CN109886990A - A kind of image segmentation system based on deep learning - Google Patents
A kind of image segmentation system based on deep learning Download PDFInfo
- Publication number
- CN109886990A CN109886990A CN201910083679.8A CN201910083679A CN109886990A CN 109886990 A CN109886990 A CN 109886990A CN 201910083679 A CN201910083679 A CN 201910083679A CN 109886990 A CN109886990 A CN 109886990A
- Authority
- CN
- China
- Prior art keywords
- component
- segmentation
- building block
- feature
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image segmentation system based on deep learning, including feature extraction unit, component cutting units, parts match unit, component integrated unit and optimization unit, wherein feature extraction unit, the feature that input picture is extracted using convolutional neural networks, generates multiple characteristic patterns；Component cutting unit is split respectively using each building block of the convolutional neural networks to the object in image, generates the component feature figure and component segmentation figure of each building block；Parts match unit is respectively derived by the component feature figure of each building block using convolutional neural networks and generates the other segmentation figure of a universal class, calculates and output block matches score value；The component feature figure correspondence of each building block is merged using convolutional neural networks, generates the other segmentation figure of a universal class, calculate and output block merges score value by component integrated unit；Optimize unit, merges score value using the costing bio disturbance layer, parts match score value and component of convolutional neural networks, update the component feature figure of convolutional neural networks parameter and each building block, update and export the component segmentation figure after optimization.It can be realized the less Automatic image segmentation manually participated in through the invention, image segmentation is accurate, high-efficient.
Description
Technical field
The present invention relates to field of image processings, more particularly, to a kind of image segmentation system based on deep learning.
Background technique
Solar panel is one of most important equipment of photovoltaic plant, its maintenance, maintenance is extremely important.Due to photovoltaic
Power station is in the farming region of complicated landform mostly, how automatically to carry out the current check of solar panel and maintenance is one
A urgent problem to be solved.In addition, one piece of solar panel is usually made of several pieces of daughter boards, the specific location to break down
It is information necessary to subsequent maintenance, hot spot failure has occurred in the kth block daughter board such as on n-th piece of solar panel.At present
Fault detection method usually require largely have markd training dataset, be generally difficult to obtain and need to spend a large amount of people
Work marks work.In addition, these methods usually require additional step to obtain specific abort situation.In order to reduce artificial mark
Note work simultaneously obtains specific abort situation, and the invention proposes a kind of weak prisons of daughter board segmentation that can be used for solar panel
The image segmentation system based on deep learning superintended and directed, the system can be easily extended to the segmentation and solar-electricity of arbitrary objects
Normal, the anomaly classification task of pond plate.It can be realized the purpose of the less Automatic image segmentation manually participated in through the invention, scheme
Accurately as segmentation, high-efficient.
Summary of the invention
Purpose to realize the present invention, is achieved using following technical scheme:
A kind of image segmentation system based on deep learning, including feature extraction unit, component cutting unit, parts match
Unit, component integrated unit and optimization unit, in which: feature extraction unit, for extracting input figure using convolutional neural networks
The feature of picture generates multiple characteristic patterns；Component cutting unit, for utilizing convolutional neural networks to each of the object in image
Building block is split respectively, generates the component feature figure and component segmentation figure of each building block；Parts match unit is used
The other segmentation figure of a universal class is generated in respectively being derived using convolutional neural networks by the component feature figure of each building block, is calculated simultaneously
Output block matches score value；Component integrated unit, for utilizing convolutional neural networks by the component feature figure of each building block
Correspondence is merged, and the other segmentation figure of a universal class is generated, and calculates and output block merges score value；Optimize unit, for utilizing volume
The product costing bio disturbance layer of neural network, parts match score value and component merge score value, update convolutional neural networks parameter and each
The component feature figure of building block updates and exports the component segmentation figure after optimization.
The image segmentation system, wherein it is characterized by: feature extraction unit, utilizes the volume of convolutional neural networks
Lamination, pond layer, normalization layer, extract the feature of input picture, generate multiple characteristic patterns；Component cutting unit, utilizes convolution
Warp lamination, convolutional layer, costing bio disturbance layer and the multiple characteristic patterns of neural network, to each composition of the object in image
Component is split respectively, generates the component feature figure and component segmentation figure of each building block；Parts match unit, utilizes volume
The product warp lamination of neural network, convolutional layer, costing bio disturbance layer and each building block component feature figure, in image
Object carry out whole segmentation, i.e., respectively derives the generation other segmentation figure of one universal class by the component feature figure of each building block, and
According to the similitude between the other segmentation figure of multiple universal class of generation, calculates and output block matches score value；Component integrated unit, benefit
With the component feature figure of the warp lamination of convolutional neural networks, convolutional layer, costing bio disturbance layer and each building block, to figure
Object as in carries out whole segmentation, i.e., merges the component feature figure correspondence of each building block, generate a universal class
Other segmentation figure, and according to the integrality for generating the other segmentation figure of universal class, it calculates and output block merges score value；Optimize unit, utilizes
The costing bio disturbance layers of convolutional neural networks, parts match score value and component merge score value, update convolutional neural networks parameter and each
The component feature figure of a building block updates and exports the component segmentation figure after optimization.
The image segmentation system, in which: feature extraction unit, using the convolutional layer of convolutional neural networks, pond layer,
Layer is normalized, the feature of input picture is extracted, generates multiple characteristic patterns, comprising:
A. convolution operation is carried out to input picture, generates trellis diagram；
B. operation is normalized to trellis diagram, the trellis diagram after generating normalization；
C. pondization operation is carried out to the trellis diagram after normalization, generates characteristic pattern；
D. it is computed repeatedly several times after adjusting the parameter of above-mentioned steps；
When e. reaching preset calculation times, the characteristic pattern generated under multiple groups parameter is exported.
The image segmentation system, in which: component cutting unit utilizes the warp lamination of convolutional neural networks, convolution
Layer, costing bio disturbance layer and multiple characteristic patterns, are split each building block of the object in image respectively, generate each
The component feature figure and component segmentation figure of a building block, comprising:
A. deconvolution operation is carried out to characteristic pattern, generates deconvolution figure；
B. operation is normalized to deconvolution figure, the deconvolution figure after generating normalization；
C. up-sampling operation, generating unit characteristic pattern are carried out to the deconvolution figure after normalization；
D. it is computed repeatedly several times after adjusting the parameter of above-mentioned steps；
It when e. training network, repeats the above steps, until loss reaches threshold value；
F. the component feature figure and component segmentation figure of each building block are exported.
The image segmentation system, in which: parts match unit utilizes the warp lamination of convolutional neural networks, convolution
The component feature figure of layer, costing bio disturbance layer and each building block carries out whole segmentation to the object in image, i.e., by every
The component feature figure of a building block, which respectively derives, generates the other segmentation figure of a universal class, and according to the other segmentation figure of multiple universal class of generation
Between similitude, calculate and output block match score value, comprising:
A. convolution operation is carried out to the component feature figure of each building block, generates the trellis diagram of each building block；
B. up-sampling operation is carried out to the trellis diagram of each building block, generates the volume for each building block that size increases
Product figure；
C. the trellis diagram for each building block that size increases is merged；
D. it is computed repeatedly several times after adjusting the parameter of above-mentioned steps；
It when e. training network, repeats the above steps, until loss reaches threshold value；
F. the other segmentation figure of universal class derived respectively by each building block is exported；
G. the similitude of the other segmentation figure of universal class of calculating current part and remaining part between any two, and putting down similitude
Mean value is exported as the parts match score value of current part.
The image segmentation system, in which: component integrated unit utilizes the warp lamination of convolutional neural networks, convolution
The component feature figure of layer, costing bio disturbance layer and each building block carries out whole segmentation to the object in image, i.e., will be every
The component feature figure correspondence of a building block is merged, and generates the other segmentation figure of a universal class, and do not divide according to universal class is generated
The integrality of figure, calculates and output block merges score value, comprising:
A. convolution operation is carried out to the component feature figure of each building block, generates the trellis diagram of each building block；
B. mixing operation is carried out to the trellis diagram of each building block, generates fusion figure；
C. fusion is schemed to carry out convolution operation, generates fusion feature figure；
D. up-sampling operation is carried out to fusion feature figure, generates the fusion feature figure that size increases；
E. it is computed repeatedly several times after adjusting the parameter of above-mentioned steps；
It when f. training network, repeats the above steps, until loss reaches threshold value；
G. the other segmentation figure of universal class merged out by each building block is exported；
H. the integrality of the other segmentation figure of above-mentioned universal class is calculated, combining position information, which calculates and exports the component of current part, to be melted
Close score value.
The image segmentation system, in which: optimization unit utilizes costing bio disturbance layer, the component of convolutional neural networks
Score value is merged with score value and component, the component feature figure of network parameter and each building block is updated, updates and export multiple excellent
Component segmentation figure after change, comprising:
A. score value is merged according to parts match score value and component, calculates the overall score value of each building block；
B. using the overall score value of each building block as its weight, the component segmentation figure of each building block is melted
It closes, generates fusion segmentation figure, and calculate loss；
C. the whole loss of network is calculated；
D. network parameter is updated；
Component segmentation figure when e. reaching preset calculation times, after output optimization.
Detailed description of the invention
Fig. 1 is one and is integrated with application system schematic diagram of the invention.
Fig. 2 is image segmentation system functional schematic of the present invention.
Fig. 3 is the flow chart of feature of present invention extraction unit.
Fig. 4 is the flow chart of component cutting unit of the present invention.
Fig. 5 is the flow chart of the other segmentation figure generation step of universal class of parts match unit of the present invention.
Fig. 6 is the flow chart that the matching score value of parts match unit of the present invention calculates.
Fig. 7 is the flow chart of component integrated unit of the present invention.
Fig. 8 is the flow chart of present invention optimization unit.
Fig. 9 is the key that the general hardware block diagram according to point positioning system of the embodiment of the present invention.
Specific embodiment
Embodiment of the present invention is described in detail with reference to the accompanying drawing.
Fig. 1 is one and is integrated with the schematic diagram of application system of the invention, and the application system includes unmanned plane and photovoltaic
Power station.Unmanned plane is embedded in common camera to capture the photovoltaic plant image in range of visibility.The present invention passes through analysis capture
Photovoltaic plant image be partitioned into the overall region and its sub-tile areas of every piece of solar panel, and result fed back to therefore
Hinder detection system, and then judges whether solar panel faulty and exact position of failure.System shown in FIG. 1
Only of the invention one applies example, may be more or less than the equipment number that it is included in practical application, or makes
With different equipment, or for different scenes.
In general, work, the labeled data that network training process must be automatically generated manually are marked in order to reduce
The step of, these data inevitably include noise, because it is easy in the case where no supervision by error label.Together
When, each solar panel only includes boundary (grid) and core (daughter board) component.Target core component is limited in boundary
In component.Simultaneously, it should search for border components except core component region.Based on these observations, the present invention proposes a kind of base
It is designed in the structure of component, forms portion to force network correctly to learn each of object out from noisy training data
Part region.
According to Fig.2, present invention input is a frame image, is exported as the whole area of every piece of solar panel in image
Domain and its sub-tile areas.According to Fig.2, of the invention, a kind of image segmentation system based on deep learning, using dividing and rule
Strategy, study only focuses on a building block of cutting object every time, then again merges result, can be according to having
The parameter of the data label study convolutional neural networks of noise, reduces artificial mark cost.More specifically, of the invention based on depth
The image segmentation system of degree study includes feature extraction unit, component cutting unit, parts match unit, component integrated unit and
Optimize unit.Wherein: feature extraction unit is extracted using layers such as the convolutional layer of convolutional neural networks, pond layer, normalization layers
The feature (such as the characteristics of image such as color, texture, shape and spatial relationship) of input picture, generates multiple characteristic patterns；Component point
Unit is cut, using the warp lamination of convolutional neural networks, convolutional layer, costing bio disturbance layer and multiple characteristic patterns, in image
Each building block of object is split respectively, generates the component feature figure and component segmentation figure of each building block；Component
Matching unit utilizes the component of the warp lamination of convolutional neural networks, convolutional layer, costing bio disturbance layer and each building block
Characteristic pattern carries out whole segmentation to the object in image, i.e., derives generation one respectively by the component feature figure of each building block
The other segmentation figure of a universal class, and according to the similitude between the other segmentation figure of multiple universal class of generation, it calculates and output block matching divides
Value；Component integrated unit utilizes the warp lamination of convolutional neural networks, convolutional layer, costing bio disturbance layer and each building block
Component feature figure, whole segmentation is carried out to the object in image, i.e., is carried out the component feature figure of each building block is corresponding
Fusion generates the other segmentation figure of a universal class, and according to the integrality for generating the other segmentation figure of universal class, calculates simultaneously output block fusion point
Value.Optimize unit, merges score value using the costing bio disturbance layer, parts match score value and component of convolutional neural networks, update network
The component feature figure of parameter and each building block updates and exports the component segmentation figure after multiple optimizations.
According to Fig.3, feature extraction unit utilizes the convolutional layer of convolutional neural networks first stage, pond layer, normalizing
Change the layers such as layer, extracts the feature of input picture.As shown in figure 3, this unit utilizes the convolution of multiple groups different parameters, Chi Hua, normalizing
Change the feature that different levels are extracted in operation to input picture.This unit implements convolution operation to input picture first with convolution kernel
And obtain trellis diagram.Later, this unit utilizes linearity correction unit (the Rectified Linear in convolutional neural networks
Unit) and operation is normalized to trellis diagram in local acknowledgement's method for normalizing.Further, after this unit is to above-mentioned normalization
Trellis diagram carries out maximum or average pondization processing.In order to obtain richer multiple dimensioned feature, this unit is different using multiple groups
Parameter when reaching preset calculation times, exports the characteristic pattern generated under multiple groups parameter to aforesaid operations repeated several times.Tool
Body is to the embodiment of the present invention, and feature extraction unit extracts the following characteristics in image in the present invention: boundary (grid) component and
Core (daughter board) component generates boundary (grid) component feature figure and core (daughter board) component feature figure.
On the whole, the other segmentation problem of universal class is divided into the segmentation of several building blocks of an object by the present invention, and
Label noise is overcome using complementary characteristic.Based on extracted characteristic pattern, present invention use divides sub-network, component by component
The network of sub-network and component fusant network composition is matched, in Weakly supervised mode, study has the core of noise label
Part.Component cutting unit forms a component (example in multiple segmentations branch (such as Liang Ge branch) and each branch when working
Such as core (daughter board) component in boundary (grid) component and the second branch in the first branch).Parts match unit minimizes
The overlappings of all components.It generates multiple other segmentation figures of universal class according to corresponding component segmentation result, and estimates their phase
Like degree.Component integrated unit minimizes the gap of all components, it does not divide permeate universal class of component segmentation result
As a result.
According to Fig.4, component cutting unit utilizes warp lamination, convolutional layer, the costing bio disturbance of convolutional neural networks
Layer and multiple characteristic patterns, are split each building block of the object in image respectively, generate each building block
Component feature figure and component segmentation figure.In the present invention, the above-mentioned partitioning scheme of component cutting unit forms two segmentations point
, it respectively include a component in the segmentation result of each branch, i.e. boundary (grid) component and the second branch in the first branch
In core (daughter board) component).Certain branch also can according to the increasing for building block quantity of objects in images and therewith
Increase.
Workflow as shown in Figure 4, the component cutting unit utilize several groups of deconvolution, normalization with different parameters
Come extracting parts feature and generating unit characteristic pattern with up-sampling step.It is carried out first with the characteristic pattern that deconvolution verification is extracted
Deconvolution obtains deconvolution figure.Then, operation is normalized to deconvolution figure, the deconvolution figure after generating normalization.To returning
Deconvolution figure after one change carries out up-sampling operation, generating unit characteristic pattern.Using convolutional layer to the channel of component feature figure
Number is adjusted.In order to obtain be originally inputted the segmentation with same scale, the unit adjust convolutional neural networks correlation ginseng
It counts and repeats the above repeatedly.It when training network, repeats the above steps, utilizes loss function calculating unit in costing bio disturbance layer
The difference of segmentation figure and training label, until loss reaches threshold value or iteration reaches preset calculation times and (sets as needed
It is fixed, such as 500,000 times) when, which exports the component feature figure and component segmentation figure of each building block.
According to Fig. 5-6, parts match unit utilizes warp lamination, convolutional layer, the costing bio disturbance of convolutional neural networks
The component feature figure of layer and each building block carries out whole segmentation, i.e. forming by each branch to the object in image
The component feature figure of component, which derives, generates the other segmentation figure of a universal class, and according to the phase between the other segmentation figure of multiple universal class of generation
Like property, calculates and output block matches score value.The other segmentation figure of the universal class be include input picture whole building blocks point
Figure is cut, which does not include the interference patterns such as the ground unrelated with building block in input picture.
In the cutting procedure of component cutting unit, hierarchically realizes and be partitioned into a component every time.Whole network
The first stage of each member fingers only focuses on the segmentation of a component.Second stage pay close attention to the member fingers component and other
Member fingers fusion, matching export the universal class generated by the branch and do not divide.The output of all branches should be mutually matched, i.e.,
Building block in each branch is different, and the component in all branches adds up and forms input picture, this is also benefit of the invention
The place on the basis of label noise is overcome with complementary characteristic.Therefore, parts match unit forces learning process to be minimized
The constraint of overlapping between the core and border components of generation.
Workflow as shown in Figure 5, the parts match unit using multiple groups have different parameters convolution sum on adopt
Sample process is applied to mapping generated of deconvoluting.It carries out convolution behaviour to the component feature figure of each building block first
Make, generates the trellis diagram of each building block.The trellis diagram of each building block is subjected to deconvolution operation, to obtain fusion
Weight, and generate the trellis diagram of each building block of size increase.Then, the up-sampling figure that fusion adjacent layer generates.In order to
The segmentation with same scale is obtained and is originally inputted, unit adjustment relevant parameter simultaneously repeats the above repeatedly.Training net
It when network, repeats the above steps, calculates the difference of the other segmentation figure of universal class between any two using loss function in costing bio disturbance layer, until
Loss reaches threshold value or when iteration reaches preset calculation times (being set as needed, such as 500,000 times), unit output
The other segmentation figure of universal class derived respectively by each building block.Workflow as shown in Figure 6, the unit calculate current part
With the similitude of the other segmentation figure of universal class of remaining part between any two, and using the average value of similitude as the component of current part
Match score value output.
According to Fig.7, component integrated unit utilizes warp lamination, convolutional layer, the costing bio disturbance of convolutional neural networks
The component feature figure of layer and each building block carries out whole segmentation to the object in image, and specific partitioning scheme is:
The component feature figure correspondence of each building block is merged, generates the other segmentation figure of a universal class, and other according to universal class is generated
The integrality of segmentation figure, calculates and output block merges score value.
Since the training label of core component and border components is generated in the case where unsupervised, individually
All there is noise in the component segmentation result that study comes out, and the fusion results of component segmentation should be close to combination tag.Therefore,
Component integrated unit forces learning process by the constraint for minimizing the gap between the core and border components generated.
Workflow as shown in Figure 7, the unit is by the component feature on the corresponding scale of the component in each branch
Figure is fused together.It using several groups there is the convolution sum upper sampling process of different parameters to reflect come the deconvolution for being applied to generate
It penetrates.It carries out convolution operation to the component feature figure of each building block first, generates the trellis diagram of each building block.It will be each
The trellis diagram of a building block carries out deconvolution operation, to obtain the weight of fusion, and generates each composition portion of size increase
The trellis diagram of part.Mixing operation is carried out to the trellis diagram of each building block, generates fusion figure.In order to obtain be originally inputted tool
There is the segmentation of same scale, unit adjustment relevant parameter simultaneously repeats the above repeatedly.When training network, above-mentioned step is repeated
Suddenly, in costing bio disturbance layer, the difference of universal class other segmentation figure and combination tag is calculated using loss function, until loss reaches threshold value
Or iteration, when reaching preset calculation times (being set as needed, such as 500,000 times), unit output forms portion by each
The other segmentation figure of the universal class that part merges out.The unit calculates the integrality of the other segmentation figure of above-mentioned universal class, and combining position information calculates simultaneously
The component fusion score value for the other segmentation figure of universal class that the characteristic pattern of the current each branch components of output merges.
According to Fig.8, optimize unit, utilize the costing bio disturbance layer, parts match score value and component of convolutional neural networks
Score value is merged, the component feature figure of network parameter and each building block is updated, update and exports the component after multiple optimizations point
Cut figure.
In order to realize while optimizing the segmentation result of all parts, this optimization unit goes out net according to the costing bio disturbance of each branch
Result anti-pass is returned network and updates the parameter of related each layer by the whole loss of network.This unit is according to parts match score value and portion
Part merges score value, calculates the overall score value of each building block.It, will be each using the overall score value of each building block as its weight
The component segmentation figure of a building block is merged, and generates fusion segmentation figure, and calculate the loss of this time fusion.According to each point
The loss of branch, calculates the whole loss of network, such as by the loss weighted sum of each branch.When reaching preset calculation times
When, the component segmentation figure after the output optimization of this unit.
Fig. 9 show a hardware block diagram of the present invention, which includes a video camera, and five processing are single
Member, a CPU, a RAM, a display equipment.Feature extraction unit in the processing unit, that is, foregoing description, component point
Unit, parts match unit, component integrated unit and optimization unit are cut, either hardware device can also be program unit,
CPU controls the work of image segmentation system, and controls the operation of each processing unit, and display equipment is for showing input picture, place
Manage result etc..RAM is used to support the operation of CPU.
Claims (7)
1. a kind of image segmentation system based on deep learning, including feature extraction unit, component cutting unit, parts match list
Member, component integrated unit and optimization unit, it is characterised in that: feature extraction unit, it is defeated for being extracted using convolutional neural networks
The feature for entering image generates multiple characteristic patterns；Component cutting unit, for utilizing convolutional neural networks to the object in image
Each building block is split respectively, generates the component feature figure and component segmentation figure of each building block；Parts match list
Member generates the other segmentation figure of a universal class for respectively being derived using convolutional neural networks by the component feature figure of each building block,
It calculates and output block matches score value；Component integrated unit, for utilizing convolutional neural networks by the component of each building block
Characteristic pattern correspondence is merged, and the other segmentation figure of a universal class is generated, and calculates and output block merges score value；Optimize unit, is used for
Score value is merged using the costing bio disturbance layer, parts match score value and component of convolutional neural networks, updates convolutional neural networks parameter
And the component feature figure of each building block, it updates and exports the component segmentation figure after optimization.
2. the image segmentation system according to claim 1 based on deep learning, it is characterised in that: feature extraction unit,
Using the convolutional layer of convolutional neural networks, pond layer, normalization layer, the feature of input picture is extracted, multiple characteristic patterns are generated；Portion
Part cutting unit, using the warp lamination of convolutional neural networks, convolutional layer, costing bio disturbance layer and multiple characteristic patterns, to image
In each building block of object be split respectively, generate the component feature figure and component segmentation figure of each building block；
Parts match unit utilizes the warp lamination of convolutional neural networks, convolutional layer, costing bio disturbance layer and each building block
Component feature figure carries out whole segmentation to the object in image, i.e., respectively derives generation by the component feature figure of each building block
The other segmentation figure of one universal class, and according to the similitude between the other segmentation figure of multiple universal class of generation, it calculates and output block matches
Score value；Component integrated unit utilizes the warp lamination of convolutional neural networks, convolutional layer, costing bio disturbance layer and each composition portion
The component feature figure of part carries out whole segmentation to the object in image, i.e., by the component feature figure of each building block it is corresponding into
Row fusion generates the other segmentation figure of a universal class, and according to the integrality for generating the other segmentation figure of universal class, calculates simultaneously output block fusion
Score value；Optimize unit, merges score value using the costing bio disturbance layer, parts match score value and component of convolutional neural networks, update volume
The component feature figure of product neural network parameter and each building block updates and exports the component segmentation figure after optimization.
3. image segmentation system according to claim 2, it is characterised in that: feature extraction unit utilizes convolutional Neural net
Convolutional layer, pond layer, the normalization layer of network, extract the feature of input picture, generate multiple characteristic patterns, comprising:
A. convolution operation is carried out to input picture, generates trellis diagram；
B. operation is normalized to trellis diagram, the trellis diagram after generating normalization；
C. pondization operation is carried out to the trellis diagram after normalization, generates characteristic pattern；
D. it is computed repeatedly several times after adjusting the parameter of above-mentioned steps；
When e. reaching preset calculation times, the characteristic pattern generated under multiple groups parameter is exported.
4. image segmentation system according to claim 2, it is characterised in that: component cutting unit utilizes convolutional Neural net
Warp lamination, convolutional layer, costing bio disturbance layer and the multiple characteristic patterns of network, to each building block point of the object in image
It is not split, generates the component feature figure and component segmentation figure of each building block, comprising:
A. deconvolution operation is carried out to characteristic pattern, generates deconvolution figure；
B. operation is normalized to deconvolution figure, the deconvolution figure after generating normalization；
C. up-sampling operation, generating unit characteristic pattern are carried out to the deconvolution figure after normalization；
D. it is computed repeatedly several times after adjusting the parameter of above-mentioned steps；
It when e. training network, repeats the above steps, until loss reaches threshold value；
F. the component feature figure and component segmentation figure of each building block are exported.
5. image segmentation system according to claim 2, it is characterised in that: parts match unit utilizes convolutional Neural net
The warp lamination of network, convolutional layer, costing bio disturbance layer and each building block component feature figure, to the object in image into
The whole segmentation of row, i.e., respectively derived by the component feature figure of each building block and generate the other segmentation figure of a universal class, and according to generation
The other segmentation figure of multiple universal class between similitude, calculate and output block match score value, comprising:
A. convolution operation is carried out to the component feature figure of each building block, generates the trellis diagram of each building block；
B. up-sampling operation is carried out to the trellis diagram of each building block, generates the convolution for each building block that size increases
Figure；
C. the trellis diagram for each building block that size increases is merged；
D. it is computed repeatedly several times after adjusting the parameter of above-mentioned steps；
It when e. training network, repeats the above steps, until loss reaches threshold value；
F. the other segmentation figure of universal class derived respectively by each building block is exported；
G. the similitude of the other segmentation figure of universal class of current part and remaining part between any two is calculated, and by the average value of similitude
Parts match score value as current part exports.
6. image segmentation system according to claim 2, it is characterised in that: component integrated unit utilizes convolutional Neural net
The warp lamination of network, convolutional layer, costing bio disturbance layer and each building block component feature figure, to the object in image into
The component feature figure correspondence of each building block is merged, generates the other segmentation figure of a universal class, and root by the whole segmentation of row
According to the integrality for generating the other segmentation figure of universal class, calculates and output block merges score value, comprising:
A. convolution operation is carried out to the component feature figure of each building block, generates the trellis diagram of each building block；
B. mixing operation is carried out to the trellis diagram of each building block, generates fusion figure；
C. fusion is schemed to carry out convolution operation, generates fusion feature figure；
D. up-sampling operation is carried out to fusion feature figure, generates the fusion feature figure that size increases；
E. it is computed repeatedly several times after adjusting the parameter of above-mentioned steps；
It when f. training network, repeats the above steps, until loss reaches threshold value；
G. the other segmentation figure of universal class merged out by each building block is exported；
H. the integrality of the other segmentation figure of above-mentioned universal class is calculated, combining position information calculates and exports the component fusion point of current part
Value.
7. image segmentation system according to claim 1, it is characterised in that: optimization unit utilizes convolutional neural networks
Costing bio disturbance layer, parts match score value and component merge score value, update the component feature figure of network parameter and each building block,
It updates and exports the component segmentation figure after multiple optimizations, comprising:
A. score value is merged according to parts match score value and component, calculates the overall score value of each building block；
B. using the overall score value of each building block as its weight, the component segmentation figure of each building block is merged,
Fusion segmentation figure is generated, and calculates loss；
C. the whole loss of network is calculated；
D. network parameter is updated；
Component segmentation figure when e. reaching preset calculation times, after output optimization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910083679.8A CN109886990B (en) | 2019-01-29 | 2019-01-29 | Image segmentation system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910083679.8A CN109886990B (en) | 2019-01-29 | 2019-01-29 | Image segmentation system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109886990A true CN109886990A (en) | 2019-06-14 |
| CN109886990B CN109886990B (en) | 2021-04-13 |
Family
ID=66927105
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910083679.8A Active CN109886990B (en) | 2019-01-29 | 2019-01-29 | Image segmentation system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109886990B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106409711A (en) * | 2016-09-12 | 2017-02-15 | 佛山市南海区广工大数控装备协同创新研究院 | Solar silicon wafer defect detecting system and method |
| US20170347110A1 (en) * | 2015-02-19 | 2017-11-30 | Magic Pony Technology Limited | Online Training of Hierarchical Algorithms |
| CN108631727A (en) * | 2018-03-26 | 2018-10-09 | 河北工业大学 | A kind of solar panel defect identification method based on convolutional neural networks |
| CN108765425A (en) * | 2018-05-15 | 2018-11-06 | 深圳大学 | Image partition method, device, computer equipment and storage medium |
| CN109145769A (en) * | 2018-08-01 | 2019-01-04 | 辽宁工业大学 | The target detection network design method of blending image segmentation feature |
-
2019
- 2019-01-29 CN CN201910083679.8A patent/CN109886990B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170347110A1 (en) * | 2015-02-19 | 2017-11-30 | Magic Pony Technology Limited | Online Training of Hierarchical Algorithms |
| CN106409711A (en) * | 2016-09-12 | 2017-02-15 | 佛山市南海区广工大数控装备协同创新研究院 | Solar silicon wafer defect detecting system and method |
| CN108631727A (en) * | 2018-03-26 | 2018-10-09 | 河北工业大学 | A kind of solar panel defect identification method based on convolutional neural networks |
| CN108765425A (en) * | 2018-05-15 | 2018-11-06 | 深圳大学 | Image partition method, device, computer equipment and storage medium |
| CN109145769A (en) * | 2018-08-01 | 2019-01-04 | 辽宁工业大学 | The target detection network design method of blending image segmentation feature |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109886990B (en) | 2021-04-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107123123B (en) | Image segmentation quality evaluating method based on convolutional neural networks | |
| US10062207B2 (en) | Method and system for reconstructing a three-dimensional model of point clouds | |
| CN111144364B (en) | Twin network target tracking method based on channel attention updating mechanism | |
| CN110119728A (en) | Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network | |
| CN109685067A (en) | A kind of image, semantic dividing method based on region and depth residual error network | |
| CN108319972A (en) | A kind of end-to-end difference online learning methods for image, semantic segmentation | |
| CN104008384B (en) | Character identifying method and character recognition device | |
| EP2637139A1 (en) | Method and apparatus for bi-layer segmentation | |
| Elfiky et al. | Automation of dormant pruning in specialty crop production: An adaptive framework for automatic reconstruction and modeling of apple trees | |
| CN109285179A (en) | A kind of motion target tracking method based on multi-feature fusion | |
| CN106981080A (en) | Night unmanned vehicle scene depth method of estimation based on infrared image and radar data | |
| CN109544537A (en) | The fast automatic analysis method of hip joint x-ray image | |
| CN106875373A (en) | Mobile phone screen MURA defect inspection methods based on convolutional neural networks pruning algorithms | |
| CN106951836A (en) | Crop cover degree extracting method based on priori threshold optimization convolutional neural networks | |
| CN103903006A (en) | Crop pest identification method and system based on Android platform | |
| CN109840553A (en) | The extracting method and system, storage medium, electronic equipment for agrotype of ploughing | |
| CN110852393A (en) | Remote sensing image segmentation method and system | |
| CN106570874A (en) | Image marking method combining local image constraint and overall target constraint | |
| CN108038857A (en) | A kind of foreground target detection method based on semantic information and edge constraint | |
| CN110147833A (en) | Facial image processing method, apparatus, system and readable storage medium storing program for executing | |
| CN109978888A (en) | A kind of image partition method, device and computer readable storage medium | |
| CN107563406A (en) | A kind of image sophisticated category method of autonomous learning | |
| CN111260687B (en) | Aerial video target tracking method based on semantic perception network and related filtering | |
| CN103793054B (en) | A kind of action identification method simulating declarative memory process | |
| CN110472608A (en) | Image recognition tracking processing method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |