CN110211061A - List depth camera depth map real time enhancing method and device neural network based - Google Patents
List depth camera depth map real time enhancing method and device neural network based Download PDFInfo
- Publication number
- CN110211061A CN110211061A CN201910417886.2A CN201910417886A CN110211061A CN 110211061 A CN110211061 A CN 110211061A CN 201910417886 A CN201910417886 A CN 201910417886A CN 110211061 A CN110211061 A CN 110211061A
- Authority
- CN
- China
- Prior art keywords
- depth
- image
- neural network
- feedforward neural
- illumination
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 72
- 238000000034 method Methods 0.000 title claims abstract description 46
- 230000002708 enhancing effect Effects 0.000 title claims abstract description 42
- 238000012549 training Methods 0.000 claims abstract description 32
- 239000011159 matrix material Substances 0.000 claims abstract description 19
- 230000005855 radiation Effects 0.000 claims abstract description 13
- 230000009466 transformation Effects 0.000 claims abstract description 12
- 238000005286 illumination Methods 0.000 claims description 43
- 241000208340 Araliaceae Species 0.000 claims description 12
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 12
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 12
- 235000008434 ginseng Nutrition 0.000 claims description 12
- 230000008569 process Effects 0.000 abstract description 9
- 230000002452 interceptive effect Effects 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 35
- 230000004927 fusion Effects 0.000 description 7
- 238000012360 testing method Methods 0.000 description 6
- 238000013461 design Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 4
- 238000010276 construction Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000000670 limiting effect Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000001902 propagating effect Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 description 1
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000003416 augmentation Effects 0.000 description 1
- 230000002146 bilateral effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000004438 eyesight Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000002829 reductive effect Effects 0.000 description 1
- 238000010008 shearing Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/02—Affine transformations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/38—Registration of image sequences
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/80—Analysis of captured images to determine intrinsic or extrinsic camera parameters, i.e. camera calibration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of single depth camera depth map real time enhancing method and devices neural network based, wherein, this method comprises: passing through the depth image and RGB image of depth camera collecting sample, depth image is aligned with RGB image according to matrix is joined inside and outside the depth camera of calibration, and RGB image is transformed into gray space and carries out radiation transformation, obtain the grayscale image being aligned with the feature of depth image;Feedforward neural network and loss function are constructed, depth image and grayscale image input feedforward neural network are trained, and update the weight of feedforward neural network according to loss function backpropagation;Feedforward neural network after grayscale image to be input to training update weight is to export enhancing depth image.This method carries out shooting, collecting depth image to sample using depth camera, without high-precision scanning device acquisition real depth map as supervision message, eliminates the process manually demarcated, provides good interactive three-dimensional for user and rebuild experience.
Description
Technical field
The present invention relates to computer visions and computer graphics techniques field, in particular to a kind of neural network based
Single depth camera depth map real time enhancing method and device.
Background technique
Consumer level depth camera is gradually popularized in recent years, and being even more built-in in especially newest Iphone X is based on structure
The depth camera of light.This makes from 3-D scanning to virtual reality, and many brand-new mobile terminal applications are possibly realized in mixed reality.
Although the resolution ratio and quality of the initial data of sensor acquisition have promotion, the depth obtained at this stage from consumer level depth camera
Degree figure still has many noises, also lacks enough details.Such as body three-dimensional reconstruction is computer graphics and computer view
The Important Problems in feel field.The human 3d model of high quality has extensively in fields such as video display amusement, demographic data statistical analysis
General application prospect and important application value.But the acquisition of high quality human 3d model usually rely on it is expensive swash
Photoscanner or polyphaser array system are realized, at high cost also significantly there is can not be real-time although precision is higher
The disadvantages of, it can not spread in general public daily life.There are many methods to believe using the high frequency in high resolution R GB image
Breath, while removing the distinctive a large amount of structured noises of depth transducer.Traditional class bilateral filtering method can not be in enhancing details
While guarantee the authenticity of depth map, complicated optimization algorithm is usually required by the method that light and shade restores the 3D shape of object
And to specific input failure, based on the smooth fusion method of time domain can not single frames real time enhancing depth map, and data-driven
Machine learning algorithm can not carry out unsupervisedly in the case where no real depth diagram data.
Summary of the invention
The present invention is directed to solve at least some of the technical problems in related technologies.
For this purpose, an object of the present invention is to provide a kind of single depth camera depth maps neural network based to increase in real time
Strong method, this method can directly shoot with a large amount of Quick Acquisition depth images object using depth camera, without height
The scanning device acquisition real depth map of precision eliminates the process manually demarcated as supervision message.
It is another object of the present invention to propose a kind of single depth camera depth map real time enhancing neural network based
Device.
In order to achieve the above objectives, it is deep to propose a kind of single depth camera neural network based for one aspect of the present invention embodiment
Spend figure real time enhancing method, comprising: by the depth image and RGB image of depth camera collecting sample, according to the depth of calibration
The depth image is aligned by the inside and outside ginseng matrix of camera with the RGB image, and the RGB image is transformed into gray space
And radiation transformation is carried out, obtain the grayscale image being aligned with the feature of the depth image;Construct feedforward neural network and loss letter
The depth image and the grayscale image are inputted the feedforward neural network and are trained by number, and according to the loss function
Backpropagation updates the weight of the feedforward neural network;The grayscale image is input to the forward direction after training updates the weight
Neural network is to export enhancing depth image.
Single depth camera depth map real time enhancing method neural network based of the embodiment of the present invention, by utilizing depth
Camera acquires RGB image and depth image;RGB image and depth image are aligned according to the camera internal reference of calibration;Building fusion is more
The neural network of the multiple dimensioned output of grade and unsupervised loss function;RGB image and depth image training nerve are inputted simultaneously
Network simultaneously updates network weight according to loss function backpropagation;Fixed network weight is for test and deployment phase, according to depth
Spend the RGB image real time enhancing depth map of camera acquisition.Thus, it is possible to directly be shot object with big using depth camera
Quick Acquisition depth image is measured, without high-precision scanning device acquisition real depth map as supervision message, is eliminated simultaneously
The process manually demarcated.Data needed for the data-driven method are very easy to acquisition, can be simply using end-to-end unsupervised
Training completes training in the PC machine of single video card.
In addition, single depth camera depth map real time enhancing method neural network based according to the above embodiment of the present invention
There can also be following additional technical characteristic:
Further, the radiation transformation are as follows:
Zcpc=RZdpd+T
Wherein, (Rc,Tc) and (Rd,Td) it is respectively the colour of calibration and the outer ginseng matrix of depth transducer, KcAnd KdRespectively
For the colour of calibration and the internal reference matrix of depth transducer, ZdFor the depth value of pixel, ZcIndicate corresponding on colour or gray level image
The homogeneous coordinates value of point.
Further, the loss function formula are as follows:
L(Ddt, D, I) and=λshlsh+λfidlfid+λsmolsmo
Wherein, DdtFor the depth map, D is the depth image, and I is the gray level image;
lshItem is lost for illumination, according to the depth map D of enhancingdtThe normal direction figure of calculating is Ndt, lshIs defined as:
Wherein, B () is that irradiation level calculates function, l*It is the illumination tensor of estimation, R is albedo figure, and I is the gray scale
Figure,Indicate the difference of gradient;
lfidItem is lost for value preserving, is defined as:
lfid(Ddt, D)=| Ddt-D|1
lsmoSmoothly to lose item, it is defined as DdtIn anisotropic full variation:
Further, the RGB image assists the depth image to denoise and in the feedforward neural network that enhances, packet
The stack operation of characteristic pattern is exported after convolution containing multiple identical or adjacent scale feature figure.
Further, the illumination tensor is estimated by the depth image, the light is calculated according to the illumination tensor
According to the numerical value of loss item.
In order to achieve the above objectives, another aspect of the present invention embodiment proposes a kind of single depth camera neural network based
Depth map real time enhancing square law device, comprising:
Alignment module, for passing through the depth image and RGB image of depth camera collecting sample, according to the depth phase of calibration
The depth image is aligned by the inside and outside ginseng matrix of machine with the RGB image, and the RGB image is transformed into gray space simultaneously
Radiation transformation is carried out, the grayscale image being aligned with the feature of the depth image is obtained;
Training update module, for constructing feedforward neural network and loss function, by the depth image and the gray scale
Figure inputs the feedforward neural network and is trained, and updates the feedforward neural network according to the loss function backpropagation
Weight;
Enhance module, trains the feedforward neural network after updating the weight to export for the grayscale image to be input to
Enhance depth image.
Single depth camera depth map real time enhancing square law device neural network based of the embodiment of the present invention, passes through utilization
Depth camera acquires RGB image and depth image;RGB image and depth image are aligned according to the camera internal reference of calibration;Building is melted
The neural network of the multistage multiple dimensioned output of conjunction and unsupervised loss function;RGB image and depth image training are inputted simultaneously
Neural network simultaneously updates network weight according to loss function backpropagation;Fixed network weight is for test and deployment phase, root
The RGB image real time enhancing depth map acquired according to depth camera.Thus, it is possible to directly be shot using depth camera to object
It is saved simultaneously with a large amount of Quick Acquisition depth images without high-precision scanning device acquisition real depth map as supervision message
The process of artificial calibration is gone.Data needed for the data-driven method are very easy to acquisition, can simply use end-to-end nothing
Supervised training completes training in the PC machine of single video card.
In addition, single depth camera depth map real time enhancing method neural network based according to the above embodiment of the present invention
Device can also have following additional technical characteristic:
Further, the radiation transformation are as follows:
Zcpc=RZdpd+T
Wherein, (Rc,Tc) and (Rd,Td) it is respectively the colour of calibration and the outer ginseng matrix of depth transducer, KcAnd KdRespectively
For the colour of calibration and the internal reference matrix of depth transducer, ZdFor the depth value of pixel, ZcIndicate corresponding on colour or gray level image
The homogeneous coordinates value of point.
Further, the loss function formula are as follows:
L(Ddt, D, I) and=λshlsh+λfidlfid+λsmolsmo
Wherein, DdtFor the depth map, D is the depth image, and I is the gray level image;
lshItem is lost for illumination, according to the depth map D of enhancingdtThe normal direction figure of calculating is Ndt, lshIs defined as:
Wherein, B () is that irradiation level calculates function, l*It is the illumination tensor of estimation, R is albedo figure, and I is the gray scale
Figure,Indicate the difference of gradient;
lfidItem is lost for value preserving, is defined as:
lfid(Ddt, D)=| Ddt-D|1
lsmoSmoothly to lose item, it is defined as DdtIn anisotropic full variation:
Further, the RGB image assists the depth image to denoise and in the feedforward neural network that enhances, packet
The stack operation of characteristic pattern is exported after convolution containing multiple identical or adjacent scale feature figure.
Further, the illumination tensor is estimated by the depth image, the light is calculated according to the illumination tensor
According to the numerical value of loss item.
The additional aspect of the present invention and advantage will be set forth in part in the description, and will partially become from the following description
Obviously, or practice through the invention is recognized.
Detailed description of the invention
Above-mentioned and/or additional aspect and advantage of the invention will become from the following description of the accompanying drawings of embodiments
Obviously and it is readily appreciated that, in which:
Fig. 1 is single depth camera depth map real time enhancing method neural network based according to one embodiment of the invention
Flow chart;
Fig. 2 is single depth camera depth map real time enhancing side neural network based according to another embodiment of the invention
Method flow chart;
Fig. 3 is the neural network structure figure according to the multistage multiple dimensioned output of fusion of the building of one embodiment of the invention;
Fig. 4 is single depth camera depth map real time enhancing device neural network based according to one embodiment of the invention
Structural schematic diagram.
Specific embodiment
The embodiment of the present invention is described below in detail, examples of the embodiments are shown in the accompanying drawings, wherein from beginning to end
Same or similar label indicates same or similar element or element with the same or similar functions.Below with reference to attached
The embodiment of figure description is exemplary, it is intended to is used to explain the present invention, and is not considered as limiting the invention.
The single depth camera depth map neural network based proposed according to embodiments of the present invention is described with reference to the accompanying drawings
Real time enhancing method and device.
The single depth camera depth neural network based proposed according to embodiments of the present invention is described with reference to the accompanying drawings first
Figure real time enhancing method.
Fig. 1 is single depth camera depth map real time enhancing method neural network based according to one embodiment of the invention
Flow chart.
As shown in Figure 1, single depth camera depth map real time enhancing method neural network based the following steps are included:
In step s101, by the depth image and RGB image of depth camera collecting sample, according to the depth phase of calibration
Depth image is aligned by the inside and outside ginseng matrix of machine with RGB image, and RGB image is transformed into gray space and carries out radiation change
It changes, obtains the grayscale image being aligned with the feature of depth image.
Further, by the depth image and RGB image of depth camera collecting sample, sample can be human body or object
Body can be static or dynamic.Human body or object are shot to obtain depth image and corresponding RGB image, according to
The parameter of calibration is aligned two inlet flows and is pre-processed.
Specifically, as shown in Fig. 2, being shot using single depth camera to static or dynamic object, connected
Individual continuous depth image and corresponding RGB image sequence.According to the colour of calibration and the outer ginseng matrix (R of depth transducerc,
Tc) and (Rd,Td), and colored and depth transducer the internal reference matrix K of calibrationcAnd Kd, gray scale is carried out to the RGB image of input
The pretreatment such as conversion and affine transformation.Radiation transformation are as follows:
Zcpc=RZdpd+T
Wherein, ZdFor the depth value of some pixel, ZcIndicate the homogeneous coordinates value of corresponding points on colored (gray scale) image.It is right
Depth map and RGB image size having the same after neat, in order to which augmentation training data introduces randomness, random shearing fixed edge
A length of 256 square region is as network inputs.
In step s 102, feedforward neural network and loss function are constructed, depth image and grayscale image input is preceding Godwards
It is trained through network, and updates the weight of feedforward neural network according to loss function backpropagation.
Specifically, the feedforward neural network of the multistage multiple dimensioned output of building fusion, input are respectively depth map D and gray scale
Scheme I, exports as without the depth map D to make an uproar and details enhancesdt, the network have simultaneously be fitted depth map denoising and details enhance letter
Several abilities.
A variety of unsupervised loss function L are constructed, illumination loss item l is specifically includedsh(Ddt, I), value preserving loss item l_fid
(Ddt, D) and smooth loss item lsmo(Ddt)。
Wherein, loss function L is calculated, the independent variable of loss function is all from the data of acquisition, is a kind of unsupervised damage
Lose function, formula are as follows:
L(Ddt, D, I) and=λshlsh+λfidlfid+λsmolsmo
L in loss functionshItem is lost for illumination, if according to the depth map D of enhancingdtThe normal direction figure of calculating is Ndt, the Xiang Ding
Justice are as follows:
Wherein B () is that irradiation level calculates function, l*It is the illumination tensor of estimation, R is albedo figure, and Section 2 indicates two
The difference of person's gradient,Indicate the difference of gradient.
L in loss functionfidItem is lost for value preserving, is defined as:
lfid(Ddt, D)=| Ddt-D|1
L in loss functionsmoSmoothly to lose item, it is defined as DdtIn anisotropic full variation:
Construct a kind of neural network of the multistage multiple dimensioned output of fusion of special construction and the loss letter that one kind is unsupervised
It is several, it is therefore intended that the high frequency detail in different scale characteristic pattern in gray level image to be made full use of, with depth in corresponding scale characteristic pattern
Degree figure blends, to achieve the purpose that the noise that removal is mixed on depth map essence low dimensional manifold, while in corresponding position
The details that depth map is lost is recovered from gray level image according to illumination equation.In an example of the present invention, fusion is multistage more
The Neural Network Structure Design of scale output is as shown in Figure 3.
Specifically, D and C of the network inputs for left side, conv expression convolution, concat are indicated along the last one dimension in figure
(characteristic pattern port number) merges two inputs, and pool indicates the operation of maximum value pondization, and resize is indicated using bilinear interpolation
Up-sampling operation.All equal differentiables of operation in network, i.e. error can update convolution nuclear parameter by backpropagation.For ruler
The input that degree is 256, network carry out down-sampling twice and up-sample twice, and after three different scales incorporate convolution
Characteristic pattern, this design have fully merged the feature of gray level image and original depth-map during a propagated forward.Mirror
It can cut down, but can be inferred by the depth value of part, one embodiment of the present of invention after multiple convolution in high frequency detail
In do not have using profound convolutional neural networks.This design can save calculating money in the sizeable situation of receptive field
Shorten net training time, and be effectively prevented from over-fitting in source.
In the training stage, loss function is by three Xiang Zucheng: item l is lost in illuminationsh, value preserving loss item lfidWith smooth loss item
lsmo.The weighted sum that training objective total losses function is three:
L(Ddt, D, I) and=λshlsh+λfidlfid+λsmolsmo
Every depth map and greyscale image data stream for only relying upon input in loss function is a kind of unsupervised instruction
Practice method.
Illustrate that item is lost in illumination first.Surface primary for youth and low frequency illumination, object irradiation level can use the humorous letter of second order ball
Several and illumination tensor is approximate:
Wherein Hb:R3→ R is one group of basic function of spheric harmonic function, and l is the humorous light of second order ball of 9 decomposition scene low frequency illumination
According to coefficient, R is the albedo of object, and N is the normal direction figure calculated according to depth map.Based on above-mentioned formula, the present invention constructs one
Kind loses item without the unsupervised illumination of truthful data:
Wherein NdtIt is according to depth map DdtThe normal direction figure of calculating, illumination lose Section 2 in item and constrain the two gradient
Difference, gradient difference value can more robustly enhance depth map in complex illumination.It is anti-for general object when calculating penalty values
It is assumed to be definite value according to rate R, illumination tensor l can be by least-squares estimation:
In order to limit enhanced depth map close to the noise-containing depth map being originally inputted, the present invention designs value preserving damage
Lose item:
lfid(Ddt, D)=| Ddt-D|1
Wherein norm loss item can drive network to export depth map DdtIt is as sparse as possible with input D error, it ensure that increasing
Depth map after strong has restored the low frequency part in D as precisely as possible, and has filtered out the high frequency knot of depth transducer generation
Structure noise.
Last in loss function smoothly loses item lsmoAs regular terms, the people of illumination item introducing is effectively reduced
Work trace.Smooth loss item is defined as DdtIn anisotropic full variation:
After specifying overall loss function, the training process of network is summarized as follows.In one embodiment of the invention, network
Trained batch processing quantity is 64, and total iteration wheel number is 20.Initial learning rate is 0.001, the optimization plan of target loss function
Adam is slightly selected, learning rate and momentum, and the demand not additional to video memory can be adaptively adjusted in the training process.
After two norms of the gradient value that training reaches the wheel number of setting or backpropagation calculates are less than certain threshold value, no longer update
Network weight saves optimal network model weight, repeats application or the further tuning under more data.
In step s 103, the feedforward neural network after grayscale image to be input to training update weight is to export enhancing depth
Image.
Further, fixed network weight W is walked for test and deployment phase according to the data that depth camera acquires
Rapid B enhances original depth map using the high-frequency information in grayscale image I, the depth map D enhanceddt, network forward calculation speed
Degree can achieve the requirement applied in real time, provides good interactive three-dimensional for user and rebuilds experience, before possessing wide application
Scape.
Specifically, after training has updated feedforward neural network, in test or deployment phase, due to convolution operation and pond
It is partial operation, model can be extended to arbitrary size, and the input depth map and RGB image of the ratio of width to height are further widened
The application scenarios of the method for the present invention.Similarly, the input of network is pretreated depth map and greyscale image data stream.This rank
Section fixed network weight, enhances original depth map D using the high-frequency information in grayscale image I, the depth map D enhanceddt.By
Compact in the network structure of design, one embodiment of the present of invention entirety weight quantity is only 130,000, and offline model occupies empty
Between it is very small, can be deployed in various mobile devices completely.For the 640 × 480 of the acquisition of general consumer level depth camera
The depth map of size, constructed network forward calculation speed fully achieves real-time application in one embodiment of the present of invention
Requirement, it might even be possible to run on the hardware systems such as specific smart phone processing unit or mobile PCs.
It is understood that object noiseless and the depth comprising enhancing details can be obtained in real time in deployment phase
Figure ensure that the accuracy of depth map, neural network have good generalization ability, can be enhanced not while enhancing depth map
The depth map of same human body, object, effect is good, and the speed of service is more than in real time, to gather around and have broad application prospects, the model ginseng after training
Number can be deployed on the hardware systems such as specific smart phone processing unit or mobile PCs.
Further, RGB image auxiliary depth map denoise and in the network that enhances, includes multiple identical or adjacent scale spy
Sign figure exports the stack operation of characteristic pattern after convolution.
Further, item l is lost in illumination in unsupervised loss functionshEvaluation method, in the hypothesis of uniform albedo
Under, illumination tensor l*It can be estimated first by the D inputted, recycle illumination l*Calculate the numerical value l of loss itemsh(l*,Ndt, I),
This process can iteration repeatedly to improve estimation precision.That is, illumination tensor is estimated by depth image, according to illumination tensor
Calculate the numerical value of illumination loss item.
Further, the denoising enhancing of depth map or details generation cannot substantially deviate original depth-map D, value preserving loss
Item has larger impact for the constringency performance of network training.
The single depth camera depth map real time enhancing method neural network based proposed according to embodiments of the present invention, passes through
Special object is acquired to obtain individual depth image and RGB image or depth map and RGB image stream;RGB image is transformed into
Gray space is simultaneously aligned with deepness image registration;RGB image and depth image are aligned according to the camera internal reference of calibration;Construct one kind
The neural network of the multistage multiple dimensioned output of the fusion of special construction, constructs a kind of nothing for not needing depth map true value in the training stage
Supervise loss function;In the training stage, while inputting RGB image and depth image training neural network and anti-according to loss function
Network weight is updated to propagating;In test or actual deployment stage, fixed network weight only carries out the forward calculation of network, utilizes
High-frequency information in the RGB image of depth camera acquisition, real time enhancing depth map.
Thus, it is possible to directly be shot object with a large amount of Quick Acquisition depth images using depth camera, without height
The scanning device acquisition real depth map of precision eliminates the process manually demarcated as supervision message.The data-driven
Data needed for method are very easy to acquisition, can simply use end-to-end unsupervised training, complete instruction in the PC machine of single video card
Practice.
The single depth camera depth map neural network based proposed according to embodiments of the present invention referring next to attached drawing description
Real time enhancing square law device.
Fig. 4 is single depth camera depth map real time enhancing method neural network based according to one embodiment of the invention
Apparatus structure schematic diagram.
As shown in figure 4, single depth camera depth map real time enhancing square law device neural network based includes: alignment mould
Block 100, training update module 200 and enhancing module 300.
Wherein, alignment module 100, for passing through the depth image and RGB image of depth camera collecting sample, according to calibration
Depth camera inside and outside ginseng matrix depth image is aligned with RGB image, and RGB image is transformed into gray space and is carried out
Radiation transformation, obtains the grayscale image being aligned with the feature of depth image.
Training update module 200 inputs depth image and grayscale image for constructing feedforward neural network and loss function
Feedforward neural network is trained, and the weight of feedforward neural network is updated according to loss function backpropagation.
Enhance module 300, trains the feedforward neural network after updating weight to export enhancing for grayscale image to be input to
Depth image.
The device can provide good interactive three-dimensional for user and rebuild experience, gather around and have broad application prospects.
Further, radiation transformation are as follows:
Zcpc=RZdpd+T
Wherein, (Rc,Tc) and (Rd,Td) it is respectively the colour of calibration and the outer ginseng matrix of depth transducer, KcAnd KdRespectively
For the colour of calibration and the internal reference matrix of depth transducer, ZdFor the depth value of pixel, ZcIndicate corresponding on colour or gray level image
The homogeneous coordinates value of point.
Further, loss function formula are as follows:
L(Ddt, D, I) and=λshlsh+λfidlfid+λsmolsmo
Wherein, DdtFor depth map, D is depth image, and I is gray level image;
lshItem is lost for illumination, according to the depth map D of enhancingdtThe normal direction figure of calculating is Ndt, lshIs defined as:
Wherein, B () is that irradiation level calculates function, l*It is the illumination tensor of estimation, R is albedo figure, and I is grayscale image,Indicate the difference of gradient;
lfidItem is lost for value preserving, is defined as:
lfid(Ddt, D)=| Ddt-D|1
lsmoSmoothly to lose item, it is defined as DdtIn anisotropic full variation:
Further, RGB image auxiliary depth image denoise and in the feedforward neural network that enhances, include it is multiple identical or
Adjacent scale feature figure exports the stack operation of characteristic pattern after convolution.
Further, illumination tensor is estimated by depth image, the numerical value of illumination loss item is calculated according to illumination tensor.
It should be noted that aforementioned to single depth camera depth map real time enhancing embodiment of the method neural network based
The device for being also applied for the embodiment is illustrated, details are not described herein again.
The single depth camera depth map real time enhancing device neural network based proposed according to embodiments of the present invention, passes through
Special object is acquired to obtain individual depth image and RGB image or depth map and RGB image stream;RGB image is transformed into
Gray space is simultaneously aligned with deepness image registration;RGB image and depth image are aligned according to the camera internal reference of calibration;Construct one kind
The neural network of the multistage multiple dimensioned output of the fusion of special construction, constructs a kind of nothing for not needing depth map true value in the training stage
Supervise loss function;In the training stage, while inputting RGB image and depth image training neural network and anti-according to loss function
Network weight is updated to propagating;In test or actual deployment stage, fixed network weight only carries out the forward calculation of network, utilizes
High-frequency information in the RGB image of depth camera acquisition, real time enhancing depth map.
Thus, it is possible to directly be shot object with a large amount of Quick Acquisition depth images using depth camera, without height
The scanning device acquisition real depth map of precision eliminates the process manually demarcated as supervision message.The data-driven
Data needed for method are very easy to acquisition, can simply use end-to-end unsupervised training, complete instruction in the PC machine of single video card
Practice.
In addition, term " first ", " second " are used for descriptive purposes only and cannot be understood as indicating or suggesting relative importance
Or implicitly indicate the quantity of indicated technical characteristic.Define " first " as a result, the feature of " second " can be expressed or
Implicitly include at least one this feature.In the description of the present invention, the meaning of " plurality " is at least two, such as two, three
It is a etc., unless otherwise specifically defined.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example
Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not
It must be directed to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be in office
It can be combined in any suitable manner in one or more embodiment or examples.In addition, without conflicting with each other, the skill of this field
Art personnel can tie the feature of different embodiments or examples described in this specification and different embodiments or examples
It closes and combines.
Although the embodiments of the present invention has been shown and described above, it is to be understood that above-described embodiment is example
Property, it is not considered as limiting the invention, those skilled in the art within the scope of the invention can be to above-mentioned
Embodiment is changed, modifies, replacement and variant.
Claims (10)
1. a kind of single depth camera depth map real time enhancing method neural network based, which comprises the following steps:
By the depth image and RGB image of depth camera collecting sample, according to joining matrix inside and outside the depth camera of calibration for institute
It states depth image to be aligned with the RGB image, and the RGB image is transformed into gray space and carries out radiation transformation, obtain
To the grayscale image being aligned with the feature of the depth image;
Feedforward neural network and loss function are constructed, the depth image and the grayscale image are inputted into the feedforward neural network
It is trained, and updates the weight of the feedforward neural network according to the loss function backpropagation;
Feedforward neural network after the grayscale image to be input to the training update weight is to export enhancing depth image.
2. the method according to claim 1, wherein the radiation converts are as follows:
Zcpc=RZdpd+T
Wherein, (Rc, Tc) and (Rd, Td) it is respectively the colour of calibration and the outer ginseng matrix of depth transducer, KcAnd KdRespectively mark
The internal reference matrix of fixed colour and depth transducer, ZdFor the depth value of pixel, ZcIndicate corresponding points on colour or gray level image
Homogeneous coordinates value.
3. the method according to claim 1, wherein the loss function formula are as follows:
L(Ddt, D, I) and=λshlsh+λfidlfid+λsmolsmo
Wherein, DdtFor the depth map, D is the depth image, and I is the gray level image;
lshItem is lost for illumination, according to the depth map D of enhancingdtThe normal direction figure of calculating is Ndt, lshIs defined as:
Wherein, B () is that irradiation level calculates function, l*It is the illumination tensor of estimation, R is albedo figure, and I is the grayscale image,Indicate the difference of gradient;
lfidItem is lost for value preserving, is defined as:
lfid(Ddt, D)=| Ddt-D|1
lsmoSmoothly to lose item, it is defined as DdtIn anisotropic full variation:
4. the method according to claim 1, wherein
The RGB image assists the depth image to denoise and in the feedforward neural network that enhances, comprising multiple identical or
Adjacent scale feature figure exports the stack operation of characteristic pattern after convolution.
5. according to the method described in claim 3, it is characterized in that,
The illumination tensor is estimated by the depth image, and the number of the illumination loss item is calculated according to the illumination tensor
Value.
6. a kind of single depth camera depth map real time enhancing device neural network based characterized by comprising
Alignment module, for passing through the depth image and RGB image of depth camera collecting sample, according in the depth camera of calibration
The depth image is aligned by outer ginseng matrix with the RGB image, and the RGB image is transformed into gray space and is carried out
Radiation transformation, obtains the grayscale image being aligned with the feature of the depth image;
Training update module, it is for constructing feedforward neural network and loss function, the depth image and the grayscale image is defeated
Enter the feedforward neural network to be trained, and updates the power of the feedforward neural network according to the loss function backpropagation
Weight;
Enhance module, trains the feedforward neural network after updating the weight to export enhancing for the grayscale image to be input to
Depth image.
7. device according to claim 6, which is characterized in that the radiation transformation are as follows:
Zcpc=RZdPd+T
Wherein, (Rc, Tc) and (Rd, Td) it is respectively the colour of calibration and the outer ginseng matrix of depth transducer, KcAnd KdRespectively mark
The internal reference matrix of fixed colour and depth transducer, ZdFor the depth value of pixel, ZcIndicate corresponding points on colour or gray level image
Homogeneous coordinates value.
8. device according to claim 6, which is characterized in that the loss function formula are as follows:
L(Ddt, D, I) and=λshlsh+λfidlfid+λsmolsmo
Wherein, DdtFor the depth map, D is the depth image, and I is the gray level image;
lshItem is lost for illumination, according to the depth map D of enhancingdtThe normal direction figure of calculating is Ndt, lshIs defined as:
Wherein, B () is that irradiation level calculates function, l*It is the illumination tensor of estimation, R is albedo figure, and I is the grayscale image,Indicate the difference of gradient;
lfidItem is lost for value preserving, is defined as:
lfid(Ddt, D)=| Ddt-D|1
lsmoSmoothly to lose item, it is defined as DdtIn anisotropic full variation:
9. device according to claim 6, which is characterized in that
The RGB image assists the depth image to denoise and in the feedforward neural network that enhances, comprising multiple identical or
Adjacent scale feature figure exports the stack operation of characteristic pattern after convolution.
10. device according to claim 6, which is characterized in that
The illumination tensor is estimated by the depth image, and the number of the illumination loss item is calculated according to the illumination tensor
Value.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910417886.2A CN110211061A (en) | 2019-05-20 | 2019-05-20 | List depth camera depth map real time enhancing method and device neural network based |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910417886.2A CN110211061A (en) | 2019-05-20 | 2019-05-20 | List depth camera depth map real time enhancing method and device neural network based |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110211061A true CN110211061A (en) | 2019-09-06 |
Family
ID=67787810
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910417886.2A Pending CN110211061A (en) | 2019-05-20 | 2019-05-20 | List depth camera depth map real time enhancing method and device neural network based |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110211061A (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110689558A (en) * | 2019-09-30 | 2020-01-14 | 清华大学 | Multi-sensor image enhancement method and device |
CN111272290A (en) * | 2020-03-13 | 2020-06-12 | 西北工业大学 | Temperature measurement thermal infrared imager calibration method and device based on deep neural network |
CN111275751A (en) * | 2019-10-12 | 2020-06-12 | 浙江省北大信息技术高等研究院 | Unsupervised absolute scale calculation method and system |
CN111652966A (en) * | 2020-05-11 | 2020-09-11 | 北京航空航天大学 | Three-dimensional reconstruction method and device based on multiple visual angles of unmanned aerial vehicle |
CN111784757A (en) * | 2020-06-30 | 2020-10-16 | 北京百度网讯科技有限公司 | Training method of depth estimation model, depth estimation method, device and equipment |
CN112767294A (en) * | 2021-01-14 | 2021-05-07 | Oppo广东移动通信有限公司 | Depth image enhancement method and device, electronic equipment and storage medium |
CN112927154A (en) * | 2021-03-05 | 2021-06-08 | 上海炬佑智能科技有限公司 | ToF device, depth camera and gray scale image enhancement method |
CN113052884A (en) * | 2021-03-17 | 2021-06-29 | Oppo广东移动通信有限公司 | Information processing method, information processing apparatus, storage medium, and electronic device |
CN113096172A (en) * | 2021-03-22 | 2021-07-09 | 西安交通大学 | Reverse generation method from iToF depth data to original raw data |
CN113096228A (en) * | 2021-06-09 | 2021-07-09 | 上海影创信息科技有限公司 | Real-time illumination estimation and rendering method and system based on neural network |
CN113126944A (en) * | 2021-05-17 | 2021-07-16 | 北京的卢深视科技有限公司 | Depth map display method, display device, electronic device, and storage medium |
CN113362241A (en) * | 2021-06-03 | 2021-09-07 | 太原科技大学 | Depth map denoising method combining high-low frequency decomposition and two-stage fusion strategy |
CN113658037A (en) * | 2021-08-24 | 2021-11-16 | 凌云光技术股份有限公司 | Method and device for converting depth image into gray image |
CN114359123A (en) * | 2022-01-12 | 2022-04-15 | 广东汇天航空航天科技有限公司 | Image processing method and device |
CN115375827A (en) * | 2022-07-21 | 2022-11-22 | 荣耀终端有限公司 | Illumination estimation method and electronic equipment |
CN116612357A (en) * | 2023-07-11 | 2023-08-18 | 睿尔曼智能科技(北京)有限公司 | Method, system and storage medium for constructing unsupervised RGBD multi-mode data set |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016704A (en) * | 2017-03-09 | 2017-08-04 | 杭州电子科技大学 | A kind of virtual reality implementation method based on augmented reality |
CN108399610A (en) * | 2018-03-20 | 2018-08-14 | 上海应用技术大学 | A kind of depth image enhancement method of fusion RGB image information |
-
2019
- 2019-05-20 CN CN201910417886.2A patent/CN110211061A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016704A (en) * | 2017-03-09 | 2017-08-04 | 杭州电子科技大学 | A kind of virtual reality implementation method based on augmented reality |
CN108399610A (en) * | 2018-03-20 | 2018-08-14 | 上海应用技术大学 | A kind of depth image enhancement method of fusion RGB image information |
Non-Patent Citations (1)
Title |
---|
SHI YAN 等: "DDRNet:Depth Map Denoising and Refinement for Consumer Depth Cameras Using Cascaded CNNs", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 * |
Cited By (29)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110689558A (en) * | 2019-09-30 | 2020-01-14 | 清华大学 | Multi-sensor image enhancement method and device |
CN110689558B (en) * | 2019-09-30 | 2022-07-22 | 清华大学 | Multi-sensor image enhancement method and device |
CN111275751A (en) * | 2019-10-12 | 2020-06-12 | 浙江省北大信息技术高等研究院 | Unsupervised absolute scale calculation method and system |
CN111275751B (en) * | 2019-10-12 | 2022-10-25 | 浙江省北大信息技术高等研究院 | Unsupervised absolute scale calculation method and system |
CN111272290B (en) * | 2020-03-13 | 2022-07-19 | 西北工业大学 | Temperature measurement thermal infrared imager calibration method and device based on deep neural network |
CN111272290A (en) * | 2020-03-13 | 2020-06-12 | 西北工业大学 | Temperature measurement thermal infrared imager calibration method and device based on deep neural network |
CN111652966A (en) * | 2020-05-11 | 2020-09-11 | 北京航空航天大学 | Three-dimensional reconstruction method and device based on multiple visual angles of unmanned aerial vehicle |
CN111652966B (en) * | 2020-05-11 | 2021-06-04 | 北京航空航天大学 | Three-dimensional reconstruction method and device based on multiple visual angles of unmanned aerial vehicle |
CN111784757A (en) * | 2020-06-30 | 2020-10-16 | 北京百度网讯科技有限公司 | Training method of depth estimation model, depth estimation method, device and equipment |
CN111784757B (en) * | 2020-06-30 | 2024-01-23 | 北京百度网讯科技有限公司 | Training method of depth estimation model, depth estimation method, device and equipment |
CN112767294A (en) * | 2021-01-14 | 2021-05-07 | Oppo广东移动通信有限公司 | Depth image enhancement method and device, electronic equipment and storage medium |
CN112767294B (en) * | 2021-01-14 | 2024-04-26 | Oppo广东移动通信有限公司 | Depth image enhancement method and device, electronic equipment and storage medium |
CN112927154B (en) * | 2021-03-05 | 2023-06-02 | 上海炬佑智能科技有限公司 | ToF device, depth camera and gray image enhancement method |
CN112927154A (en) * | 2021-03-05 | 2021-06-08 | 上海炬佑智能科技有限公司 | ToF device, depth camera and gray scale image enhancement method |
CN113052884A (en) * | 2021-03-17 | 2021-06-29 | Oppo广东移动通信有限公司 | Information processing method, information processing apparatus, storage medium, and electronic device |
CN113096172A (en) * | 2021-03-22 | 2021-07-09 | 西安交通大学 | Reverse generation method from iToF depth data to original raw data |
CN113096172B (en) * | 2021-03-22 | 2023-10-27 | 西安交通大学 | Reverse generation method from iToF depth data to original raw data |
CN113126944A (en) * | 2021-05-17 | 2021-07-16 | 北京的卢深视科技有限公司 | Depth map display method, display device, electronic device, and storage medium |
CN113362241B (en) * | 2021-06-03 | 2022-04-05 | 太原科技大学 | Depth map denoising method combining high-low frequency decomposition and two-stage fusion strategy |
CN113362241A (en) * | 2021-06-03 | 2021-09-07 | 太原科技大学 | Depth map denoising method combining high-low frequency decomposition and two-stage fusion strategy |
CN113096228B (en) * | 2021-06-09 | 2021-08-31 | 上海影创信息科技有限公司 | Real-time illumination estimation and rendering method and system based on neural network |
CN113096228A (en) * | 2021-06-09 | 2021-07-09 | 上海影创信息科技有限公司 | Real-time illumination estimation and rendering method and system based on neural network |
CN113658037A (en) * | 2021-08-24 | 2021-11-16 | 凌云光技术股份有限公司 | Method and device for converting depth image into gray image |
CN113658037B (en) * | 2021-08-24 | 2024-05-14 | 凌云光技术股份有限公司 | Method and device for converting depth image into gray level image |
CN114359123A (en) * | 2022-01-12 | 2022-04-15 | 广东汇天航空航天科技有限公司 | Image processing method and device |
CN115375827A (en) * | 2022-07-21 | 2022-11-22 | 荣耀终端有限公司 | Illumination estimation method and electronic equipment |
CN115375827B (en) * | 2022-07-21 | 2023-09-15 | 荣耀终端有限公司 | Illumination estimation method and electronic equipment |
CN116612357A (en) * | 2023-07-11 | 2023-08-18 | 睿尔曼智能科技(北京)有限公司 | Method, system and storage medium for constructing unsupervised RGBD multi-mode data set |
CN116612357B (en) * | 2023-07-11 | 2023-11-24 | 睿尔曼智能科技(北京)有限公司 | Method, system and storage medium for constructing unsupervised RGBD multi-mode data set |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110211061A (en) | List depth camera depth map real time enhancing method and device neural network based | |
CN108921926B (en) | End-to-end three-dimensional face reconstruction method based on single image | |
Bilinski et al. | Dense decoder shortcut connections for single-pass semantic segmentation | |
CN105787439B (en) | A kind of depth image human synovial localization method based on convolutional neural networks | |
CN109410261B (en) | Monocular image depth estimation method based on pyramid pooling module | |
CN109101975A (en) | Image, semantic dividing method based on full convolutional neural networks | |
CN107358626A (en) | A kind of method that confrontation network calculations parallax is generated using condition | |
CN111542861A (en) | System and method for rendering an avatar using a depth appearance model | |
CN108022213A (en) | Video super-resolution algorithm for reconstructing based on generation confrontation network | |
CN109584290A (en) | A kind of three-dimensional image matching method based on convolutional neural networks | |
CN109087243A (en) | A kind of video super-resolution generation method generating confrontation network based on depth convolution | |
CN113822982A (en) | Human body three-dimensional model construction method and device, electronic equipment and storage medium | |
CN109035142A (en) | A kind of satellite image ultra-resolution method fighting network integration Aerial Images priori | |
CN110473284A (en) | A kind of moving object method for reconstructing three-dimensional model based on deep learning | |
CN107609638A (en) | A kind of method based on line decoder and interpolation sampling optimization convolutional neural networks | |
CN109461177B (en) | Monocular image depth prediction method based on neural network | |
CN110210524A (en) | A kind of training method, image enchancing method and the device of image enhancement model | |
CN107944551A (en) | One kind is used for electrowetting display screen defect identification method | |
CN110168572A (en) | Information processing method, information processing unit, computer readable storage medium | |
CN113658040A (en) | Face super-resolution method based on prior information and attention fusion mechanism | |
CN112634456B (en) | Real-time high-realism drawing method of complex three-dimensional model based on deep learning | |
CN110599585A (en) | Single-image human body three-dimensional reconstruction method and device based on deep learning | |
CN109447897A (en) | A kind of real scene image composition method and system | |
CN114897728A (en) | Image enhancement method and device, terminal equipment and storage medium | |
CN116670720A (en) | Method and system for generating a three-dimensional (3D) model of an object |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190906 |