutional layers of the first convolution sub-network
The output of layer is as a result, deep layer output result is the output result of the deep layer convolutional layer positioned at the second depth.Here, the first depth is small
In the second depth, that is, shallow-layer convolutional layer can be any of the convolutional layer that depth is smaller or shallower in multiple convolutional layers,
Deep layer convolutional layer can be that depth is larger or any of deeper convolutional layer.
According to an illustrative embodiment of the invention, to the first segmentation result of acquisition carry out the second process of convolution it
Before, the first segmentation result and shallow-layer output result and/or deep layer output result are merged into processing, obtain the first merging knot
Fruit.Wherein, shallow-layer segmentation result is the output of the shallow-layer convolutional layer in the first convolution sub-network as a result, containing shallow-layer convolutional layer
The feature such as image texture, material extracted from training sample image；Deep layer segmentation result is in the first convolution sub-network
Deep layer convolutional layer output as a result, containing such as image subject details that deep layer convolutional layer is extracted from training sample image
Deng more abstract, advanced information.Preferably, by shallow-layer export result and deep layer output result simultaneously with the first segmentation result into
Row merging treatment so that the first amalgamation result includes the detailed information of training sample image and whole semantic information.
Merging treatment mode is shown in Fig. 3, is to tie the first segmentation result and shallow-layer output result and/or deep layer output
Fruit inputs the second convolution sub-network Refine Conv together, and merging treatment is executed by the second convolution sub-network.In practical application
In, the second convolution subnet can also be inputted after executing merging treatment and obtaining the first amalgamation result, then by the first amalgamation result
Network carries out the second process of convolution.
In step S230, carried out by the first amalgamation result of at least one second convolution sub-network pair of nerve network system
Second process of convolution obtains the second segmentation result of the preceding background segment of training sample image.
In the embodiment of the present invention, the first convolution sub-network of nerve network system is the convolutional network of first stage, is used for
The first process of convolution of preliminary preceding background segment is carried out to training sample image；At least one second convolution sub-network is second-order
The convolutional network of section, the second process of convolution for background segment accuracy before being optimized to the first segmentation result.
Nerve network system shown in Fig. 3 includes a second convolution sub-network, passes through the second convolution sub-network pair
First amalgamation result carries out the second process of convolution, obtains the second segmentation result.Wherein, the first amalgamation result includes training sample figure
The detailed information of picture and whole semantic information, the second process of convolution is carried out based on the first amalgamation result, can further increase the
The accuracy of two segmentation results.
Optionally, when nerve network system includes at least two the second convolution sub-networks, pass through at least two volume Twos
Second the first segmentation result of convolution sub-network pair for being located at minimum-depth in product sub-network carries out the second process of convolution, obtains second
Segmentation result.Here, the second convolution sub-network for being located at minimum-depth can be the second convolution sub-network shown in Fig. 3, use
In the second process of convolution to the progress of the first segmentation result for the first time.
And the second convolution sub-network positioned at the 4th depth is obtained by the second convolution sub-network positioned at third depth
The second segmentation result taken carries out the second new process of convolution, obtains the second new segmentation result；Wherein, third depth is more than the
Four depth.That is, being obtained to the second convolution sub-network positioned at smaller depth by the second convolution sub-network positioned at larger depth
The second segmentation result taken, carries out the second process of convolution again, to obtain higher the second new segmentation result of accuracy.This
In, the second segmentation result for being located at the second convolution sub-network output of depth capacity is preceding background segment result finally.
Optionally, the shallow-layer output result and/or deep layer output in the first convolution sub-network are obtained as a result, the 4th will be located at
The second segmentation result that second convolution sub-network of depth obtains is closed with the first output result and/or the second output result
And obtain the second amalgamation result；By carrying out new the positioned at second the second amalgamation result of convolution sub-network pair of third depth
Two process of convolution realize the second process of convolution for being executed to the second segmentation result and optimizing preceding background segment into accuracy.Here, it obtains
The deep layer output result taken can also be second of the second convolution sub-network output positioned at smaller depth (being less than the 4th depth)
Segmentation result.
Illustrate herein, the quantity of the second convolution sub-network can be according to actual processing speed in the processing of preceding background segment
Demand is set.For example, if the processing speed of demand is higher, 2 or 3 the second convolution sub-networks can be set, to not
Increase excessive processing time, improves processing speed.
That is, carrying out primary segmentation processing to training sample image by the first convolution sub-network, obtain preliminary
After segmentation result (the first segmentation result), primary segmentation result can be optimized by the way that a second convolution sub-network is arranged
Processing, to improve the accuracy of separation；And by the way that multiple second convolution sub-networks are arranged, to the second segmentation result after optimization into
The multiple optimization processing of row, to further increase the accuracy of separation.
In step S240, obtain the preceding background segment of training sample image mark segmentation result and the first segmentation result it
Between the first variance data, and the second variance data between the second segmentation result, according to the first variance data and second
Variance data and adjustment nerve network system network parameter.
Here, according to the difference between mark segmentation result and the first segmentation result and the second segmentation result, training god
When through network system, by the first variance data and the second variance data and, the final segmentation as nerve network system acquisition
As a result the metric of the difference between object segmentation result, and then joined according to the network of metric adjustment nerve network system
Number.
Wherein, if nerve network system includes multiple second convolution sub-networks, according to mark segmentation result and first
Difference between segmentation result, and the difference between the second segmentation result of each second convolution sub-network acquisition, training god
Through network system.Optionally, the first variance data between mark segmentation result and the first segmentation result is obtained, with the second segmentation
As a result the second variance data between, and each third variance data between each the second new segmentation result；According to first
Variance data, the second variance data and each third variance data and adjustment nerve network system network parameter.
The neural network training method that the embodiment of the present invention is executed based on a large amount of training sample image, according to obtaining every time
Metric adjustment nerve network system network parameter, until nerve network system preferably restrains, that is, until obtaining
Metric is reduced to preset discrepancy threshold or is not reducing.
According to embodiments of the present invention two neural network training method, by carrying out preliminary convolution to training sample image
Processing, obtains the primary segmentation result of the preceding background segment of training sample image；And by primary segmentation result into advance one
The process of convolution of step obtains the Optimized Segmentation result of the preceding background segment of training sample image；To according to training sample image
Difference between primary segmentation result and Optimized Segmentation result trains nerve network system, improves the neural network that training is completed
The accuracy of background segment before system carries out；And by when optimizing processing to primary segmentation result, it is contemplated that preliminary
Shallow-layer output result and deep layer in deconvolution process export result, that is to say, that consider the details letter of training sample image
Breath and whole semantic information further improve the accuracy of background segment before the nerve network system that training is completed carries out；With
And by carrying out again multiple optimization processing to the segmentation result after optimization, further improve the Optimized Segmentation knot of acquisition
The accuracy of fruit, the accuracy of background segment before the nerve network system to further increase training completion carries out.Using this
The nerve network system that the neural network training method of embodiment is trained, background segment is accurate before can effectively improving
Degree, and ensure higher processing speed.
The neural network training method of the present embodiment can have corresponding image or data processing energy by any suitable
The equipment of power executes, including but not limited to：Such as the terminal device of computer, and integrated computer journey on the terminal device
Sequence, processor etc..
Embodiment three
Fig. 4 is the flow chart for the image processing method for showing according to embodiments of the present invention three.
With reference to Fig. 4, in step S410, pending image is inputted into nerve network system, is treated by nerve network system
Handle background segment processing before image carries out.
In the embodiment of the present invention, nerve network system be used for image carry out before background segment processing, be specifically as follows through
Cross the nerve network system that the neural network training method of the embodiment of the present invention one or embodiment two is trained.
Optionally, the first process of convolution is carried out to pending image by nerve network system, obtains pending image
First segmentation result of preceding background segment；And the second process of convolution is carried out to the first segmentation result, obtain pending image
Second segmentation result of preceding background segment；And background segment processing before being carried out to pending image according to the second segmentation result, with
Pending image is carried out to handle relevant processing with preceding background segment.
Wherein, the first process of convolution, which can be considered, carries out at preliminary convolution pending image by nerve network system
Reason, obtains the primary segmentation result of the preceding background segment of pending image；Second process of convolution can be considered by primary segmentation
As a result the process of convolution advanced optimized obtains the Optimized Segmentation of the preceding background segment of pending image as a result, to carry
The accuracy of background segment before height, and ensure higher processing speed.
Also, it can be obtained when carrying out the second process of convolution to the first segmentation result for carrying out the first process of convolution
The output result of multiple convolutional layer middle-shallow layer convolutional layers and the output of deep layer convolutional layer are as a result, that is, obtain at the first convolution
Shallow-layer output result and deep layer output during reason is as a result, and export the first segmentation result and shallow-layer output result and deep layer
As a result merge, to by obtain amalgamation result carry out the second process of convolution, further increase nerve network system into
The accuracy of background segment before row.
And the second segmentation result can again be carried out for excellent after carrying out the second process of convolution to the first segmentation result
The process of convolution for changing accuracy, further to improve the accuracy of background segment before nerve network system carries out.
In step S420, the background parts of the pending image before carrying out after background segment processing are obtained, to background portion
Divide and carries out segmentation virtualization processing.
For example, being based on preceding background segment handling result, background parts in pending image are determined, it will be each in background parts
The fuzzy dynamics of pixel is set as directly proportional at a distance from pixel to image lower edge, and the fuzzy dynamics based on setting
Fuzzy Processing is carried out to background parts, so that the virtualization degree of the virtualization degree of background parts and foreground area carries out area
Point, improve the validity of background blurring processing.
Optionally, before executing this step, the pending image after background segment processing before progress can be led
To being filtered and/or alpha matting processing.Wherein, foreground part of the Steerable filter processing for being partitioned into image
Smooth noise reduction is carried out with the edge of background parts, improves reservation degree of the segmentation result to edge details.Alpha matting processing
For making the transitional region between foreground part and background parts more natural.
That is, the image processing method of the present embodiment to image carry out before background segment processing after, can be right
Image carries out including but not limited to Steerable filter processing, background segmentation virtualization processing etc. and the relevant processing of preceding background segment.
The image processing method of the present embodiment can be by any suitable with corresponding image and data-handling capacity
Equipment executes, including but not limited to：Terminal device and integrated computer program, processor etc. on the terminal device, for example,
The mobile terminal devices such as smart mobile phone, tablet computer.
Example IV
With reference to Fig. 5, a kind of structure diagram of according to embodiments of the present invention four neural metwork training device is shown.
The neural metwork training device of the embodiment of the present invention includes：First segmentation module 502, for passing through neural network system
System carries out the first process of convolution to training sample image, obtains the first segmentation knot of the preceding background segment of the training sample image
Fruit；And the second process of convolution is carried out to first segmentation result, obtain the preceding background segment of the training sample image
Second segmentation result；Training module 504, for according to the mark segmentation result of the preceding background segment of the training sample image with
The difference between difference and the mark segmentation result and second segmentation result between first segmentation result,
The training nerve network system.
Optionally, the training module 504 is for obtaining between the mark segmentation result and first segmentation result
The first variance data and it is described mark segmentation result and second segmentation result between the second variance data；According to
First variance data and second variance data and the adjustment nerve network system network parameter.
Optionally, the neural network includes system the first convolution sub-network and the second convolution sub-network；Dress shown in Fig. 5
On the basis of setting, the first segmentation module 502 includes：First cutting unit 5021, for passing through the first convolution subnet
Network carries out first process of convolution to the training sample image, obtains first segmentation result；Second cutting unit
5022, for carrying out second process of convolution to first segmentation result by the second convolution sub-network, obtain institute
State the second segmentation result.
Optionally, the first convolution sub-network includes multiple convolutional layers positioned at different depth；The first segmentation mould
Block 502 further includes：First combining unit 5023, for obtaining the shallow of the convolutional layer in the multiple convolutional layer positioned at the first depth
Layer exports result and/or the deep layer of the convolutional layer positioned at the second depth exports result；Wherein, first depth is less than described the
Two depth；And first segmentation result and shallow-layer output result and/or deep layer output result are closed
And obtain the first amalgamation result；Second cutting unit 5022 is used to carry out at the second convolution first amalgamation result
Reason.
Optionally, the nerve network system includes at least two the second convolution sub-networks positioned at different depth；It is described
Second cutting unit 5022 is used for the second convolution by being located at minimum-depth at least two second convolution sub-networks
Network carries out the second process of convolution to first segmentation result, obtains second segmentation result；The first segmentation module
502 further include third cutting unit 5024, for by the second convolution sub-network positioned at third depth to being located at the 4th depth
The second segmentation result for obtaining of the second convolution sub-network carry out the second new process of convolution, obtain the second new segmentation result；
Wherein, the third depth is more than the 4th depth；Before the training module 504 is used for according to the training sample image
Difference between the mark segmentation result and first segmentation result of scape segmentation, the mark segmentation result and described second point
Cut the difference between difference and the mark segmentation result and the second new segmentation result between result, training institute
State nerve network system.
Optionally, the first segmentation module 502 further includes the second combining unit 5025, for that will be located at the 4th depth
The second segmentation result that second convolution sub-network obtains exports result with the shallow-layer and/or deep layer output result is closed
And obtain the second amalgamation result；The third cutting unit 5024 is for passing through the second convolution sub-network positioned at third depth
The second new process of convolution is carried out to stating the second amalgamation result.
Optionally, the training module 504 is for obtaining between the mark segmentation result and first segmentation result
The first variance data, the second variance data and described between the mark segmentation result and second segmentation result
Mark the third variance data between segmentation result and the second new segmentation result；According to first variance data, institute
State the network parameter with the adjustment nerve network system of the second variance data and the third variance data.
The neural metwork training device of the present embodiment is for realizing corresponding neural metwork training in preceding method embodiment
Method, and the advantageous effect with corresponding embodiment of the method, details are not described herein.
The present embodiment also provides a kind of computer program comprising has computer program instructions, described program instruction to be located
The step of managing when device executes for realizing any neural network training method provided in an embodiment of the present invention.
The present embodiment also provides a kind of computer readable storage medium, is stored thereon with computer program instructions, the program
The step of instruction realizes any neural network training method provided in an embodiment of the present invention when being executed by processor.
Embodiment five
With reference to Fig. 7, a kind of structure diagram of according to embodiments of the present invention five image processing apparatus is shown.
The image processing apparatus of the embodiment of the present invention includes：Second segmentation module 702, for pending image to be inputted god
Through network system, background segment processing before being carried out to the pending image by the nerve network system；Blurring module
704, for obtaining the background parts for carrying out the pending image after the preceding background segment is handled, to the background parts into
Row segmentation virtualization processing.
Optionally, the second segmentation module 702 includes：Third cutting unit 7021, for passing through nerve network system
First process of convolution is carried out to pending image, obtains the first segmentation result of the preceding background segment of the pending image；The
Four cutting units 7022, for carrying out the second process of convolution to first segmentation result, before obtaining the pending image
Second segmentation result of background segment；5th cutting unit 7023 is used for according to second segmentation result to described pending
Background segment processing before image carries out, to carry out handling relevant processing with preceding background segment to the pending image.
Optionally, the nerve network system is to train to obtain using the neural metwork training device of the embodiment of the present invention four
Nerve network system.
Optionally, described device further includes：Processing module 706, after to the progress preceding background segment processing
Pending image carries out Steerable filter processing and/or alpha matting processing.
Optionally, the blurring module 704 is used for according to each pixel in the background parts to the pending figure
The distance of the designated edge of picture carries out Fuzzy Processing to each pixel.
The neural metwork training device of the present embodiment for realizing corresponding image processing method in preceding method embodiment,
And the advantageous effect with corresponding embodiment of the method, details are not described herein.
The present embodiment also provides a kind of computer program comprising has computer program instructions, described program instruction to be located
The step of managing when device executes for realizing any image processing method provided in an embodiment of the present invention.
The present embodiment also provides a kind of computer readable storage medium, is stored thereon with computer program instructions, the program
The step of instruction realizes any image processing method provided in an embodiment of the present invention when being executed by processor.
Embodiment six
The embodiment of the present invention six provides a kind of electronic equipment, such as can be mobile terminal, personal computer (PC), put down
Plate computer, server etc..Below with reference to Fig. 8, it illustrates suitable for for realizing terminal device or the service of the embodiment of the present invention
The structural schematic diagram of the electronic equipment 800 of device：As shown in figure 8, electronic equipment 800 includes one or more processors, communication member
Part etc., one or more of processors are for example：One or more central processing unit (CPU) 801, and/or it is one or more
Image processor (GPU) 813 etc., processor can according to the executable instruction being stored in read-only memory (ROM) 802 or
From the executable instruction that storage section 808 is loaded into random access storage device (RAM) 803 execute it is various it is appropriate action and
Processing.Communication device includes communication component 812 and/or communication interface 809.Wherein, communication component 812 may include but be not limited to net
Card, the network interface card may include but be not limited to IB (Infiniband) network interface card, and communication interface 809 includes such as LAN card, modulation /demodulation
The communication interface of the network interface card of device etc., communication interface 809 execute communication process via the network of such as internet.
Processor can be communicated with read-only memory 802 and/or random access storage device 803 to execute executable instruction,
It is connected with communication component 812 by communication bus 804 and is communicated with other target devices through communication component 812, to completes this
The corresponding operation of any one neural network training method that inventive embodiments provide, for example, by nerve network system to training
Sample image carries out the first process of convolution, obtains the first segmentation result of the preceding background segment of the training sample image；And
Second process of convolution is carried out to first segmentation result, obtains the second segmentation of the preceding background segment of the training sample image
As a result；According to the difference between the mark segmentation result of the preceding background segment of the training sample image and first segmentation result
Difference different and between the mark segmentation result and second segmentation result, the training nerve network system.
In addition, in RAM 803, it can also be stored with various programs and data needed for device operation.CPU801 or
GPU813, ROM802 and RAM803 are connected with each other by communication bus 804.In the case where there is RAM803, ROM802 is can
Modeling block.RAM803 stores executable instruction, or executable instruction is written into ROM802 at runtime, and executable instruction makes place
It manages device and executes the corresponding operation of above-mentioned communication means.Input/output (I/O) interface 805 is also connected to communication bus 804.Communication
Component 812 can be integrally disposed, may be set to be with multiple submodule (such as multiple IB network interface cards), and in communication bus chain
It connects.
It is connected to I/O interfaces 805 with lower component：Importation 806 including keyboard, mouse etc.；It is penetrated including such as cathode
The output par, c 807 of spool (CRT), liquid crystal display (LCD) etc. and loud speaker etc.；Storage section 808 including hard disk etc.；
And the communication interface 809 of the network interface card including LAN card, modem etc..Driver 810 is also according to needing to connect
It is connected to I/O interfaces 805.Detachable media 811, such as disk, CD, magneto-optic disk, semiconductor memory etc. are pacified as needed
On driver 810, in order to be mounted into storage section 808 as needed from the computer program read thereon.
It should be noted that framework as shown in Figure 8 is only a kind of optional realization method, it, can root during concrete practice
The component count amount and type of above-mentioned Fig. 8 are selected, are deleted, increased or replaced according to actual needs；It is set in different function component
It sets, separately positioned or integrally disposed and other implementations, such as separable settings of GPU and CPU or can be by GPU collection can also be used
At on CPU, the separable setting of communication device, can also be integrally disposed on CPU or GPU, etc..These interchangeable embodiment party
Formula each falls within protection scope of the present invention.
Particularly, according to embodiments of the present invention, it is soft to may be implemented as computer for the process above with reference to flow chart description
Part program.For example, the embodiment of the present invention includes a kind of computer program products comprising be tangibly embodied in machine readable media
On computer program, computer program includes the program code for method shown in execution flow chart, and program code can wrap
The corresponding instruction of corresponding execution neural network training method step provided in an embodiment of the present invention is included, for example, passing through neural network
System carries out the first process of convolution to training sample image, obtains the first segmentation of the preceding background segment of the training sample image
As a result；And the second process of convolution is carried out to first segmentation result, obtain the preceding background segment of the training sample image
The second segmentation result；It is tied according to the mark segmentation result of the preceding background segment of the training sample image and first segmentation
The difference between difference and the mark segmentation result and second segmentation result between fruit, the training nerve net
Network system.In such embodiments, which can be downloaded and installed by communication device from network, and/
Or it is mounted from detachable media 811.When the computer program is executed by processor, in the method for executing the embodiment of the present invention
The above-mentioned function of limiting.
Embodiment seven
The embodiment of the present invention seven provides a kind of electronic equipment, such as can be mobile terminal, personal computer (PC), put down
Plate computer, server etc..Below with reference to Fig. 9, it illustrates suitable for for realizing terminal device or the service of the embodiment of the present invention
The structural schematic diagram of the electronic equipment 900 of device：As shown in figure 9, electronic equipment 900 includes one or more processors, communication member
Part etc., one or more of processors are for example：One or more central processing unit (CPU) 901, and/or it is one or more
Image processor (GPU) 913 etc., processor can according to the executable instruction being stored in read-only memory (ROM) 902 or
From the executable instruction that storage section 908 is loaded into random access storage device (RAM) 903 execute it is various it is appropriate action and
Processing.Communication device includes communication component 912 and/or communication interface 909.Wherein, communication component 912 may include but be not limited to net
Card, the network interface card may include but be not limited to IB (Infiniband) network interface card, and communication interface 909 includes such as LAN card, modulation /demodulation
The communication interface of the network interface card of device etc., communication interface 909 execute communication process via the network of such as internet.
Processor can be communicated with read-only memory 902 and/or random access storage device 903 to execute executable instruction,
It is connected with communication component 912 by communication bus 904 and is communicated with other target devices through communication component 912, to completes this
The corresponding operation of any one image processing method that inventive embodiments provide, for example, by pending image input neural network system
System, background segment processing before being carried out to the pending image by the nerve network system；It obtains and carries out the preceding background
The background parts of pending image after dividing processing carry out segmentation virtualization processing to the background parts.
In addition, in RAM 903, it can also be stored with various programs and data needed for device operation.CPU901 or
GPU913, ROM902 and RAM903 are connected with each other by communication bus 904.In the case where there is RAM903, ROM902 is can
Modeling block.RAM903 stores executable instruction, or executable instruction is written into ROM902 at runtime, and executable instruction makes place
It manages device and executes the corresponding operation of above-mentioned communication means.Input/output (I/O) interface 905 is also connected to communication bus 904.Communication
Component 912 can be integrally disposed, may be set to be with multiple submodule (such as multiple IB network interface cards), and in communication bus chain
It connects.
It is connected to I/O interfaces 905 with lower component：Importation 906 including keyboard, mouse etc.；It is penetrated including such as cathode
The output par, c 907 of spool (CRT), liquid crystal display (LCD) etc. and loud speaker etc.；Storage section 908 including hard disk etc.；
And the communication interface 909 of the network interface card including LAN card, modem etc..Driver 910 is also according to needing to connect
It is connected to I/O interfaces 905.Detachable media 911, such as disk, CD, magneto-optic disk, semiconductor memory etc. are pacified as needed
On driver 910, in order to be mounted into storage section 908 as needed from the computer program read thereon.
It should be noted that framework as shown in Figure 9 is only a kind of optional realization method, it, can root during concrete practice
The component count amount and type of above-mentioned Fig. 9 are selected, are deleted, increased or replaced according to actual needs；It is set in different function component
It sets, separately positioned or integrally disposed and other implementations, such as separable settings of GPU and CPU or can be by GPU collection can also be used
At on CPU, the separable setting of communication device, can also be integrally disposed on CPU or GPU, etc..These interchangeable embodiment party
Formula each falls within protection scope of the present invention.
Particularly, according to embodiments of the present invention, it is soft to may be implemented as computer for the process above with reference to flow chart description
Part program.For example, the embodiment of the present invention includes a kind of computer program products comprising be tangibly embodied in machine readable media
On computer program, computer program includes the program code for method shown in execution flow chart, and program code can wrap
The corresponding instruction of corresponding execution image processing method step provided in an embodiment of the present invention is included, for example, pending image is inputted
Nerve network system, background segment processing before being carried out to the pending image by the nerve network system；It obtains and carries out
The background parts of pending image after the preceding background segment processing carry out segmentation virtualization processing to the background parts.
In such embodiments, which can be downloaded and installed by communication device from network, and/or from removable
Medium 911 is unloaded to be mounted.When the computer program is executed by processor, execute limited in the method for the embodiment of the present invention it is upper
State function.
It may be noted that according to the needs of implementation, all parts/step described in the embodiment of the present invention can be split as more
The part operation of two or more components/steps or components/steps can be also combined into new component/step by multi-part/step
Suddenly, to realize the purpose of the embodiment of the present invention.
It is above-mentioned to be realized in hardware, firmware according to the method for the embodiment of the present invention, or be implemented as being storable in note
Software or computer code in recording medium (such as CD ROM, RAM, floppy disk, hard disk or magneto-optic disk), or it is implemented through net
The original storage that network is downloaded in long-range recording medium or nonvolatile machine readable media and will be stored in local recording medium
In computer code, can be stored in using all-purpose computer, application specific processor or can compile to method described here
Such software processing in journey or the recording medium of specialized hardware (such as ASIC or FPGA).It is appreciated that computer, processing
Device, microprocessor controller or programmable hardware include can store or receive software or computer code storage assembly (for example,
RAM, ROM, flash memory etc.), when the software or computer code are by computer, processor or hardware access and execute, realize
Processing method described here.In addition, when all-purpose computer accesses the code for realizing the processing being shown here, code
Execute special purpose computer all-purpose computer is converted to for executing the processing being shown here.
Those of ordinary skill in the art may realize that lists described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and method and step can be realized with the combination of electronic hardware or computer software and electronic hardware.These functions are actually
It is implemented in hardware or software, depends on the specific application and design constraint of 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 range of the embodiment of the present invention.
Embodiment of above is merely to illustrate the embodiment of the present invention, and is not the limitation to the embodiment of the present invention, related skill
The those of ordinary skill in art field can also make various in the case where not departing from the spirit and scope of the embodiment of the present invention
Variation and modification, therefore all equivalent technical solutions also belong to the scope of the embodiment of the present invention, the patent of the embodiment of the present invention
Protection domain should be defined by the claims.
Claims (10)
1. a kind of neural network training method, including：
The first process of convolution is carried out to training sample image by nerve network system, obtains the preceding back of the body of the training sample image
First segmentation result of scape segmentation；And the second process of convolution is carried out to first segmentation result, obtain the training sample
Second segmentation result of the preceding background segment of image；
According to the difference between the mark segmentation result of the preceding background segment of the training sample image and first segmentation result
Difference different and between the mark segmentation result and second segmentation result, the training nerve network system.
2. a kind of image processing method, including：
Pending image is inputted into nerve network system, the back of the body before being carried out to the pending image by the nerve network system
Scape dividing processing；
The background parts for carrying out the pending image after the preceding background segment processing are obtained, the background parts are divided
Section virtualization processing.
3. a kind of neural metwork training device, including：
First segmentation module, for carrying out the first process of convolution to training sample image by nerve network system, described in acquisition
First segmentation result of the preceding background segment of training sample image；And first segmentation result is carried out at the second convolution
Reason, obtains the second segmentation result of the preceding background segment of the training sample image；
Training module, the mark segmentation result for the preceding background segment according to the training sample image are divided with described first
As a result the difference between difference and the mark segmentation result and second segmentation result between, the training nerve
Network system.
4. a kind of image processing apparatus, including：
Second segmentation module, for pending image to be inputted nerve network system, by the nerve network system to described
Background segment processing before pending image carries out；
Blurring module, the background parts for obtaining the pending image after carrying out the preceding background segment processing, to described
Background parts carry out segmentation virtualization processing.
5. a kind of computer program comprising there is computer program instructions, wherein described program instruction is used when being executed by processor
In step corresponding to realization neural network training method described in claim 1.
6. a kind of computer program comprising there is computer program instructions, wherein described program instruction is used when being executed by processor
In realizing step corresponding to image processing method described in claim 2.
7. a kind of computer readable storage medium, is stored thereon with computer program instructions, wherein described program instruction is handled
For realizing step corresponding to neural network training method described in claim 1 when device executes.
8. a kind of computer readable storage medium, is stored thereon with computer program instructions, wherein described program instruction is handled
For realizing step corresponding to the image processing method described in claim 2 when device executes.
9. a kind of electronic equipment, including：Processor, memory, communication device and communication bus, the processor, the storage
Device and the communication device complete mutual communication by the communication bus；
The memory makes the processor execute as right is wanted for storing an at least executable instruction, the executable instruction
The step for asking 1 neural network training method corresponding.
10. a kind of electronic equipment, including：Processor, memory, communication device and communication bus, the processor, the storage
Device and the communication device complete mutual communication by the communication bus；
The memory makes the processor execute as right is wanted for storing an at least executable instruction, the executable instruction
The step for asking 2 described image processing methods corresponding.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810463674.3A CN108665475A (en) | 2018-05-15 | 2018-05-15 | Neural metwork training, image processing method, device, storage medium and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810463674.3A CN108665475A (en) | 2018-05-15 | 2018-05-15 | Neural metwork training, image processing method, device, storage medium and electronic equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108665475A true CN108665475A (en) | 2018-10-16 |
Family
ID=63779599
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810463674.3A Pending CN108665475A (en) | 2018-05-15 | 2018-05-15 | Neural metwork training, image processing method, device, storage medium and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108665475A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109491784A (en) * | 2018-10-18 | 2019-03-19 | 北京旷视科技有限公司 | Reduce method, apparatus, the electronic equipment, readable storage medium storing program for executing of EMS memory occupation amount |
| CN110110778A (en) * | 2019-04-29 | 2019-08-09 | 腾讯科技（深圳）有限公司 | Image processing method, device, electronic equipment and computer readable storage medium |
| CN110175966A (en) * | 2019-05-30 | 2019-08-27 | 上海极链网络科技有限公司 | Non-mated images generation method, system, server and storage medium |
| CN111583264A (en) * | 2020-05-06 | 2020-08-25 | 上海联影智能医疗科技有限公司 | Training method for image segmentation network, image segmentation method, and storage medium |
| CN112330709A (en) * | 2020-10-29 | 2021-02-05 | 奥比中光科技集团股份有限公司 | Foreground image extraction method and device, readable storage medium and terminal equipment |
| CN112749801A (en) * | 2021-01-22 | 2021-05-04 | 上海商汤智能科技有限公司 | Neural network training and image processing method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170243053A1 (en) * | 2016-02-18 | 2017-08-24 | Pinscreen, Inc. | Real-time facial segmentation and performance capture from rgb input |
| CN107392933A (en) * | 2017-07-12 | 2017-11-24 | 维沃移动通信有限公司 | A kind of method and mobile terminal of image segmentation |
| CN107545571A (en) * | 2017-09-22 | 2018-01-05 | 深圳天琴医疗科技有限公司 | A kind of image detecting method and device |
| CN107749046A (en) * | 2017-10-27 | 2018-03-02 | 维沃移动通信有限公司 | A kind of image processing method and mobile terminal |
-
2018
- 2018-05-15 CN CN201810463674.3A patent/CN108665475A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170243053A1 (en) * | 2016-02-18 | 2017-08-24 | Pinscreen, Inc. | Real-time facial segmentation and performance capture from rgb input |
| CN107392933A (en) * | 2017-07-12 | 2017-11-24 | 维沃移动通信有限公司 | A kind of method and mobile terminal of image segmentation |
| CN107545571A (en) * | 2017-09-22 | 2018-01-05 | 深圳天琴医疗科技有限公司 | A kind of image detecting method and device |
| CN107749046A (en) * | 2017-10-27 | 2018-03-02 | 维沃移动通信有限公司 | A kind of image processing method and mobile terminal |
Non-Patent Citations (3)
| Title |
|---|
| JONATHAN LONG,ET AL: ""Fully Convolutional Networks"", 《IEEE COMPUTER VISION AND PATTERN RECOGNITION》 * |
| SHENXIAOLU1984: ""【图像分割】Fully Convolutional Networks for Semantic Segmentation"", 《CSDN》 * |
| 刘晓阳，薛纯: ""基于多尺度的水下图像显著性区域检测"", 《微型机与应用》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109491784A (en) * | 2018-10-18 | 2019-03-19 | 北京旷视科技有限公司 | Reduce method, apparatus, the electronic equipment, readable storage medium storing program for executing of EMS memory occupation amount |
| CN109491784B (en) * | 2018-10-18 | 2021-01-22 | 北京旷视科技有限公司 | Method and device for reducing memory occupation amount, electronic equipment and readable storage medium |
| CN110110778A (en) * | 2019-04-29 | 2019-08-09 | 腾讯科技（深圳）有限公司 | Image processing method, device, electronic equipment and computer readable storage medium |
| WO2020220858A1 (en) * | 2019-04-29 | 2020-11-05 | 腾讯科技（深圳）有限公司 | Image processing method and apparatus, electronic device, and computer-readable storage medium |
| CN110175966A (en) * | 2019-05-30 | 2019-08-27 | 上海极链网络科技有限公司 | Non-mated images generation method, system, server and storage medium |
| CN111583264A (en) * | 2020-05-06 | 2020-08-25 | 上海联影智能医疗科技有限公司 | Training method for image segmentation network, image segmentation method, and storage medium |
| CN111583264B (en) * | 2020-05-06 | 2024-02-27 | 上海联影智能医疗科技有限公司 | Training method for image segmentation network, image segmentation method, and storage medium |
| CN112330709A (en) * | 2020-10-29 | 2021-02-05 | 奥比中光科技集团股份有限公司 | Foreground image extraction method and device, readable storage medium and terminal equipment |
| CN112749801A (en) * | 2021-01-22 | 2021-05-04 | 上海商汤智能科技有限公司 | Neural network training and image processing method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108665475A (en) | Neural metwork training, image processing method, device, storage medium and electronic equipment | |
| Zhu et al. | A benchmark for edge-preserving image smoothing | |
| CN105184249B (en) | Method and apparatus for face image processing | |
| CN111950723B (en) | Neural network model training method, image processing method, device and terminal equipment | |
| US8692830B2 (en) | Automatic avatar creation | |
| CN106971165B (en) | A kind of implementation method and device of filter | |
| CN108229279A (en) | Face image processing process, device and electronic equipment | |
| CN109325988A (en) | A kind of facial expression synthetic method, device and electronic equipment | |
| CN110378235A (en) | A kind of fuzzy facial image recognition method, device and terminal device | |
| CN109712144A (en) | Processing method, training method, equipment and the storage medium of face-image | |
| CN108647634A (en) | Framing mask lookup method, device, computer equipment and storage medium | |
| CN107395960A (en) | Photographic method and device, computer installation and computer-readable recording medium | |
| CN111127309B (en) | Portrait style migration model training method, portrait style migration method and device | |
| CN108234858A (en) | Image virtualization processing method, device, storage medium and electronic equipment | |
| CN111091493B (en) | Image translation model training method, image translation method and device and electronic equipment | |
| CN107610046A (en) | Background-blurring method, apparatus and system | |
| CN107464217A (en) | A kind of image processing method and device | |
| CN108230384A (en) | Picture depth computational methods, device, storage medium and electronic equipment | |
| CN106650795A (en) | Sorting method of hotel room type images | |
| CN108596070A (en) | Character recognition method, device, storage medium, program product and electronic equipment | |
| CN110009573A (en) | Model training, image processing method, device, electronic equipment and computer readable storage medium | |
| CN107610675A (en) | A kind of image processing method and device based on dynamic level | |
| CN115100334B (en) | Image edge tracing and image animation method, device and storage medium | |
| CN108769545A (en) | A kind of image processing method, image processing apparatus and mobile terminal | |
| CN107851309A (en) | A kind of image enchancing 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: 20181016 |
|
| RJ01 | Rejection of invention patent application after publication |