CN109615588A - A method of image reconstruction is solved the problems, such as based on depth autoregression model - Google Patents
A method of image reconstruction is solved the problems, such as based on depth autoregression model Download PDFInfo
- Publication number
- CN109615588A CN109615588A CN201811281916.3A CN201811281916A CN109615588A CN 109615588 A CN109615588 A CN 109615588A CN 201811281916 A CN201811281916 A CN 201811281916A CN 109615588 A CN109615588 A CN 109615588A
- Authority
- CN
- China
- Prior art keywords
- image
- reconstruction
- depth
- pixel
- model
- 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.)
- Withdrawn
Links
- 238000000034 method Methods 0.000 title claims abstract description 86
- 238000012549 training Methods 0.000 claims abstract description 27
- 230000008859 change Effects 0.000 claims abstract description 15
- 230000009466 transformation Effects 0.000 claims abstract description 15
- 238000007476 Maximum Likelihood Methods 0.000 claims abstract description 13
- 238000005192 partition Methods 0.000 claims abstract description 10
- 150000001875 compounds Chemical class 0.000 claims abstract description 8
- 230000008569 process Effects 0.000 claims abstract description 8
- 238000005457 optimization Methods 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 35
- 230000006835 compression Effects 0.000 claims description 19
- 238000007906 compression Methods 0.000 claims description 19
- 230000003416 augmentation Effects 0.000 claims description 7
- 230000003190 augmentative effect Effects 0.000 claims description 6
- 238000004088 simulation Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 5
- 238000009792 diffusion process Methods 0.000 claims description 3
- 238000003384 imaging method Methods 0.000 claims description 3
- 239000003086 colorant Substances 0.000 claims description 2
- 238000013528 artificial neural network Methods 0.000 claims 1
- 238000006243 chemical reaction Methods 0.000 claims 1
- 238000000605 extraction Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000015572 biosynthetic process Effects 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 238000003786 synthesis reaction Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 241000196324 Embryophyta Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 206010052428 Wound Diseases 0.000 description 1
- 208000027418 Wounds and injury Diseases 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000009659 non-destructive testing Methods 0.000 description 1
- 230000003252 repetitive effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- 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
- G06T3/20—Linear translation of whole images or parts thereof, e.g. panning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
Abstract
A kind of method for solving the problems, such as image reconstruction based on depth autoregression model proposed in the present invention, its main contents includes: depth autoregression model, the realization details that linear transformation, optimization depth autoregression model and reconstruction image are carried out to input picture, its process is, Maximum-likelihood estimation training first is carried out to depth autoregression model with the image of database, it is made to have the ability for solving reconstruction image task；Then, specific projection operator is introduced in reconstruction image expression formula, and by using gradual change updating method and partition and compound training method, to improve the performance of depth autoregression model；Linear transformation is carried out to input picture finally, realizing by forward model, to realize image reconstruction.Present invention employs depth autoregression models, its performance is optimized, and ensure that the pixel consistency of reconstruction image and input picture, and improve the gradually changeable of reconstruction image.
Description
Technical field
The present invention relates to field of image processings, more particularly, to a kind of solution image weight based on depth autoregression model
The method for building problem.
Background technique
Image Reconstruction Technology is a kind of non-destructive testing technology, mainly by being computed to DATA REASONING is carried out outside object
Machine digitized processing is trained and is simulated by depth convolutional neural networks model, to realize the reconstruction to subject image.?
Monitoring field rebuilds the photo that monitoring device is filmed using Image Reconstruction Technology, the pixel of photo is improved, to obtain
Take the face feature information of suspect；In satellite remote sensing field, geographic pattern is rebuild using Image Reconstruction Technology, to obtain
Obtain more accurate geography information；In the field of medical imaging, X-ray machine, nuclear magnetic resonane scanne etc. are set using Image Reconstruction Technology
The image of standby scanning is rebuild, and be can get the state of an illness image being more clear, is provided convenience for diagnosis；In addition, image
Reconstruction technique also has a wide range of applications in fields such as aerospace, virtual realities.However, existing image rebuilding method exists
Pixel it is not high, the problems such as image morphing is poor.
A kind of method for solving the problems, such as image reconstruction based on depth autoregression model proposed in the present invention, first uses data
The image in library carries out Maximum-likelihood estimation training to depth autoregression model, it is made to have the ability for solving reconstruction image task；
Then, specific projection operator is introduced in reconstruction image expression formula, and by using gradual change updating method and partition and synthesis
Coaching method, to improve the performance of depth autoregression model；Finally, linearly being turned by forward model realization to input picture
It changes, to realize image reconstruction.Present invention employs depth autoregression models, its performance is optimized, and ensure that reconstruction
The pixel consistency of image and input picture, and improve the gradually changeable of reconstruction image.
Summary of the invention
The problems such as that there is pixels is not high for existing image rebuilding method, and image morphing is poor, the purpose of the present invention
It is to provide a kind of method for solving the problems, such as image reconstruction based on depth autoregression model, process is, first with database
Image carries out Maximum-likelihood estimation training to depth autoregression model, it is made to have the ability for solving reconstruction image task；Then,
Specific projection operator is introduced in reconstruction image expression formula, and by using gradual change updating method and partition and compound training
Method, to improve the performance of depth autoregression model；Linear transformation is carried out to input picture finally, realizing by forward model, from
And realize image reconstruction.
To solve the above problems, the present invention provides a kind of side for solving the problems, such as image reconstruction based on depth autoregression model
Method, main contents include:
(1) depth autoregression model；
(2) linear transformation is carried out to input picture；
(3) optimize depth autoregression model；
(4) the realization details of reconstruction image.
Wherein, the depth autoregression model, the depth convolutional neural networks with residual error connection, using the general of orientation
Rate opinion chain rule carrys out the dependence between simulation pixel；Enabling X is input picture, is indicated with n × n matrix, to every a line
Pixel carries out dot matrixed processingWherein vec () indicates vector, xiIndicate pixel, wherein
I ∈ [1,2 ..., n2],For delimiter；Pixel xiDepending on all pixels being designated as under in X before i, it is expressed as x< i, thus
The density of simultaneous distribution of image pixel indicates are as follows:
Wherein, ρ () indicates the distribution density of pixel；Depth autoregression model is simulated using Joint Distribution, is formed
Logic distribution；Then Maximum-likelihood estimation training is carried out to model on RGB image, so that model has solution reconstruction image
The ability of task；Reconstruction image task is carried out to input picture with the measured value Y reconstruction image X obtained in forward model
Linear transformation, this is a possibility predication problem, is usedIndicate reconstruction image:
Wherein, ρ () indicates the distribution density of pixel, and x represents pixel；Reconstruction imageAs Maximum-likelihood estimation
As a result；Log () is logarithmic function, and ρ (Y | X) indicates the density of Y under the premise of known X.
Wherein, described that linear transformation is carried out to input picture, forward model is mainly used, forward model is compression image
Input picture is multiplexed on independent image detector by generation system using programmable digital micromirror array, is surveyed with obtaining
Magnitude Y is realized and is carried out linear transformation to input picture；To digital micromirror array carry out it is different pre-set, it is different to obtain
The bandwidth of measured value, measured value is determined by the speed of service of digital micromirror array.
Further, the compression image generation system mainly includes SPC and LiSens；SPC has efficient full frame
Sensor and compression sensing device, the high-resolution imaging for non-visible wave band；Using SPC as forward model, expression formula
For y=Φ x, wherein Φ indicates that m × n compresses sensing matrix, y indicate by m independent pixel measurement set at vector, x expression
Pixel；LiSens is the compression image generation system based on line sensor, and each pixel in line sensor is mapped to number
The row of micro mirror array, therefore be only multiplexed on the row to independent image detector of input picture, to obtain measured value Y；It is right
LiSens is optimized, and obtains compression image generation system FlatCam, respectively will be from scene not using amplitude and diffusion mask
With on the pumped FIR laser to sensor of part, the information for being thus located at certain point in image is propagated in entire sensor, so that rebuilding
Image has higher accuracy, expression formula are as follows: Y=ΦLXΦR, wherein X is input picture, is indicated with n × n matrix, and Y is to survey
Magnitude, ΦLAnd ΦRIt is the column matrix and row matrix of input picture X respectively.
Wherein, the optimization depth autoregression model realizes process are as follows: first according to the noise content and direct die of measured value
This method is divided by the form of type: hard constraint and soft-constraint method, augmentation Lagrangian method；Then the estimation to three kinds of modes respectively
Specific projection operator is introduced in image expression formula, projection operator is auxiliary operation matrix, to reduce the appearance of low score image,
And guarantee to carry out image reconstruction in the case where meeting the constraint condition of forward model；In addition, by using gradual change updating method and
Partition and compound training method, further increase the performance of depth autoregression model.
Further, the hard constraint and soft-constraint method, hard constraint method refer in the case where measured value noise is low into
The method of row image reconstruction, reconstruction image are expressed asIts constraint condition is Y=f (X), WhereinFor reconstruction image, Y is measured value, and ρ indicates pixel distribution density, and θ is picture signal ginseng
Number, X is input picture, is indicated with n × n matrix, XijFor the i-th row jth list in matrix X；Soft-constraint method refers to makes an uproar in measured value
The method that image reconstruction is carried out in the case where sound pitch, strong noise measured value are usedIt indicates,η is measured value noise, clothes
From Gaussian Profile, reconstruction image is expressed asλ is measured value noise
Standard deviation.
Further, the augmentation Lagrangian method, for specifically generating compression image generation system FlatCam,
Projection operator does not have closed-form solution, carries out image reconstruction using Augmented Lagrangian Functions:
Wherein, Y is measured value, and ρ indicates pixel distribution density, and θ is picture signal parameter, and X is input picture, with n × n
Matrix indicates that λ is the standard deviation of measured value noise, and L () is Augmented Lagrangian Functions, ΦLAnd ΦRIt is difference input picture X
Column matrix and row matrix,For ΦRTransposition；F is expressed as compressing image generation system FlatCam.
Further, the gradual change updating method, depth autoregression model can tend to when simulating adjacent pixel
Identical value, which is assigned, to two o'clock therefore, quantitative pixel is extracted during each simulation so that two points do not have distinction
Point is updated, remaining holding ortho states, and the pixel thus updated and the pixel not updated produce difference, has image
There is gradually changeable；Renewal step by step method mainly passes through stochastic gradient iterative formula.
Further, partition and compound training method, in 64 × 64 image block training depth autoregression model, therefore to mould
Before type is optimized with Renewal step by step method, image is first divided into 64 × 64 image block, is later again synthesized image block former
The size of beginning is to carry out Maximum-likelihood estimation training.
Wherein, the realization details of the reconstruction image, firstly, image of the model in the ImageNet of down-sampled data library
Carry out the training of 6 batches, wherein the size of the image in database ImageNet is 64 × 64, the sample number of every batch of
Amount control is at 36, filter channel 100；After the completion of the training of depth autoregression model, sampled just from uniformly random distribution
Beginning image accelerates the speed of service using momentum in gradual change update；When rebuilding color image, to three color wounds of RGB
It builds three channels to be trained, the corresponding compression sensing matrix in each channel finally obtains three measured value Y.
Detailed description of the invention
Fig. 1 is a kind of system framework of the method for solving the problems, such as image reconstruction based on depth autoregression model of the present invention
Figure.
Fig. 2 is a kind of flow chart of the method for solving the problems, such as image reconstruction based on depth autoregression model of the present invention.
Fig. 3 is a kind of reconstruction image effect of method for solving the problems, such as image reconstruction based on depth autoregression model of the present invention
Fruit figure.
Fig. 4 is a kind of reconstruction image pair of the method for solving the problems, such as image reconstruction based on depth autoregression model of the present invention
Than figure.
Specific embodiment
It should be noted that in the absence of conflict, the features in the embodiments and the embodiments of the present application can phase
It mutually combines, invention is further described in detail in the following with reference to the drawings and specific embodiments.
Fig. 1 is a kind of system framework of the method for solving the problems, such as image reconstruction based on depth autoregression model of the present invention
Figure.Mainly include depth autoregression model, carry out linear transformation, optimization depth autoregression model and reconstruction image to input picture
Realization details.
Depth autoregression model, the depth convolutional neural networks with residual error connection, using the probability theory chain method of orientation
Then carry out the dependence between simulation pixel；Enabling X is input picture, is indicated with n × n matrix, is carried out to the pixel of every a line
Dot matrixed processingWherein vec () indicates vector, xiIndicate pixel, wherein i ∈ [1,
2 ..., n2],For delimiter；Pixel xiDepending on all pixels being designated as under in X before i, it is expressed as x< i, thus image slices
The density of simultaneous distribution of element indicates are as follows:
Wherein, ρ () indicates the distribution density of pixel；Depth autoregression model is simulated using Joint Distribution, is formed
Logic distribution；Then Maximum-likelihood estimation training is carried out to model on RGB image, so that model has solution reconstruction image
The ability of task；Reconstruction image task is carried out to input picture with the measured value Y reconstruction image X obtained in forward model
Linear transformation, this is a possibility predication problem, is usedIndicate reconstruction image:
Wherein, ρ () indicates the distribution density of pixel, and x represents pixel；Reconstruction imageAs Maximum-likelihood estimation
As a result；Log () is logarithmic function, and ρ (Y | X) indicates the density of Y under the premise of known X.
Linear transformation is carried out to input picture, mainly uses forward model, forward model is compression image generation system, is made
Input picture is multiplexed on independent image detector with programmable digital micromirror array, to obtain measured value Y, realization pair
Input picture carries out linear transformation；To digital micromirror array carry out it is different pre-set, obtain different measured values, measured value
Bandwidth is determined by the speed of service of digital micromirror array.
Wherein, image generation system is compressed, mainly includes SPC and LiSens；SPC has efficient full frame sensor and pressure
Contracting sensing device, the high-resolution imaging for non-visible wave band；Using SPC as forward model, expression formula is y=Φ x,
Middle Φ indicates that m × n compresses sensing matrix, y indicate by m independent pixel measurement set at vector, x expression pixel；LiSens
It is the compression image generation system based on line sensor, each pixel in line sensor is mapped to digital micromirror array
Row, therefore be only multiplexed on the row to independent image detector of input picture, to obtain measured value Y；LiSens is optimized
Compression image generation system FlatCam is obtained, uses amplitude and diffusion mask by the pumped FIR laser from scene different piece respectively
Onto sensor, the information for being thus located at certain point in image is propagated in entire sensor, so that reconstruction image has higher standard
True property, expression formula are as follows: Y=ΦLXΦR, wherein X is input picture, indicates that Y is measured value, Φ with n × n matrixLAnd ΦR
It is the column matrix and row matrix of input picture X respectively.
Optimize depth autoregression model, realize process are as follows: first will according to the form of the noise content of measured value and forward model
This method is divided into hard constraint and soft-constraint method, augmentation Lagrangian method；Then respectively to the estimation image expression formula of three kinds of modes
Middle to introduce specific projection operator, projection operator is auxiliary operation matrix, to reduce the appearance of low score image, and is guaranteed full
Image reconstruction is carried out in the case where the constraint condition of sufficient forward model；In addition, by using gradual change updating method and partition and synthesis
Coaching method further increases the performance of depth autoregression model.
Wherein, hard constraint and soft-constraint method, hard constraint method, which refers to, carries out image reconstruction in the case where measured value noise is low
Method, reconstruction image is expressed asIts constraint condition is Y=f (X),WhereinFor reconstruction image, Y is measured value, and ρ indicates pixel distribution density, and θ is picture signal
Parameter, X are input picture, are indicated with n × n matrix, XijFor the i-th row jth list in matrix X；Soft-constraint method refers in measured value
The method that image reconstruction is carried out in the case that noise is high, strong noise measured value are usedIt indicates,η is measured value noise,
Gaussian distributed, reconstruction image are expressed asλ makes an uproar for measured value
The standard deviation of sound.
Wherein, augmentation Lagrangian method, for specifically generating compression image generation system FlatCam, projection operator does not have
There is closed-form solution, carry out image reconstruction using Augmented Lagrangian Functions:
Wherein, Y is measured value, and ρ indicates pixel distribution density, and θ is picture signal parameter, and X is input picture, with n × n
Matrix indicates that λ is the standard deviation of measured value noise, and L () is Augmented Lagrangian Functions, ΦLAnd ΦRIt is difference input picture X
Column matrix and row matrix,For ΦRTransposition；F is expressed as compressing image generation system FlatCam.
Wherein, gradual change updating method, depth autoregression model can tend to assign two o'clock when simulating adjacent pixel
Therefore identical value, quantitative pixel is extracted during each simulation and is carried out more so that two points do not have distinction
Newly, remaining holding ortho states, the pixel thus updated and the pixel not updated produce difference, and image is made to have gradual change
Property；Renewal step by step method mainly passes through stochastic gradient iterative formula.
Wherein, partition and compound training method in 64 × 64 image block training depth autoregression model, therefore are used model
Before Renewal step by step method optimizes, image is first divided into 64 × 64 image block, is later again synthesized image block original
Size is to carry out Maximum-likelihood estimation training.
The realization details of reconstruction image, firstly, image of the model in the ImageNet of down-sampled data library carries out 6 batches
Training, wherein the size of the image in database ImageNet is 64 × 64, and the sample size of every batch of is controlled 36
It is a, filter channel 100；After the completion of the training of depth autoregression model, initial pictures are sampled from uniformly random distribution, gradually
Accelerate the speed of service using momentum during change is new；When rebuilding color image, three channels are created to three colors of RGB
It is trained, the corresponding compression sensing matrix in each channel finally obtains three measured value Y.
Fig. 2 is a kind of flow chart of the method for solving the problems, such as image reconstruction based on depth autoregression model of the present invention.Its
Process is first to carry out Maximum-likelihood estimation training to depth autoregression model with the image of database, so that it is had solution and rebuilds
The ability of image task；Then, specific projection operator is introduced in reconstruction image expression formula, and updated by using gradual change
Method and partition and compound training method, to improve the performance of depth autoregression model；Input is schemed finally, being realized by forward model
As carrying out linear transformation, to realize image reconstruction.
Fig. 3 is a kind of reconstruction image effect of method for solving the problems, such as image reconstruction based on depth autoregression model of the present invention
Fruit figure.This figure shows this method to the reconstruction effect, including animal, plant, building etc. of various types image, to different form
Image, depth autoregression model according to the difference of the noise content of measured value and the form of forward model, using hard constraint with it is soft
Leash law, augmentation Lagrangian method rebuild image, so that image reconstruction process is more targeted, and are rebuilding
Repetitive exercise several times is carried out to depth autoregression model in journey so that reconstruction image information is more accurate, as the result is shown for
The reconstruction effect of various types image all has good effect.
Fig. 4 is a kind of reconstruction image pair of the method for solving the problems, such as image reconstruction based on depth autoregression model of the present invention
Than figure.Be compared with the traditional method, depth autoregression model is rebuild image using gradual change updating method so that image pixel it
Between have distinction, to generate clean mark, with gradually changeable image；Pixel Loss Rate is to measure the pixel not being updated
Ratio, the effect for generating image promotes with the promotion of pixel Loss Rate.
For those skilled in the art, the present invention is not limited to the details of above-described embodiment, without departing substantially from essence of the invention
In the case where mind and range, the present invention can be realized in other specific forms.In addition, those skilled in the art can be to this hair
Bright to carry out various modification and variations without departing from the spirit and scope of the present invention, these improvements and modifications also should be regarded as of the invention
Protection scope.Therefore, it includes preferred embodiment and all changes for falling into the scope of the invention that the following claims are intended to be interpreted as
More and modify.
Claims (10)
1. a kind of method for solving the problems, such as image reconstruction based on depth autoregression model, which is characterized in that mainly include depth
Autoregression model (one)；Linear transformation (two) are carried out to input picture；Optimize depth autoregression model (three)；The reality of reconstruction image
Existing details (four).
2. based on depth autoregression model (one) described in claims 1, which is characterized in that the depth volume with residual error connection
Product neural network, using the probability theory chain rule of orientation come the dependence between simulation pixel；Enabling X is input picture, uses n
× n matrix indicates, carries out dot matrixed processing to the pixel of every a lineWherein vec () table
Show vector, xiIndicate pixel, wherein i ∈ [1,2 ..., n2],For delimiter；Pixel xiDepending on the institute being designated as under in X before i
There is pixel, is expressed as x< i, thus the density of simultaneous distribution of image pixel indicates are as follows:
Wherein, ρ () indicates the distribution density of pixel；Depth autoregression model is simulated using Joint Distribution, forms logic
Distribution；Then Maximum-likelihood estimation training is carried out to model on RGB image, so that model has solution reconstruction image task
Ability；Reconstruction image task is carried out to input picture linear with the measured value Y reconstruction image X obtained in forward model
Conversion, this is a possibility predication problem, is usedIndicate reconstruction image:
Wherein, ρ () indicates the distribution density of pixel, and x represents pixel；Reconstruction imageThe as result of Maximum-likelihood estimation；
Log () is logarithmic function, and ρ (Y | X) indicates the density of Y under the premise of known X.
3. based on linear transformation (two) are carried out to input picture described in claims 1, which is characterized in that main using positive
Model, forward model are compression image generation systems, and input picture is multiplexed into independence using programmable digital micromirror array
Image detector on, to obtain measured value Y, realize and linear transformation carried out to input picture；Digital micromirror array is carried out not
Same pre-sets, and obtains different measured values, the bandwidth of measured value is determined by the speed of service of digital micromirror array.
4. based on compression image generation system described in claims 3, which is characterized in that mainly include SPC and LiSens；
SPC has efficient full frame sensor and compression sensing device, the high-resolution imaging for non-visible wave band；Using SPC as
Forward model, expression formula are y=Φ x, and wherein Φ indicates that m × n compresses sensing matrix, and y is indicated by m independent pixel measured value
The vector of composition, x indicate pixel；LiSens is the compression image generation system based on line sensor, each of line sensor
Pixel is mapped to the row of digital micromirror array, therefore is only multiplexed on the row to independent image detector of input picture, to obtain
Take measured value Y；LiSens is optimized to obtain compression image generation system FlatCam, respectively using amplitude and diffusion mask
By on the pumped FIR laser to sensor from scene different piece, the information for being thus located at certain point in image passes in entire sensor
It broadcasts, so that reconstruction image has higher accuracy, expression formula are as follows: Y=ΦLXΦR, wherein X is input picture, with n × n square
Matrix representation, Y are measured value, ΦLAnd ΦRIt is the column matrix and row matrix of input picture X respectively.
5. based on optimization depth autoregression model (three) described in claims 1, which is characterized in that realize process are as follows: first root
This method is divided into according to the noise content of measured value and the form of forward model: hard constraint and soft-constraint method, augmentation Lagrangian method；
Then respectively to specific projection operator is introduced in the estimation image expression formula of three kinds of modes, projection operator is auxiliary operation square
Battle array to reduce the appearance of low score image, and carries out image reconstruction in the case where guaranteeing to meet the constraint condition of forward model；
In addition, further increasing the performance of depth autoregression model by using gradual change updating method and partition and compound training method.
6. based on hard constraint described in claims 5 and soft-constraint method, which is characterized in that hard constraint method refers to makes an uproar in measured value
The method that image reconstruction is carried out in the case that sound is low, reconstruction image are expressed asIt constrains item
Part is Y=f (X),WhereinFor reconstruction image, Y is measured value, and ρ indicates pixel distribution density, θ
For picture signal parameter, X is input picture, is indicated with n × n matrix, XijFor the i-th row jth list in matrix X；Soft-constraint method is
Refer to the method that image reconstruction is carried out in the case where measured value noise is high, strong noise measured value is usedIt indicates,η is to survey
Magnitude noise, Gaussian distributed, reconstruction image are expressed asλ is
The standard deviation of measured value noise.
7. based on augmentation Lagrangian method described in claims 5, which is characterized in that raw for specifically generating compression image
At system FlatCam, projection operator does not have closed-form solution, carries out image reconstruction using Augmented Lagrangian Functions:
Wherein, Y is measured value, and ρ indicates pixel distribution density, and θ is picture signal parameter, and X is input picture, with n × n matrix
It indicates, λ is the standard deviation of measured value noise, and L () is Augmented Lagrangian Functions, ΦLAnd ΦRIt is the column of input picture X respectively
Matrix and row matrix,For ΦRTransposition；F is expressed as compressing image generation system FlatCam.
8. based on gradual change updating method described in claims 5, which is characterized in that depth autoregression model is simulating adjacent picture
When vegetarian refreshments, it can tend to assign identical value to two o'clock, so that two points do not have distinction, therefore, in the process of each simulation
The quantitative pixel of middle extraction is updated, and remaining holding ortho states, the pixel thus updated is produced with the pixel not updated
Difference has been given birth to, has made image that there is gradually changeable；Renewal step by step method mainly passes through stochastic gradient iterative formula.
9. based on partition described in claims 5 and compound training method, which is characterized in that the image block training 64 × 64 is deep
Autoregression model is spent, therefore before being optimized to model with Renewal step by step method, image is first divided into 64 × 64 image block,
Image block is synthesized into original size to carry out Maximum-likelihood estimation training again later.
10. the realization details (four) based on reconstruction image described in claims 1, which is characterized in that firstly, model is adopted under
Image in sample database ImageNet carries out the training of 6 batches, wherein the size of the image in database ImageNet is equal
It is 64 × 64, the control of the sample size of every batch of is at 36, filter channel 100；After the completion of the training of depth autoregression model,
Initial pictures are sampled from uniformly random distribution, accelerate the speed of service using momentum in gradual change update；Rebuilding cromogram
When picture, three channels are created to three colors of RGB and are trained, the corresponding compression sensing matrix in each channel, finally
To three measured value Y.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811281916.3A CN109615588A (en) | 2018-10-31 | 2018-10-31 | A method of image reconstruction is solved the problems, such as based on depth autoregression model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811281916.3A CN109615588A (en) | 2018-10-31 | 2018-10-31 | A method of image reconstruction is solved the problems, such as based on depth autoregression model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109615588A true CN109615588A (en) | 2019-04-12 |
Family
ID=66003016
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811281916.3A Withdrawn CN109615588A (en) | 2018-10-31 | 2018-10-31 | A method of image reconstruction is solved the problems, such as based on depth autoregression model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109615588A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113034634A (en) * | 2021-03-04 | 2021-06-25 | 北京大学 | Adaptive imaging method, system and computer medium based on pulse signal |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102722863A (en) * | 2012-04-16 | 2012-10-10 | 天津大学 | Super-resolution reconstruction method for depth map by adopting autoregressive model |
| US20170272639A1 (en) * | 2016-03-16 | 2017-09-21 | Ramot At Tel-Aviv University Ltd. | Reconstruction of high-quality images from a binary sensor array |
-
2018
- 2018-10-31 CN CN201811281916.3A patent/CN109615588A/en not_active Withdrawn
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102722863A (en) * | 2012-04-16 | 2012-10-10 | 天津大学 | Super-resolution reconstruction method for depth map by adopting autoregressive model |
| US20170272639A1 (en) * | 2016-03-16 | 2017-09-21 | Ramot At Tel-Aviv University Ltd. | Reconstruction of high-quality images from a binary sensor array |
Non-Patent Citations (1)
| Title |
|---|
| AKSHAT DAVE ET AL.: "Solving Inverse Computational Imaging Problems using Deep Pixel-level Prior", 《ARXIV:1802.0985V1》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113034634A (en) * | 2021-03-04 | 2021-06-25 | 北京大学 | Adaptive imaging method, system and computer medium based on pulse signal |
| CN113034634B (en) * | 2021-03-04 | 2022-05-06 | 脉冲视觉(北京)科技有限公司 | Adaptive imaging method, system and computer medium based on pulse signal |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2023138471A1 (en) | Three-dimensional scene rendering method, device, and storage medium | |
| CN112465955B (en) | Dynamic human body three-dimensional reconstruction and visual angle synthesis method | |
| CN110599395B (en) | Target image generation method, device, server and storage medium | |
| US11508107B2 (en) | Additional developments to the automatic rig creation process | |
| CN108694740A (en) | Information processing equipment, information processing method and user equipment | |
| WO2024032464A1 (en) | Three-dimensional face reconstruction method, apparatus, and device, medium, and product | |
| CN109829849A (en) | A kind of generation method of training data, device and terminal | |
| US11126915B2 (en) | Information processing apparatus and information processing method for volume data visualization | |
| CN117036569B (en) | Three-dimensional model color generation network training method, color generation method and device | |
| CN106127818A (en) | A kind of material appearance based on single image obtains system and method | |
| CN115359163A (en) | Three-dimensional model generation system, three-dimensional model generation method, and three-dimensional model generation device | |
| CN110717978B (en) | Three-dimensional head reconstruction method based on single image | |
| CN117611542B (en) | Fetal intrauterine craniocerebral image-based detection method and system | |
| CN109145984A (en) | Method and apparatus for machine training | |
| CN117173315A (en) | Neural radiation field-based unbounded scene real-time rendering method, system and equipment | |
| CN109615588A (en) | A method of image reconstruction is solved the problems, such as based on depth autoregression model | |
| CN117095132B (en) | Three-dimensional reconstruction method and system based on implicit function | |
| CN117274446A (en) | Scene video processing method, device, equipment and storage medium | |
| CN112561782A (en) | Method for improving reality degree of simulation picture of offshore scene | |
| CN107221019B (en) | Chart conversion method and device | |
| JP2024502326A (en) | How to train a neural network configured to convert 2D images into 3D models | |
| CN115082636A (en) | Single image three-dimensional reconstruction method and equipment based on hybrid Gaussian network | |
| CN115035252A (en) | Three-dimensional reconstruction method and device for gas plant station based on nerve radiation field | |
| CN115049559A (en) | Model training method, human face image processing method, human face model processing device, electronic equipment and readable storage medium | |
| CN118172496B (en) | Three-dimensional reconstruction method, system, medium, device and program product |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190412 |