CN107610140A - Near edge detection method, device based on depth integration corrective networks - Google Patents
Near edge detection method, device based on depth integration corrective networks Download PDFInfo
- Publication number
- CN107610140A CN107610140A CN201710666537.5A CN201710666537A CN107610140A CN 107610140 A CN107610140 A CN 107610140A CN 201710666537 A CN201710666537 A CN 201710666537A CN 107610140 A CN107610140 A CN 107610140A
- Authority
- CN
- China
- Prior art keywords
- feature
- edge detection
- mrow
- detection method
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 89
- 238000003708 edge detection Methods 0.000 title claims abstract description 43
- 230000010354 integration Effects 0.000 title claims abstract description 15
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 37
- 238000004458 analytical method Methods 0.000 claims abstract description 27
- 230000000644 propagated effect Effects 0.000 claims abstract description 26
- 230000006870 function Effects 0.000 claims abstract description 16
- 230000009467 reduction Effects 0.000 claims abstract description 8
- 230000004927 fusion Effects 0.000 claims description 25
- 230000008521 reorganization Effects 0.000 claims description 8
- 238000012549 training Methods 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 4
- 230000001537 neural effect Effects 0.000 claims description 3
- 238000001514 detection method Methods 0.000 abstract description 14
- 238000013135 deep learning Methods 0.000 abstract description 7
- 230000004438 eyesight Effects 0.000 abstract description 5
- 238000012800 visualization Methods 0.000 abstract description 4
- 238000003909 pattern recognition Methods 0.000 abstract description 2
- 238000005070 sampling Methods 0.000 description 15
- 230000008859 change Effects 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 230000004044 response Effects 0.000 description 3
- 238000012937 correction Methods 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241000406668 Loxodonta cyclotis Species 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000002401 inhibitory effect Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000007935 neutral effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The present invention relates to pattern-recognition, computer vision, deep learning field, propose a kind of near edge detection method based on depth integration corrective networks, device, the problem of edge positioning is not accurate enough, the edge of detection is not fine enough in image is aimed to solve the problem that, this method includes:Step S1, by the propagated forward subnetwork of convolutional neural networks, obtain the Analysis On Multi-scale Features of input picture;Step S2, by the reverse amendment subnetwork of convolutional neural networks, the method for gradually increase feature resolution is taken to obtain the final image feature that there is equal resolution with input picture;Step S3, it is single channel by the feature passage dimensionality reduction of final image feature, edge detection results is generated by fitting function.Inventive network structure is simpler, and the image feature representation obtained more retains minutia, and Detection results are more preferable, and edge visualization result is finer.
Description
Technical field
It is more particularly to a kind of to be repaiied based on depth integration the present invention relates to pattern-recognition, computer vision, deep learning field
The near edge detection method of positive network, device.
Background technology
Have benefited from the field fast development such as depth convolutional neural networks, computer vision, artificial intelligence, machine perception.Side
Edge detection is used as a basic problem in computer vision, has also obtained significant progress.Rim detection is exactly to utilize computer
Image is analyzed, and then obtains the marginal information of objects in images.Rim detection is passed through frequently as instrument, aids in other visions
Task.Traditional edge detection method generally relies on engineer's feature, it is easy to changed by light, object color change with
And the interference that background is noisy, very robust, precision also are not difficult to user's request in practice.Based on depth convolutional Neural net
The edge detection method of network overcomes drawbacks described above to a certain extent, there is provided the more preferable edge detector of robustness.
Although depth convolutional neural networks achieve success, or even some high-caliber methods specific in rim detection
Reach on data set manually to the level at image detection edge, but the edge of its detection is not fine enough, it is very coarse, visual
Change in effect, and still there is larger gap at the edge manually extracted.On the other hand, the existing edge based on depth convolutional neural networks
Positioning of the detection method to objects in images edge is not accurate enough, can not obtain fine and accurate positioning edge detection results.
Therefore, the present invention starts with from finer edge is detected, and makes full use of Image Multiscale feature, proposes that one kind is based on depth integration
The near edge detection method of corrective networks, this method is by reversely correcting subnetwork, in gradually fusion propagated forward part
While the Analysis On Multi-scale Features comprising different abstraction hierarchies of network, gradually increase the resolution ratio of feature, it is final to obtain and input
Image has the feature representation of equal resolution.Existing hand-designed characterization method can be so made up well and is only included
The deficiency of the edge detection method based on depth convolutional neural networks of propagated forward subnetwork, obtain a kind of more robust,
The feature representation of abundant detailed information is included, caused edge is more fine.
The content of the invention
It has been to solve edge in image to position not accurate enough, detection to solve above mentioned problem of the prior art
The problem of edge is not fine enough, an aspect of of the present present invention, it is proposed that a kind of near edge inspection based on depth integration corrective networks
Survey method, comprises the following steps:
Step S1, by the propagated forward subnetwork of convolutional neural networks, obtain the Analysis On Multi-scale Features of input picture;
Step S2, by the reverse amendment subnetwork of convolutional neural networks, take the side for gradually increasing feature resolution
Method obtains the final image feature for having equal resolution with input picture;It is described gradually increase feature resolution method be:
While merging the Analysis On Multi-scale Features of the input picture, the resolution ratio of feature is increased by the method for sub-pixel convolution;
Step S3, it is single channel by the feature passage dimensionality reduction of final image feature, rim detection is generated by fitting function
As a result.
Preferably, the Analysis On Multi-scale Features of input picture described in step S1, it is the different chis that the convolutional layer of different depth obtains
The feature pyramid that characteristics of image under degree is formed.
Preferably, " take the method for gradually increase feature resolution to obtain has identical point with input picture in step S2
The final image feature of resolution ", its method are:
To all features in the feature pyramid, top-down selection adjacent feature, repeat step S21 and step
S22, until obtaining the fusion feature F ' of first layer feature1;
Wherein,
Step S21, give feature FnAnd Fn-1, by the less feature F of yardstickn, pass through the method increase point of sub-pixel convolution
Resolution obtains feature Fαn, feature FαnYardstick and Fn-1Unanimously;Wherein feature FnFor the characteristics of image of n-th layer in Analysis On Multi-scale Features;
Step S22:By convolution, concatenation by feature FαnAnd Fn-1Merged, produce (n-1)th layer of fusion feature
Fβn-1, and be F by (n-1)th layer in the feature pyramid of feature replacementβn-1。
Preferably, in step S22 " by convolution, concatenation by feature FαnAnd Fn-1Merged ", its method includes:
Step S221, by convolution operation, by FαnAnd Fn-1Dimensionality reduction is carried out, the feature that furthers port number, obtains feature representation
F′αnWith F 'n-1, size is respectively wn-1×hn-1×d′αnAnd wn-1×hn-1×d′n-1;
Step S222, by F 'αnWith F 'n-1Splicing, generation size is wn-1×hn-1×(d′αn+d′n-1) intermediate features, then
The intermediate features are used with convolution operation, generation size is wn-1×hn-1×dβn-1(n-1)th layer of fusion feature Fβn-1, and
It is F by (n-1)th layer in the feature pyramid of feature replacementβn-1。
Preferably, " given feature F described in step S21nAnd Fn-1, by the less feature F of yardstickn, pass through sub-pixel convolution
Method increase resolution ratio obtain feature Fαn", its method is:
Step S211, by feature FnCarry out multilayer convolution operation, generation new feature F "n, wherein F "nSize be wn
×hn×(dn×r×r);Wherein FnSize be wn×hn×dn, r is characterized Fn-1And FnScale size multiple;
Step S212, to new feature F "nPixel reorganization is carried out in a fixed order, and it is rw to obtain sizen×rhn×dn's
Feature Fαn。
Preferably, " to new feature F " in step S212nPixel reorganization is carried out in a fixed order ", its method is:
By wn×hnD on the two-dimensional coordinate of determinationn× r × r value is divided into r2Part, then by this r2Part is laid into r × r area
Domain, each point in the region contain dnIndividual passage.
Preferably, convolutional neural networks are same convolutional Neural net in convolutional neural networks and step S2 in step S1
Network, the convolutional neural networks are by merging the propagated forward subnetwork, the reversely amendment subnetwork is formed
Can end-to-end training near edge detection convolutional neural networks.
Preferably, it is described can end-to-end training near edge detection convolutional neural networks, its train loss function L
(w) it is:
Wherein, X represent input picture, W be propagated forward subnetwork weight, Y+Represent edge pixel point set, Y_Table
Show non-edge pixels point set, β is the ratio of non-edge pixels point sum and edge pixel point sum.
Preferably, fitting function described in step S3 is Sigmoid functions.
Another aspect of the present invention, it is proposed that a kind of storage device, wherein be stored with a plurality of program, described program be suitable to by
Processor is loaded and performed to realize the above-mentioned near edge detection method based on depth integration corrective networks.
The third aspect of the present invention, it is proposed that a kind of processing unit, including
Processor, it is adapted for carrying out each bar program;And
Storage device, suitable for storing a plurality of program;
Described program is above-mentioned based on the fine of depth integration corrective networks to realize suitable for being loaded by processor and being performed
Edge detection method.
The present invention has advantages below compared to existing edge detection method:
(1) Analysis On Multi-scale Features of propagated forward subnetwork are merged by reversely correcting subnetwork, and reversely corrected
Subnetwork increases the resolution ratio of feature when merging Analysis On Multi-scale Features, by the method for sub-pixel convolution, finally obtains and defeated
Enter the feature representation of image equal resolution, the up-sampling mode than traditional bilinear interpolation more retains minutia;
(2) propagated forward subnetwork and the reversely convolutional neural networks of the U-shaped network structure of amendment subnetwork composition,
Compared to the current existing rim detection network structure based on deep learning, network structure is simpler, and Detection results are more preferable, edge
Visualization result is finer.
Brief description of the drawings
Fig. 1 is the schematic flow sheet of near edge detection method of the embodiment of the present invention based on depth integration corrective networks;
Fig. 2 is the structural representation of the convolutional neural networks of the near edge detection of the embodiment of the present invention;
Fig. 3 is the edge detection method result visualization legend of the embodiment of the present invention.
Embodiment
The preferred embodiment of the present invention described with reference to the accompanying drawings.It will be apparent to a skilled person that this
A little embodiments are used only for explaining the technical principle of the present invention, it is not intended that limit the scope of the invention.
The present invention thought main points be:1) depth integration corrective networks network structure proposed by the present invention utilizes reversely amendment
Subnetwork gradually merges the feature of the different scale of propagated forward subnetwork;2) reversely correction portion subnetting proposed by the present invention
Network is when merging propagated forward subnetwork particular dimensions feature and current amendment feature, after reducing feature port number, using spelling
The mode connect is merged;3) reversely amendment subnetwork proposed by the present invention is in fusion propagated forward subnetwork different scale
During feature, gradually increase the resolution ratio of characteristic response, it is final to obtain and an equal amount of characteristic response of input picture;4) it is of the invention
Merge above-mentioned thought devise can end-to-end training depth convolutional neural networks structure.
The near edge detection method based on depth integration corrective networks of one embodiment of the invention, as shown in figure 1, including
Following steps:
Step S1, by the propagated forward subnetwork of convolutional neural networks, obtain the Analysis On Multi-scale Features of input picture;
Step S2, by the reverse amendment subnetwork of convolutional neural networks, take the side for gradually increasing feature resolution
Method obtains the final image feature for having equal resolution with input picture;It is described gradually increase feature resolution method be:
While merging the Analysis On Multi-scale Features of the input picture, the resolution ratio of feature is increased by the method for sub-pixel convolution;
Step S3, it is single channel by the feature passage dimensionality reduction of final image feature, rim detection is generated by fitting function
As a result.
In the embodiment of the present invention, convolutional neural networks are same convolution in convolutional neural networks and step S2 in step S1
Neutral net, as shown in Fig. 2 the convolutional neural networks are by merging the propagated forward subnetwork, the reversely amendment
Subnetwork formed can end-to-end training near edge detection convolutional neural networks, for vivider expression network
Structure, it is properly termed as U-typed convolutional neural networks.
The " u "-shaped convolutional neural networks obtain the multiple dimensioned spy of different abstract levels by the propagated forward subnetwork
Sign, then by the Analysis On Multi-scale Features reversely corrected subnetwork and gradually merge the propagated forward subnetwork, while not
Disconnected increase feature resolution, it is final to obtain the feature that there is equal resolution with input picture.
It is above-mentioned can end-to-end training near edge detection convolutional neural networks, its loss function L (w) trained is as public
Shown in formula (1):
Wherein, X represent input picture, W be propagated forward subnetwork weight, Y+Represent edge pixel point set, Y-Table
Showing non-edge pixels point set, β is the ratio of non-edge pixels point sum and edge pixel point sum, and Pr represents conditional probability,
Pr (A | B) is the conditional probability of A after known B occurs, also due to the value derived from B and the posterior probability referred to as A, yj=1 table
Show that j-th of pixel is edge pixel point, yj=0 represents that j-th of pixel is non-edge pixels point.
The Analysis On Multi-scale Features of input picture in step S1, it is the difference that the convolutional layer of different depth obtains in the present embodiment
The feature pyramid that characteristics of image under yardstick is formed.
In the present embodiment, " method of gradually increase feature resolution is taken to obtain and input picture tool described in step S2
Have the final image feature of equal resolution ", its method is:
To all features in the feature pyramid, top-down selection adjacent feature, repeat step S21 and step
S22, until obtaining the fusion feature F ' of first layer feature1;
Wherein,
Step S21, up-sampling:Given feature FnAnd Fn-1, by the less feature F of yardstickn, pass through the method for sub-pixel convolution
Increase resolution ratio obtains feature Fαn, feature FαnYardstick and Fn-1Unanimously;Wherein feature FnFor n-th layer in the Analysis On Multi-scale Features
Characteristics of image;
Step S22:Merge Analysis On Multi-scale Features:By convolution, concatenation by feature FαnAnd Fn-1Merged, produce the
The fusion feature F of n-1 layersβn-1, and be F by (n-1)th layer in feature pyramid of feature replacementβn-1。
In the present embodiment, " by convolution, concatenation by feature F in step S22αnAnd Fn-1Merged ", its method bag
Include:
Step S221, by convolution operation, by FαnAnd Fn-1Dimensionality reduction is carried out, the feature that furthers port number, obtains feature representation
F′αnWith F 'n-1, size is respectively wn-1×hn-1×d′αnAnd wn-1×hn-1×d′n-1;
Step S222, by F 'αnWith F 'n-1Splicing, generation size is wn-1×hn-1×(d′αn+d′n-1) intermediate features, then
The intermediate features are used with convolution operation, generation size is wn-1×hn-1×dβn-1(n-1)th layer of fusion feature Fβn-1, and
It is F by (n-1)th layer in the feature pyramid of feature replacementβn-1。
In the present embodiment, " given feature F in step S21nAnd Fn-1, by the less feature F of yardstickn, pass through sub-pixel convolution
Method increase resolution ratio obtain feature Fαn", its method is:
Step S211, by feature FnCarry out multilayer convolution operation, generation new feature F "n, wherein F "nSize be wn
×hn×(dn×r×r);Wherein FnSize be wn×hn×dn, r is characterized Fn-1And FnScale size multiple;
Step S212, to new feature F "nPixel reorganization is carried out in a fixed order, and it is rw to obtain sizei×rhi×di's
Feature Fαn, specific method is:By wn×hnD on the two-dimensional coordinate of determinationn× r × r value is divided into r2Part, then by this r2Part is flat
R × r region is paved into, each point in the region contains dnIndividual passage.
In the present embodiment, fitting function described in step S3 is Sigmoid functions.
In order to further clearly be illustrated to the technology of the present invention details, hereafter carried respectively from forward direction part of propagation network
Take Analysis On Multi-scale Features, reversely amendment subnetwork fusion Analysis On Multi-scale Features, reverse amendment subnetwork pass through in sub-pixel convolution
Sampling three parts are developed in details.
1st, propagated forward subnetwork extraction Analysis On Multi-scale Features
Characteristics of image is the abstractdesription to image, can representative image the characteristics of.The quality of feature directly affects side
Edge testing result.The method that traditional feature extraction typically uses hand-designed, it is difficult to the deep information of description image very well, and
It is and easily poor by the interference of the conditions such as such as illumination, angle, robustness.Different from traditional-handwork design feature method, it is based on
The mode of deep learning is learnt end to end, can be according to optimization aim, by error back propagation method, from a large amount of numbers
Feature representation according to learning to data robust.Therefore, this method employs the mode based on deep learning, to obtain input figure
As more preferable feature representation.This method is first using a series of propagated forward subnetwork knot operated comprising convolution and down-sampling
Structure, obtain the feature representation under the multiple different scales of image.
Specifically, the propagated forward subnetwork structure of this method uses depth convolutional neural networks, by image
A series of convolution and down-sampling operation are carried out, the convolutional layer of different depth can be obtained under different scale (as shown in Figure 2
1,1/2,1/4,1/8 of original image resolution etc. respectively) image feature representation, composition characteristic pyramid.It is represented by
Formula (2):
Wherein, X represents input picture, and W is the weight of propagated forward subnetwork,Represent not
With the feature of yardstick.
Wherein, more rich space characteristics being included compared with shallow-layer feature, deeper feature abstraction ability is stronger, therefore comprising more
Abundant semantic feature.These different abstraction hierarchies, the multiple dimensioned basis for being characterized in subsequent characteristics amendment.Spy in the present invention
It is a three-dimensional matrice R to levy expressionw×h×d, the resolution ratio of w and h Expressive Features, width and height are represented respectively, d represents special in addition
Levy port number.
2nd, reversely amendment subnetwork merges Analysis On Multi-scale Features
, can be although the feature that traditional side edge detection calligraphy learning based on deep learning obtains has good presentation skills
To a certain extent response diagram as the characteristics of, but still deficiency with for fine edge provides enrich detailed information.Therefore, we
On propagated forward network foundation of the tradition based on deep learning method, reversely amendment subnetwork is introduced, before preferably merging
To the Analysis On Multi-scale Features of part of propagation network, to obtain the preferably feature representation of balanced semantic feature and space characteristics.Meanwhile
Reversely amendment subnetwork gradually increases the resolution of feature in the gradually Analysis On Multi-scale Features of fusion propagated forward subnetwork
Rate, it is final to obtain the feature representation that there is equal resolution with input picture.It is represented by formula (3):
WhereinThe feature under propagated forward subnetwork current scale is represented,Represent repercussion correction portion
The upper revised feature of a yardstick, f in subnetworkfuse() represents to be merged both features, fsubConv() represents to melt
Feature after conjunction increases feature resolution in a manner of sub-pixel convolution.The feature representation so obtained is for producing fine side
Edge is particularly important.
In embodiments of the present invention, F is usednRepresent the feature of n-layer in feature pyramid.The characteristic modification step of the present invention
(the step of merging Analysis On Multi-scale Features), is top-down, i.e., since top-level feature.It is w for sizen×hn×dn's
N-th layer feature, i.e. top feature Fn, use it to (n-1)th layer of feature F of amendmentn-1.Therefore, first to FnCarry out convolution and upper
Sampling operation, it is w to obtain sizen-1×hn-1×dnFeature representation Fαn, wherein wn-1, hn-1It is consistent with the w of n-1 layer features, h.
Although by up-sampling operation by FnScale expansion most with Fn-1Yardstick it is consistent, but due to the feature of different levels
Contained feature port number is inconsistent, if directly to feature FαnAnd Fn-1Merged, can (feature passage be few to low-dimensional feature
Feature) produce inhibitory action, it is difficult to due ability to express is played in the feature after fusion.In order to more preferably retain each level
Feature generate new feature F ', it is necessary to pass through convolutional layerαnWith F 'n-1To express primitive character FαnAnd Fn-1.New feature and original spy
The yardstick of sign is the same, and the feature port number of feature to be fused approaches, and will not produce suppression to low-dimensional feature during Fusion Features.
Specifically, by convolution operation, by feature (i.e. F to be fusedαnAnd Fn-1) dimensionality reduction is carried out, the feature that furthers port number, obtain feature
Express F 'αnWith F 'n-1, size is respectively wn-1×hn-1×d′αnAnd wn-1×hn-1×d′n-1。
Obtain F 'αnWith F 'n-1Afterwards, it is necessary to use the F ' containing more global informationsαnInstruct F 'n-1Produce fusion feature Fβn-1。
In order to improve fusion after feature ability to express, on the basis of splicing, reuse convolution operation, feature after splicing taken out
As so as to produce revised fusion feature Fβn-1.Specifically, we are first by F 'αnWith F 'n-1Splicing, producing size is
wn-1×hn-1×(d′αn+d′n-1) intermediate features, then convolution operation is used to it, generation size is wn-1×hn-1×dβn-1's
Fusion feature Fβn-1, the mistake with top-level feature amendment time top-level feature is so far completed by way of merging Analysis On Multi-scale Features
Journey.
For n-2 layer features Fn-2, the fusion feature F of the invention implemented using n-1 layersβn-1Analysis On Multi-scale Features are carried out to it
Fusion, realize to feature Fn-2Amendment.Specific steps are similar to the characteristic modification to n-1 layers, and the present invention is first by Fβn-1On
Sample Fn-2Yardstick, afterwards according to the obtained F of up-samplingαn-1And Fn-2, difference convolution generation new feature F 'αn-1With F 'n-2.Most
Afterwards, by F 'αn-1With F 'n-2Generate fusion feature Fβn-2。
Said process is repeated, until completing to correct bottom feature F1, generation size is w1×h1×d1Feature F '1
(w1And h1Equal to the size of input picture).In general, the feature port number of edge graph is 1, therefore, we are needed to feature
F′1Convolution operation is carried out, its feature passage is reduced to 1, F is obtained by Sigmoid functionsout, i.e. the rim detection knot of this method
Fruit.
3rd, reversely amendment subnetwork is up-sampled by sub-pixel convolution
Edge detection method based on depth convolutional neural networks uses a series of convolution and down-sampling layer, more to be taken out
The feature representation of elephant, but because a series of down-samplings operate, the resolution ratio of feature representation is gradually lowered.It is end-to-end in order to carry out
Study, the edge detection method based on depth convolutional neural networks generally require by resolution ratio reduce feature be upsampled to it is defeated
Enter the original resolution ratio of image.Traditional edge detection method based on depth convolutional neural networks is by way of bilinearity difference
Feature representation is directly up-sampled to increase feature resolution, the top sampling method of this bilinear interpolation can lose details letter
Breath, this is unfavorable to obtaining finer feature representation.In order to obtain and input image resolution identical feature representation, sheet
Method takes the mode of gradually increase feature resolution.The reverse amendment subnetwork of this method is in gradually fusion propagated forward portion
During the Analysis On Multi-scale Features of subnetwork, gradually increase the resolution ratio of feature, final obtain has equal resolution with input picture
Feature representation.Wherein, mode of the feature resolution using sub-pixel convolution is increased.
Given size is respectively wn×hn×dnAnd wn-1×hn-1×dn-1Feature FnAnd Fn-1, the purpose of sub-pixel convolution
It is to increase FnFeature resolution, it is set to be converted into and Fn-1Feature F with same scaleαn, wherein FαnSize be wn-1×
hn-1×dn-1.The top sampling method for being conventionally used to increase resolution ratio is usually the wave filter of engineer, such as bilinearity or double three
Secondary sampler.The wave filter of engineer uses the up-sampling mode of bilinear interpolation, it is excessively smooth to up-sample result, it is difficult to protect
Spacing detailed information, and sub-pixel convolution constantly updates parameter by the study of mass data, up-sampling result retains richer
Rich spatial information.
Different from other up-sampling modes based on study, sub-pixel convolution includes multilayer convolution and pixel reorganization two
Point.The purpose of multilayer convolution is to be abstracted input feature vector from many levels, stronger so as to generate feature representation ability
New feature.For feature FnAnd Fn-1If Fn-1Yardstick be FnR times, i.e. wn-1=r × wn, hn-1=r × hn.We pass through
To FnCarry out multilayer convolution operation, generation new feature F "n.Wherein F "nSize be wn×hn×(dn×r×r).New feature
F″nWith FnCompare, maintain the uniformity of yardstick, only change feature port number in proportion.Obtain new feature F "nAfterwards, according still further to
Permanent order carries out pixel reorganization operation, makes F "nSize and Fn-1Match (i.e. F "nSize reaches wn-1And hn-1).For three
Dimensional feature matrix F "n, have d on the two-dimensional coordinate that is determined by a certain group of w and hn× r × r value.These values are divided into r2Part, then will
This r2Part is laid into r × r region, and each point in the region contains diIndividual passage.To F "nAfter tiling in a manner described, obtain
Size is rwn×rhn×dnFeature Fαn.After pixel reorganization, obtain and Fn-1The feature F of yardstick matchingαn, this feature can
To regard feature F asnMapping in target signature space.Sub-pixel convolution operation is used in addition to top feature by the present invention
The upsampling process of all hierarchy characteristics, it is final to obtain the feature representation that there is equal resolution with input picture.
Increase feature resolution by way of sub-pixel convolution, its detailed explanation may be referred to Wenzhe Shi and exist
2016 entitled in having delivered one on CVPR《By efficient sub-pixel convolutional neural networks realize real-time single image and
Video super-resolution (Real-Time Single Image and Video Super-Resolution Using an
Efficient Sub-Pixel Convolutional Neural Network)》Paper, be no longer developed in details herein
It is bright.
Fig. 3 is the edge detection method result visualization legend of the present embodiment.First behavior image to be detected, the second row
The edge detection method result designed for the standard edge testing result manually marked, the third line for the present invention.
The method that is described with reference to the embodiments described herein can use hardware, computing device the step of algorithm
Software module, or the two combination are implemented.Software module can be placed in random access memory (RAM), internal memory, read-only storage
(ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technical field
In any other form of storage medium well known to interior.
Those skilled in the art should be able to recognize that, the side of each example described with reference to the embodiments described herein
Method step, can be realized with electronic hardware, computer software or the combination of the two, in order to clearly demonstrate electronic hardware and
The interchangeability of software, the composition and step of each example are generally described according to function in the above description.These
Function is performed with electronic hardware or software mode actually, application-specific and design constraint depending on technical scheme.
Those skilled in the art can realize described function using distinct methods to each specific application, but this reality
Now it is not considered that beyond the scope of this invention.
The storage device of an embodiment of the present invention, wherein being stored with a plurality of program, described program is suitable to be added by processor
Carry and perform to realize the above-mentioned near edge detection method based on depth integration corrective networks.
The processing unit of an embodiment of the present invention, including processor, storage device;Processor, it is adapted for carrying out each bar journey
Sequence;Storage device, suitable for storing a plurality of program;Described program is above-mentioned based on depth to realize suitable for being loaded by processor and being performed
The near edge detection method of degree fusion corrective networks.
Person of ordinary skill in the field can be understood that, for convenience and simplicity of description, foregoing description
Storage device, the specific work process of processing unit and relevant explanation, may be referred to the corresponding process in preceding method embodiment,
It will not be repeated here.
So far, combined preferred embodiment shown in the drawings describes technical scheme, still, this area
Technical staff is it is easily understood that protection scope of the present invention is expressly not limited to these embodiments.Without departing from this
On the premise of the principle of invention, those skilled in the art can make equivalent change or replacement to correlation technique feature, these
Technical scheme after changing or replacing it is fallen within protection scope of the present invention.
Claims (11)
1. a kind of near edge detection method based on depth integration corrective networks, it is characterised in that comprise the following steps:
Step S1, by the propagated forward subnetwork of convolutional neural networks, obtain the Analysis On Multi-scale Features of input picture;
Step S2, by the reverse amendment subnetwork of convolutional neural networks, the method for gradually increase feature resolution is taken to obtain
Take the final image feature that there is equal resolution with input picture;It is described gradually increase feature resolution method be:Fusion
While the Analysis On Multi-scale Features of the input picture, the resolution ratio of feature is increased by the method for sub-pixel convolution;
Step S3, it is single channel by the feature passage dimensionality reduction of final image feature, edge detection results is generated by fitting function.
2. near edge detection method according to claim 1, it is characterised in that more chis of input picture described in step S1
Feature is spent, the feature pyramid being made up of the characteristics of image under different scale that the convolutional layer of different depth obtains.
3. near edge detection method according to claim 1, it is characterised in that " take gradually increase special in step S2
The method of sign resolution ratio obtains the final image feature for having equal resolution with input picture ", its method is:
To all features in the feature pyramid, top-down selection adjacent feature, repeat step S21 and step S22,
Until obtaining the fusion feature F of first layer feature1′;
Wherein,
Step S21, give feature FnAnd Fn-1, by the less feature F of yardstickn, resolution ratio is increased by the method for sub-pixel convolution
Obtain feature Fαn, feature FαnYardstick and Fn-1Unanimously;Wherein feature FnFor the characteristics of image of n-th layer in Analysis On Multi-scale Features;
Step S22:By convolution, concatenation by feature FαnAnd Fn-1Merged, produce (n-1)th layer of fusion feature Fβn-1,
And by (n-1)th layer in the feature pyramid of feature replacement it is Fβn-1。
4. near edge detection method according to claim 3, it is characterised in that " pass through convolution, splicing in step S22
Operate feature FαnAnd Fn-1Merged ", its method includes:
Step S221, by convolution operation, by FαnAnd Fn-1Dimensionality reduction is carried out, the feature that furthers port number, obtains feature representation F 'αnWith
F′n-1, size is respectively wn-1×hn-1×d′αnAnd wn-1×hn-1×d′n-1;
Step S222, by F 'αnWith F 'n-1Splicing, generation size is wn-1×hn-1×(d′αn+d′n-1) intermediate features, then to institute
State intermediate features and use convolution operation, generation size is wn-1×hn-1×dβn-1(n-1)th layer of fusion feature Fβn-1, and by institute
The feature replacement for stating (n-1)th layer in feature pyramid is Fβn-1。
5. near edge detection method according to claim 3, it is characterised in that " given feature F described in step S21n
And Fn-1, by the less feature F of yardstickn, resolution ratio is increased by the method for sub-pixel convolution and obtains feature Fαn", its method is:
Step S211, by feature FnCarry out multilayer convolution operation, generation new feature Fn", wherein Fn" size be wn×hn×
(dn×r×r);Wherein FnSize be wn×hn×dn, r is characterized Fn-1And FnScale size multiple;
Step S212, to new feature Fn" pixel reorganization is carried out in a fixed order, it is rw to obtain sizen×rhn×dnFeature
Fαn。
6. near edge detection method according to claim 5, it is characterised in that " to new feature F in step S212n" press
Pixel reorganization is carried out according to permanent order ", its method is:
By wn×hnD on the two-dimensional coordinate of determinationn× r × r value is divided into r2Part, then by this r2Part is laid into r × r region,
Each point in the region contains dnIndividual passage.
7. according to the near edge detection method described in claim any one of 1-6, it is characterised in that convolutional Neural in step S1
Convolutional neural networks are same convolutional neural networks in network and step S2, the convolutional neural networks be by merge it is described before
To part of propagation network, it is described reversely amendment subnetwork formed can end-to-end training near edge detect convolution god
Through network.
8. near edge detection method according to claim 7, it is characterised in that it is described can end-to-end training fine side
The convolutional neural networks of edge detection, its loss function L (w) trained are:
<mrow>
<mi>L</mi>
<mrow>
<mo>(</mo>
<mi>w</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>-</mo>
<mi>&beta;</mi>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>&Element;</mo>
<msub>
<mi>Y</mi>
<mo>+</mo>
</msub>
</mrow>
</munder>
<mi>log</mi>
<mi>Pr</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>=</mo>
<mn>1</mn>
<mo>|</mo>
<mi>X</mi>
<mo>,</mo>
<mi>w</mi>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<mi>&beta;</mi>
<mo>)</mo>
</mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>&Element;</mo>
<msub>
<mi>Y</mi>
<mo>-</mo>
</msub>
</mrow>
</munder>
<mi>log</mi>
<mi>Pr</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>=</mo>
<mn>0</mn>
<mo>|</mo>
<mi>X</mi>
<mo>,</mo>
<mi>w</mi>
<mo>)</mo>
</mrow>
</mrow>
Wherein, X represent input picture, W be propagated forward subnetwork weight, Y+Represent edge pixel point set, Y-Represent non-
Edge pixel point set, β are the ratios of non-edge pixels point sum and edge pixel point sum.
9. according to the near edge detection method described in claim any one of 1-6, it is characterised in that be fitted described in step S3
Function is Sigmoid functions.
10. a kind of storage device, wherein being stored with a plurality of program, it is characterised in that described program is suitable to by processor loading simultaneously
Perform to realize the near edge detection method based on depth integration corrective networks described in claim any one of 1-9.
11. a kind of processing unit, including
Processor, it is adapted for carrying out each bar program;And
Storage device, suitable for storing a plurality of program;
Characterized in that, described program is suitable to be loaded by processor and performed to realize:
The near edge detection method based on depth integration corrective networks described in claim any one of 1-9.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710666537.5A CN107610140A (en) | 2017-08-07 | 2017-08-07 | Near edge detection method, device based on depth integration corrective networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710666537.5A CN107610140A (en) | 2017-08-07 | 2017-08-07 | Near edge detection method, device based on depth integration corrective networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107610140A true CN107610140A (en) | 2018-01-19 |
Family
ID=61064751
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710666537.5A Pending CN107610140A (en) | 2017-08-07 | 2017-08-07 | Near edge detection method, device based on depth integration corrective networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107610140A (en) |
Cited By (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108805889A (en) * | 2018-05-07 | 2018-11-13 | 中国科学院自动化研究所 | The fining conspicuousness method for segmenting objects of margin guide and system, equipment |
CN108830196A (en) * | 2018-05-31 | 2018-11-16 | 上海贵和软件技术有限公司 | Pedestrian detection method based on feature pyramid network |
CN108830790A (en) * | 2018-05-16 | 2018-11-16 | 宁波大学 | It is a kind of based on the fast video super resolution ratio reconstruction method for simplifying convolutional neural networks |
CN108898609A (en) * | 2018-06-21 | 2018-11-27 | 深圳辰视智能科技有限公司 | A kind of method for detecting image edge, detection device and computer storage medium |
CN109035251A (en) * | 2018-06-06 | 2018-12-18 | 杭州电子科技大学 | One kind being based on the decoded image outline detection method of Analysis On Multi-scale Features |
CN109118504A (en) * | 2018-07-26 | 2019-01-01 | 深圳辰视智能科技有限公司 | A kind of method for detecting image edge neural network based, device and its equipment |
CN109472801A (en) * | 2018-11-22 | 2019-03-15 | 廖祥 | It is a kind of for multiple dimensioned neuromorphic detection and dividing method |
CN109741351A (en) * | 2018-12-12 | 2019-05-10 | 中国科学院深圳先进技术研究院 | A kind of classification responsive type edge detection method based on deep learning |
CN109816037A (en) * | 2019-01-31 | 2019-05-28 | 北京字节跳动网络技术有限公司 | The method and apparatus for extracting the characteristic pattern of image |
CN110033469A (en) * | 2019-04-01 | 2019-07-19 | 北京科技大学 | A kind of sub-pixel edge detection method and system |
CN110399900A (en) * | 2019-06-26 | 2019-11-01 | 腾讯科技(深圳)有限公司 | Method for checking object, device, equipment and medium |
WO2019227954A1 (en) * | 2018-05-31 | 2019-12-05 | 京东方科技集团股份有限公司 | Method and apparatus for identifying traffic light signal, and readable medium and electronic device |
CN110648316A (en) * | 2019-09-07 | 2020-01-03 | 创新奇智(成都)科技有限公司 | Steel coil end face edge detection algorithm based on deep learning |
CN110908566A (en) * | 2018-09-18 | 2020-03-24 | 珠海格力电器股份有限公司 | Information processing method and device |
CN110987189A (en) * | 2019-11-21 | 2020-04-10 | 北京都是科技有限公司 | Method, system and device for detecting temperature of target object |
CN111582353A (en) * | 2020-04-30 | 2020-08-25 | 恒睿(重庆)人工智能技术研究院有限公司 | Image feature detection method, system, device and medium |
JPWO2020225879A1 (en) * | 2019-05-08 | 2020-11-12 | ||
JPWO2020225880A1 (en) * | 2019-05-08 | 2020-11-12 | ||
US10977798B2 (en) | 2018-08-24 | 2021-04-13 | Apple Inc. | Direct thin boundary prediction |
CN113132755A (en) * | 2019-12-31 | 2021-07-16 | 北京大学 | Extensible man-machine cooperative image coding method and coding system |
CN113240584A (en) * | 2021-05-11 | 2021-08-10 | 上海大学 | Multitask gesture picture super-resolution method based on picture edge information |
CN113538484A (en) * | 2021-07-01 | 2021-10-22 | 广西科技大学 | Deep-refinement multiple-information nested edge detection method |
WO2022267046A1 (en) * | 2021-06-25 | 2022-12-29 | 京东方科技集团股份有限公司 | Un-decimated image processing method and apparatus |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105120130A (en) * | 2015-09-17 | 2015-12-02 | 京东方科技集团股份有限公司 | Image ascending frequency system and training method and image ascending frequency method thereof |
CN106340036A (en) * | 2016-08-08 | 2017-01-18 | 东南大学 | Binocular stereoscopic vision-based stereo matching method |
CN106898011A (en) * | 2017-01-06 | 2017-06-27 | 广东工业大学 | A kind of method that convolutional neural networks convolution nuclear volume is determined based on rim detection |
CN106934765A (en) * | 2017-03-14 | 2017-07-07 | 长沙全度影像科技有限公司 | Panoramic picture fusion method based on depth convolutional neural networks Yu depth information |
CN106952229A (en) * | 2017-03-15 | 2017-07-14 | 桂林电子科技大学 | Image super-resolution rebuilding method based on the enhanced modified convolutional network of data |
-
2017
- 2017-08-07 CN CN201710666537.5A patent/CN107610140A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105120130A (en) * | 2015-09-17 | 2015-12-02 | 京东方科技集团股份有限公司 | Image ascending frequency system and training method and image ascending frequency method thereof |
CN106340036A (en) * | 2016-08-08 | 2017-01-18 | 东南大学 | Binocular stereoscopic vision-based stereo matching method |
CN106898011A (en) * | 2017-01-06 | 2017-06-27 | 广东工业大学 | A kind of method that convolutional neural networks convolution nuclear volume is determined based on rim detection |
CN106934765A (en) * | 2017-03-14 | 2017-07-07 | 长沙全度影像科技有限公司 | Panoramic picture fusion method based on depth convolutional neural networks Yu depth information |
CN106952229A (en) * | 2017-03-15 | 2017-07-14 | 桂林电子科技大学 | Image super-resolution rebuilding method based on the enhanced modified convolutional network of data |
Non-Patent Citations (2)
Title |
---|
WENZHE SHI.ETC: "Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
YUPEI WANG.ETC: "Deep Crisp Boundaries", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
Cited By (35)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108805889B (en) * | 2018-05-07 | 2021-01-08 | 中国科学院自动化研究所 | Edge-guided segmentation method, system and equipment for refined salient objects |
CN108805889A (en) * | 2018-05-07 | 2018-11-13 | 中国科学院自动化研究所 | The fining conspicuousness method for segmenting objects of margin guide and system, equipment |
CN108830790B (en) * | 2018-05-16 | 2022-09-13 | 宁波大学 | Rapid video super-resolution reconstruction method based on simplified convolutional neural network |
CN108830790A (en) * | 2018-05-16 | 2018-11-16 | 宁波大学 | It is a kind of based on the fast video super resolution ratio reconstruction method for simplifying convolutional neural networks |
CN108830196A (en) * | 2018-05-31 | 2018-11-16 | 上海贵和软件技术有限公司 | Pedestrian detection method based on feature pyramid network |
US11410549B2 (en) | 2018-05-31 | 2022-08-09 | Boe Technology Group Co., Ltd. | Method, device, readable medium and electronic device for identifying traffic light signal |
WO2019227954A1 (en) * | 2018-05-31 | 2019-12-05 | 京东方科技集团股份有限公司 | Method and apparatus for identifying traffic light signal, and readable medium and electronic device |
CN109035251A (en) * | 2018-06-06 | 2018-12-18 | 杭州电子科技大学 | One kind being based on the decoded image outline detection method of Analysis On Multi-scale Features |
CN109035251B (en) * | 2018-06-06 | 2022-05-27 | 杭州电子科技大学 | Image contour detection method based on multi-scale feature decoding |
CN108898609A (en) * | 2018-06-21 | 2018-11-27 | 深圳辰视智能科技有限公司 | A kind of method for detecting image edge, detection device and computer storage medium |
CN109118504A (en) * | 2018-07-26 | 2019-01-01 | 深圳辰视智能科技有限公司 | A kind of method for detecting image edge neural network based, device and its equipment |
CN109118504B (en) * | 2018-07-26 | 2021-03-30 | 深圳辰视智能科技有限公司 | Image edge detection method, device and equipment based on neural network |
US10977798B2 (en) | 2018-08-24 | 2021-04-13 | Apple Inc. | Direct thin boundary prediction |
CN110908566A (en) * | 2018-09-18 | 2020-03-24 | 珠海格力电器股份有限公司 | Information processing method and device |
CN109472801A (en) * | 2018-11-22 | 2019-03-15 | 廖祥 | It is a kind of for multiple dimensioned neuromorphic detection and dividing method |
CN109741351A (en) * | 2018-12-12 | 2019-05-10 | 中国科学院深圳先进技术研究院 | A kind of classification responsive type edge detection method based on deep learning |
CN109816037A (en) * | 2019-01-31 | 2019-05-28 | 北京字节跳动网络技术有限公司 | The method and apparatus for extracting the characteristic pattern of image |
CN110033469A (en) * | 2019-04-01 | 2019-07-19 | 北京科技大学 | A kind of sub-pixel edge detection method and system |
CN110033469B (en) * | 2019-04-01 | 2021-08-27 | 北京科技大学 | Sub-pixel edge detection method and system |
JPWO2020225879A1 (en) * | 2019-05-08 | 2020-11-12 | ||
JPWO2020225880A1 (en) * | 2019-05-08 | 2020-11-12 | ||
JP7184176B2 (en) | 2019-05-08 | 2022-12-06 | 日本電気株式会社 | Allocation device, method and program |
JP7184175B2 (en) | 2019-05-08 | 2022-12-06 | 日本電気株式会社 | Operation unit and operation allocation method |
CN110399900A (en) * | 2019-06-26 | 2019-11-01 | 腾讯科技(深圳)有限公司 | Method for checking object, device, equipment and medium |
CN110648316A (en) * | 2019-09-07 | 2020-01-03 | 创新奇智(成都)科技有限公司 | Steel coil end face edge detection algorithm based on deep learning |
CN110987189A (en) * | 2019-11-21 | 2020-04-10 | 北京都是科技有限公司 | Method, system and device for detecting temperature of target object |
CN113132755B (en) * | 2019-12-31 | 2022-04-01 | 北京大学 | Method and system for encoding extensible man-machine cooperative image and method for training decoder |
CN113132755A (en) * | 2019-12-31 | 2021-07-16 | 北京大学 | Extensible man-machine cooperative image coding method and coding system |
CN111582353B (en) * | 2020-04-30 | 2022-01-21 | 恒睿(重庆)人工智能技术研究院有限公司 | Image feature detection method, system, device and medium |
CN111582353A (en) * | 2020-04-30 | 2020-08-25 | 恒睿(重庆)人工智能技术研究院有限公司 | Image feature detection method, system, device and medium |
CN113240584A (en) * | 2021-05-11 | 2021-08-10 | 上海大学 | Multitask gesture picture super-resolution method based on picture edge information |
CN113240584B (en) * | 2021-05-11 | 2023-04-28 | 上海大学 | Multitasking gesture picture super-resolution method based on picture edge information |
WO2022267046A1 (en) * | 2021-06-25 | 2022-12-29 | 京东方科技集团股份有限公司 | Un-decimated image processing method and apparatus |
CN113538484B (en) * | 2021-07-01 | 2022-06-10 | 广西科技大学 | Deep-refinement multiple-information nested edge detection method |
CN113538484A (en) * | 2021-07-01 | 2021-10-22 | 广西科技大学 | Deep-refinement multiple-information nested edge detection method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107610140A (en) | Near edge detection method, device based on depth integration corrective networks | |
CN111047516B (en) | Image processing method, image processing device, computer equipment and storage medium | |
CN109816012A (en) | A kind of multiscale target detection method of integrating context information | |
Köhler et al. | Mask-specific inpainting with deep neural networks | |
US20210073982A1 (en) | Medical image processing method and apparatus, electronic device, and storage medium | |
CN109829855A (en) | A kind of super resolution ratio reconstruction method based on fusion multi-level features figure | |
CN110189255A (en) | Method for detecting human face based on hierarchical detection | |
CN104574347B (en) | Satellite in orbit image geometry positioning accuracy evaluation method based on multi- source Remote Sensing Data data | |
CN110827213A (en) | Super-resolution image restoration method based on generation type countermeasure network | |
CN107133955B (en) | A kind of collaboration conspicuousness detection method combined at many levels | |
CN106600538A (en) | Human face super-resolution algorithm based on regional depth convolution neural network | |
CN109583445A (en) | Character image correction processing method, device, equipment and storage medium | |
CN105657402A (en) | Depth map recovery method | |
CN106067161A (en) | A kind of method that image is carried out super-resolution | |
FR2646729A1 (en) | Method and system for manipulating drafts of a symbol image in various dimensions and by various displacements of points in order to improve a numerical character figure on a frame output device | |
CN112116543B (en) | Image restoration method, system and device based on detection type generation framework | |
CN113160062A (en) | Infrared image target detection method, device, equipment and storage medium | |
CN110490232A (en) | Method, apparatus, the equipment, medium of training literal line direction prediction model | |
CN111652864A (en) | Casting defect image generation method for generating countermeasure network based on conditional expression | |
CN112270366B (en) | Micro target detection method based on self-adaptive multi-feature fusion | |
CN109544694A (en) | A kind of augmented reality system actual situation hybrid modeling method based on deep learning | |
CN110415284A (en) | A kind of haplopia color image depth map preparation method and device | |
Li et al. | Line drawing guided progressive inpainting of mural damages | |
CN107169498A (en) | It is a kind of to merge local and global sparse image significance detection method | |
CN115953330B (en) | Texture optimization method, device, equipment and storage medium for virtual scene image |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180119 |
|
WD01 | Invention patent application deemed withdrawn after publication |