CN106296611A - The compressed sensing image recovery method of a kind of objective attribute target attribute auxiliary and system thereof - Google Patents
The compressed sensing image recovery method of a kind of objective attribute target attribute auxiliary and system thereof Download PDFInfo
- Publication number
- CN106296611A CN106296611A CN201610647440.5A CN201610647440A CN106296611A CN 106296611 A CN106296611 A CN 106296611A CN 201610647440 A CN201610647440 A CN 201610647440A CN 106296611 A CN106296611 A CN 106296611A
- Authority
- CN
- China
- Prior art keywords
- target
- subspace
- image
- dictionary
- vector
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 23
- 238000011084 recovery Methods 0.000 title claims abstract description 23
- 238000005192 partition Methods 0.000 claims abstract description 14
- 239000011159 matrix material Substances 0.000 claims description 59
- 238000005516 engineering process Methods 0.000 claims description 9
- 239000000203 mixture Substances 0.000 claims description 9
- 230000008901 benefit Effects 0.000 description 9
- 230000006835 compression Effects 0.000 description 7
- 238000007906 compression Methods 0.000 description 7
- 230000008859 change Effects 0.000 description 5
- 238000002474 experimental method Methods 0.000 description 5
- 230000007423 decrease Effects 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000007152 ring opening metathesis polymerisation reaction Methods 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000005057 refrigeration Methods 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000000686 essence Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 230000007704 transition 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
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Aiming, Guidance, Guns With A Light Source, Armor, Camouflage, And Targets (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides the compressed sensing image recovery method that a kind of objective attribute target attribute assists, and wherein, described method includes: initialization step, Subspace partition step, atom collection update step, sparse coefficient updates step, output step.The present invention also provides for the compressed sensing image recovery system of a kind of objective attribute target attribute auxiliary.The technical scheme that the present invention provides is in the case of Small object picture signal degree of rarefication the unknown, the auxiliary information that can reflect target characteristic is incorporated in the division of subspace, thus select the dictionary subspace mated the most accurately and effectively, it is achieved efficiently and quickly reconstructing of Small object picture signal.
Description
Technical field
The present invention relates to image procossing compressed sensing field, particularly relate to the compressed sensing image of a kind of objective attribute target attribute auxiliary
Restoration methods and system thereof.
Background technology
Compressed sensing technology is widely used in image procossing neck because being expected to realize compression ratio and the balance of reconstruction quality
Territory, and be forward position and the focus of signal processing neighborhood since it is born.Image is as a kind of natural sign, between neighbor
The spatial redundancy that causes of dependency, there is the time redundancy that dependency causes, different color planes or frequency spectrum between different frame
The spectral redundancy that the dependency of band causes, i.e. also exists redundancy, and image has closely on some sparse territories in view data
Like openness, it is possible to use compressive sensing theory image to be compressed and reconstruction processing.But, existing restructing algorithm
In fail to enter at reconstruction quality and between reconstructed velocity, obtain preferable balance, they often focus on reconstruction quality and have ignored weight
The structure time, in the case, there has been proposed stator spatial class recovery algorithms.
Subspace class OMP algorithm is because of simple in construction, it is easy to accomplish, and and quilt relatively low compared to other types algorithm operation quantity
Extensively apply in the engineering project of reality.Such algorithm includes OMP, StOMP, ROMP, CoSaMP, and SP algorithm.OMP algorithm
The size of its subspace is 1, and iterations therefore will be caused many when degree of rarefication K is excessive, and operand is big, and reconstructed velocity is slow.
StOMP algorithm an iteration process may select multiple matched atoms, so effectively reducing iterations, and then reduces the fortune of algorithm
Calculation amount, reduces reconstitution time.But very strong to the dependency of degree of rarefication, the most correctly estimating degree of rarefication K could correctly reconstruct
Go out signal.ROMP algorithm is according to relevance principle and regularization process atom to carry out twice screening, but this algorithm neutron
The selection in space depends on dependency between value, and relies on degree of rarefication K transition, and the correctness that degree of rarefication is estimated will affect algorithm
Convergence rate, quality reconstruction etc..CoSaMP algorithm selects multiple uncorrelated atom also by the thought of backtracking from atom
Rejecting the uncorrelated atom of part, an iteration may select multiple matched atoms.But it equally needs known degree of rarefication K, but
Real signal degree of rarefication K the unknown often, secondly, CoSaMP algorithm increases when often walking iteration and rejects atom institute foundation
Principle is different, and the estimation that the difference of standard may result in supporting collection is inaccurate.SP algorithm is the compromise of above-mentioned algorithm, its
Performance is optimum in above-mentioned algorithm.SP algorithm computation complexity is low, reconstruction accuracy is high, and has strict theoretical guarantee, when
When observing matrix A meets limited equidistant character, SP algorithm can go out the sparse letter of any K-by Accurate Reconstruction from its noiseless is observed
Number.It has similar pluses and minuses with CoSaMP algorithm.
Although above-mentioned all algorithms all have employed the concept in stator space, except the subspace size of OMP algorithm is in addition to 1,
Remaining algorithm subspace size is all higher than 1, and these algorithms all make iterations tail off because once selecting multiple matched atoms,
And then reduce operand and reconstitution time.But, still fail to reach the requirement of real-time, and above-mentioned algorithm all depended on unduly dilute
That dredges degree K, K estimates that correctness will determine that can signal be estimated correctly the time with signal reconstruction, meanwhile, in these algorithms
The selection of dictionary subspace all complete according to dependency between value, do not consider the physical feature attribute of picture signal.
Summary of the invention
In view of this, it is an object of the invention to provide a kind of objective attribute target attribute auxiliary compressed sensing image recovery method and
Its system, it is intended to solve in prior art in the case of Small object picture signal degree of rarefication the unknown, how by the spy of target
Levy attribute (as openness in target size apriority, coherency and spatial domain etc.) and carry out the reconstruct that antithetical phrase spatial class OMP class algorithm exists
Speed compared with slow, reconstruction accuracy is relatively low and depends on the problem that the defects such as degree of rarefication carry out making up unduly.
The present invention proposes the compressed sensing image recovery method of a kind of objective attribute target attribute auxiliary, it is characterised in that described method
Including:
Initialization step: by redundancy vector rt-1, atom collection A and these parameters of iterations t initialize, wherein,
rt-1Being initialized as decrement direction finding amount y of input, A is initialized as empty set, and t is initialized as 1；
Subspace partition step: generate and image sequential matrix of the same size D ∈ Rw×h, w and h is image dimension, in institute
State the number preserving 1 to w × h in sequential matrix D, according to the target size preset, described sequential matrix D is divided into some simultaneously
Two dimension sub-block, and each sub-block divided in described sequential matrix D is all converted into column vector according to the mode of row stacking, with
Element in each column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=
1,2 ..., k), thus realize the division to dictionary space；
Atom collection updates step: obtain redundancy vector rt-1At each the dictionary subspace Φ dividediIn projectionAnd calculate PiEnergyFind out dictionary subspace corresponding to ceiling capacity simultaneously
Index subscriptAccording to calculated index subscript λtFind out the dictionary subspace of correspondenceAnd utilize word
Atom collection A is updated, i.e. by allusion quotation subspaceWherein, AtIt is the current atom collection updated, At-1It is upper one
The atom collection that secondary iteration updates.After updating atom collection A each time, by λtCorresponding dictionary subspaceSet to 0 to guarantee not weigh
Multiple value；
Sparse coefficient updates step: utilize the atom collection A after current iteration renewaltSparse coefficient is carried out least square estimate
Meter, obtains signal at atom collection AtIn sparse coefficient componentAnd update redundancy vector rt=y-
Axt；
Output step: according to the renewal redundancy vector r obtainedtIt is iterated stop technology, as redundancy vector rtNorm
During more than preset constant ε, then jump to described atom collection renewal step and carry out next iteration, otherwise terminate iteration and export letter
The final estimated value of number sparse coefficient.
Preferably, described Subspace partition step specifically includes:
Assuming that target size is B=a × b, then set template size as a × b, generate the two dimension order of correspondence image size
Matrix D ∈ Rw×h, wherein, a, b are respectively target priori length and target priori width, and w, h are respectively infrared small target image
Length and width, w is divided exactly a, h and is divided exactly b；
The row stacking value that element is correspondence image pixel position in two dimension sequential matrix D, according to the elder generation of target size
Test size two dimension sequential matrix D is divided, obtain D={B1,B2,...,Bi,...Bk, wherein
When image length w aliquant target priori length a, during picture traverse h aliquant target priori width b, then
It is set as target being completely covered by template size according to practical situation, also can be divided exactly by the length of image and width
Value；
The each sub-block divided in described sequential matrix D is all converted into column vector, with often according to the mode of row stacking
Element in one column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=1,
2 ..., k), i.e. divide dictionary subspace, Φi=Φ (Vec (Bi)), i=1,2 ..., k.
On the other hand, the present invention also provides for the compressed sensing image recovery system of a kind of objective attribute target attribute auxiliary, described system
Including:
Initialization module, for by redundancy vector rt-1, atom collection A and these parameters of iterations t initialize,
Wherein, rt-1Being initialized as decrement direction finding amount y of input, A is initialized as empty set, and t is initialized as 1；
Subspace partition module, for generating and image sequential matrix of the same size D ∈ Rw×h, w and h is image dimension,
In described sequential matrix D, preserve the number of 1 to w × h, according to the target size preset, described sequential matrix D is divided into simultaneously
Some two dimension sub-blocks, and by each sub-block divided in described sequential matrix D all according to row stacking mode change in column to
Amount, is constituted dictionary subspace with the vector that the element in each column vector picks out correspondence for index from the Φ of dictionary space
Φi(i=1,2 ..., k), thus realize the division to dictionary space；
Atom collection more new module, is used for obtaining redundancy vector rt-1At each the dictionary subspace Φ dividediIn projectionAnd calculate PiEnergyFind out the dictionary subspace that ceiling capacity is corresponding simultaneously
Index subscriptAccording to calculated index subscript λtFind out the dictionary subspace of correspondenceAnd utilize
Atom collection A is updated, i.e. by dictionary subspaceWherein, AtIt is the current atom collection updated, At-1On being
The atom collection that an iteration updates.After updating atom collection A each time, by λtCorresponding dictionary subspaceSet to 0 to guarantee not
Repeat value；
Sparse coefficient more new module, the atom collection A after utilizing current iteration to updatetSparse coefficient is carried out a young waiter in a wineshop or an inn
Take advantage of estimation, obtain signal at atom collection AtIn sparse coefficient componentAnd update redundancy vector rt=
y-Axt；
Output module, for according to the renewal redundancy vector r obtainedtIt is iterated stop technology, as redundancy vector rt's
When norm is more than preset constant ε, then jumps to described atom collection and update step and carry out next iteration, otherwise terminate iteration defeated
Go out the final estimated value of signal sparse coefficient.
Preferably, described Subspace partition module specifically for:
Assuming that target size is B=a × b, then set template size as a × b, generate the two dimension order of correspondence image size
Matrix D ∈ Rw×h, wherein, a, b are respectively target priori length and target priori width, and w, h are respectively infrared small target image
Length and width, w is divided exactly a, h and is divided exactly b；
The row stacking value that element is correspondence image pixel position in two dimension sequential matrix D, according to the elder generation of target size
Test size two dimension sequential matrix D is divided, obtain D={B1,B2,...,Bi,...Bk, wherein
When image length w aliquant target priori length a, during picture traverse h aliquant target priori width b, then
It is set as target being completely covered by template size according to practical situation, also can be divided exactly by the length of image and width
Value；
The each sub-block divided in described sequential matrix D is all converted into column vector, with often according to the mode of row stacking
Element in one column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=1,
2 ..., k), i.e. divide dictionary subspace, Φi=Φ (Vec (Bi)), i=1,2 ..., k.
The technical scheme that the present invention provides, in the case of Small object picture signal degree of rarefication the unknown, can reflect that target is special
The auxiliary information levied is incorporated in the division of subspace, thus selects the dictionary subspace mated the most accurately and effectively, it is achieved
Efficiently and quickly reconstructing of Small object picture signal.Compared with prior art, present invention have an advantage that (1) present invention not
Rely on the degree of rarefication of signal, eliminate the reliance on dependency between value and select dictionary subspace；(2) present invention can be effectively realized target
Reconstruct, improve reconstruction quality；(3) present invention effectively can reduce iterations and decrease operand, is greatly improved
The speed of reconstruct.
Accompanying drawing explanation
Fig. 1 is the compressed sensing image recovery method flow chart of objective attribute target attribute auxiliary in an embodiment of the present invention；
Fig. 2 be an embodiment of the present invention utilizes by model be IR300, frame frequency be the refrigeration mode thermal infrared imager of 50 frames
From the foreground image single infrared small target image of infrared small target image, two infrared small targets of different scenes collections
Image and multiple infrared small target image；
Fig. 3 is to arrange the quality reconstruction figure to infrared small target image according to experiment condition in an embodiment of the present invention；
Fig. 4 is the internal structure of the compressed sensing image recovery system 10 of objective attribute target attribute auxiliary in an embodiment of the present invention
Schematic diagram.
Detailed description of the invention
In order to make the purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, right
The present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, and
It is not used in the restriction present invention.
In order to solve the problem that prior art exists, the thinking that the present invention proposes is: when being reconstructed image, by
Objective attribute target attribute auxuliary subspace divides, and then assistant images reconstruct.The present invention is directed to existing stator spatial class OMP algorithm depends on
Rely degree of rarefication and according to the problem of Correlation selection subspace between value, it is proposed that divide dictionary subspace according to target size, have
Effect decreases the number of times of iteration, promotes reconstruction quality and do not rely on degree of rarefication.The present invention concrete grammar to the reconstruct of image:
First generate and correspondence image sequential matrix of the same size, according to target priori size stripe sequence matrix, the order that will divide
In matrix each piece is converted into column vector according to the form of row stacking, incompatible with the element in each column vector for indexed set
Carry out the division of selecting of dictionary subspace, referred to as subspace.After dividing subspace, obtain redundancy with certain relevance principle and exist
Projection in each subspace, and find out the dictionary subspace that maximal projection value is corresponding, expand atom with this dictionary subspace
Set, and then carry out the estimation of signal and redundancy.
The compressed sensing image recovery method of a kind of objective attribute target attribute auxiliary that the present invention provides has to be independent of signal dilute
In the case of dredging degree, it is possible to the advantage efficiently and fast realizing the reconstruct of Small object picture signal, in actual applications can be big
Reduce greatly data volume, carrying cost and transmission bandwidth, time-consuming.
Hereinafter the compressed sensing image recovery method assisting a kind of objective attribute target attribute provided by the present invention is carried out in detail
Explanation.
Refer to Fig. 1, for the compressed sensing image recovery method flow process of objective attribute target attribute auxiliary in an embodiment of the present invention
Figure.
In step sl, by redundancy vector rt-1, atom collection A and these parameters of iterations t initialize, wherein,
rt-1Being initialized as decrement direction finding amount y of input, A is initialized as empty set, and t is initialized as 1.
In step s 2, Subspace partition step: generate and image sequential matrix of the same size D ∈ Rw×h, w and h is figure
As dimension, described sequential matrix D preserves the number of 1 to w × h, simultaneously according to the target size preset by described sequential matrix D
It is divided into some two dimension sub-blocks, and each sub-block divided in described sequential matrix D is all changed according to the mode of row stacking
Become column vector, constituted dictionary with the vector that the element in each column vector picks out correspondence for index from the Φ of dictionary space
Space Φi(i=1,2 ..., k), thus realize the division to dictionary space.
In the present embodiment, described Subspace partition step S2 specifically includes:
Assuming that target size is B=a × b, then set template size as a × b, generate the two dimension order of correspondence image size
Matrix D ∈ Rw×h, wherein, a, b are respectively target priori length and target priori width, and w, h are respectively infrared small target image
Length and width, w is divided exactly a, h and is divided exactly b；
The row stacking value that element is correspondence image pixel position in two dimension sequential matrix D, according to the elder generation of target size
Test size two dimension sequential matrix D is divided, obtain D={B1,B2,...,Bi,...Bk, wherein
When image length w aliquant target priori length a, during picture traverse h aliquant target priori width b, then
It is set as target being completely covered by template size according to practical situation, also can be divided exactly by the length of image and width
Value；
The each sub-block divided in described sequential matrix D is all converted into column vector, with often according to the mode of row stacking
Element in one column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=1,
2 ..., k), i.e. divide dictionary subspace, Φi=Φ (Vec (Bi)), i=1,2 ..., k.
In step s3, atom collection updates step: obtain redundancy vector rt-1At each the dictionary subspace Φ dividedi
In projectionAnd calculate PiEnergyFind out the word that ceiling capacity is corresponding simultaneously
The index subscript of allusion quotation subspaceAccording to calculated index subscript λtFind out the dictionary subspace of correspondenceAnd utilize dictionary subspace that atom collection A is updated, i.e.Wherein, AtIt it is the current atom updated
Collection, At-1It it is the atom collection of last iteration renewal.After updating atom collection A each time, by λtCorresponding dictionary subspacePut
0 to guarantee not repeat value.
In step s 4, sparse coefficient updates step: utilize the atom collection A after current iteration renewaltSparse coefficient is entered
Row least-squares estimation, obtains signal at atom collection AtIn sparse coefficient componentAnd update redundancy
Vector rt=y-Axt。
In step s 5, output step: according to the renewal redundancy vector r obtainedtIt is iterated stop technology, works as redundancy
Vector rtNorm more than preset constant ε time, then jump to described atom collection update step carry out next iteration, otherwise terminate
Iteration the final estimated value of output signal sparse coefficient.
Below in conjunction with specific embodiment, in order to verify the inventive method in the high efficiency that picture signal is reconstructed and
Rapidity.It is IR300 by model that the present invention utilizes, frame frequency be 50 frames refrigeration mode thermal infrared imager from different scenes gather red
The foreground image single infrared small target image of outer Small object image, two infrared small target images and multiple infrared little mesh
Logo image (as shown in Figure 2) is tested, and for aesthetic property, finally result and the background of reconstruct is merged the original of formation estimation
Image.Wherein weigh reconstruction quality, with reconstitution time measure algorithm rapidity with more objectively Y-PSNR.
The experimental configuration of the present invention specifically includes: two dimension sequential matrix D ∈ Rw×h, image size is w × h, w=240, h=
320；Signal length N=w × h=76800 (the picture signal dimension of row heap poststack)；Compression measures dimension M=19200；Compression
RatioPerception matrix Φ ∈ RM×N(vector immediately of Gaussian distributed)；Compression measures y ∈ RM×1；Target chi
(w is divided exactly a, h and is divided exactly b) in very little B=a × b=4 × 4.
First according to the priori size of target size, two dimension sequential matrix D is divided, obtain D={B1,B2,...,
Bi,...Bk, whereinSo corresponding Subspace partition is Φi=Φ (Vec (Bi)), i=1,
2,...,k.Computing redundancy vector projection P on each subspacei, and obtain PiENERGY Ei, find out ceiling capacity simultaneously
Index subscript λ of corresponding dictionary subspacet.According to calculated index subscript λtFind out the dictionary subspace of correspondenceI.e.
The sub-dictionary the strongest with redundancy vector correlation, is updated atom set with this subspaceThis experiment
Just expanded c=a × b=16 dictionary atom by an iteration, the most once circulation can recover 16 sparse coefficient points.Note
After updating atom collection A each time, by λtCorresponding dictionary subspaceSet to 0 to guarantee not repeat value.
Arranging according to above-mentioned experiment condition, the present invention is to the quality reconstruction figure of infrared small target image as shown in Figure 3.
Above-mentioned experiment is only single compressed more next time than the quality reconstruction figure tested, for the impact avoiding occasionality to cause,
And more objectively explanation the present invention is directed to the effectiveness of image reconstruction, by 100 width infrared small target figures under Same Scene
As being reconstructed, and count the average peak signal to noise ratio of present invention reconstruct, the average peak of the present invention under different compression ratios
Signal to noise ratio is as shown in table 1.
The average peak signal to noise ratio of table 1 the inventive method reconstruct
Being reached a conclusion by Fig. 3 and table 1, the present invention all can realize reconstruct to the infrared small target image under different scenes, and
The Y-PSNR of reconstruct is higher, has absolutely proved effectiveness of the invention and accuracy thereof.
The present invention mainly solves the problem of the reconstitution time length that stator spatial class algorithm exists, so will pass through herein
Experiment statistics goes out under different compression ratios, and the present invention reconstructs the average time of different scene infrared image, as shown in table 2.
The reconstitution time of table 2 the inventive method
2 conclusions can be drawn by table 2, one be compression ratio relatively low time, the present invention can go out different scene with accurate reconstruction
Infrared small target image；Two is that the present invention is less to the reconstitution time of infrared small target figure.Above-mentioned conclusion has absolutely proved this
Bright rapidity.
The compressed sensing image recovery method of a kind of objective attribute target attribute auxiliary that the present invention provides is dilute in Small object picture signal
In the case of dredging degree the unknown, the auxiliary information that can reflect target characteristic is incorporated in the division of subspace, thus accurate and effective
The dictionary subspace mated the most of selection, it is achieved efficiently and quickly reconstructing of Small object picture signal.With prior art phase
Ratio, present invention have an advantage that (1) present invention is independent of the degree of rarefication of signal, eliminates the reliance on dependency between value and selects dictionary
Space；(2) present invention can be effectively realized the reconstruct of target, improves reconstruction quality；(3) present invention can effectively reduce repeatedly
Generation number and decrease operand, be greatly improved the speed of reconstruct.
The specific embodiment of the invention also provides for the compressed sensing image recovery system 10 of a kind of objective attribute target attribute auxiliary, mainly
Including:
Initialization module 11, for by redundancy vector rt-1, atom collection A and these parameters of iterations t carry out initially
Change, wherein, rt-1Being initialized as decrement direction finding amount y of input, A is initialized as empty set, and t is initialized as 1；
Subspace partition module 12, for generating and image sequential matrix of the same size D ∈ Rw×h, w and h is image dimension
Number, preserves the number of 1 to w × h in described sequential matrix D, is divided by described sequential matrix D according to the target size preset simultaneously
For some two dimension sub-blocks, and by each sub-block divided in described sequential matrix D all according to row stacking mode change in column
Vector, is constituted dictionary subspace with the vector that the element in each column vector picks out correspondence for index from the Φ of dictionary space
Φi(i=1,2 ..., k), thus realize the division to dictionary space；
Atom collection more new module 13, is used for obtaining redundancy vector rt-1At each the dictionary subspace Φ dividediIn throwing
ShadowAnd calculate PiEnergyFind out the dictionary subspace that ceiling capacity is corresponding simultaneously
Index subscriptAccording to calculated index subscript λtFind out the dictionary subspace of correspondenceAnd utilize
Atom collection A is updated, i.e. by dictionary subspaceWherein, AtIt is the current atom collection updated, At-1On being
The atom collection that an iteration updates.After updating atom collection A each time, by λtCorresponding dictionary subspaceSet to 0 to guarantee not
Repeat value；
Sparse coefficient more new module 14, the atom collection A after utilizing current iteration to updatetSparse coefficient is carried out minimum
Two take advantage of estimation, obtain signal at atom collection AtIn sparse coefficient componentAnd update redundancy vector rt
=y-Axt；
Output module 15, for according to the renewal redundancy vector r obtainedtIt is iterated stop technology, as redundancy vector rt
Norm more than preset constant ε time, then jump to described atom collection and update step and carry out next iteration, otherwise terminate iteration also
The final estimated value of output signal sparse coefficient.
The compressed sensing image recovery system 10 of a kind of objective attribute target attribute auxiliary that the present invention provides, in Small object picture signal
In the case of degree of rarefication the unknown, the auxiliary information that can reflect target characteristic is incorporated in the division of subspace, thus accurately has
The dictionary subspace that the selection of effect is mated the most, it is achieved efficiently and quickly reconstructing of Small object picture signal.
Refer to Fig. 4, show the compressed sensing image recovery system of objective attribute target attribute auxiliary in an embodiment of the present invention
The structural representation of 10.
In the present embodiment, the compressed sensing image recovery system 10 of objective attribute target attribute auxiliary, mainly include initializing mould
Block 11, Subspace partition module 12, atom collection more new module 13, sparse coefficient more new module 14 and output module 15.
Initialization module 11, for by redundancy vector rt-1, atom collection A and these parameters of iterations t carry out initially
Change, wherein, rt-1Being initialized as decrement direction finding amount y of input, A is initialized as empty set, and t is initialized as 1.
Subspace partition module 12, for generating and image sequential matrix of the same size D ∈ Rw×h, w and h is image dimension
Number, preserves the number of 1 to w × h in described sequential matrix D, is divided by described sequential matrix D according to the target size preset simultaneously
For some two dimension sub-blocks, and by each sub-block divided in described sequential matrix D all according to row stacking mode change in column
Vector, is constituted dictionary subspace with the vector that the element in each column vector picks out correspondence for index from the Φ of dictionary space
Φi(i=1,2 ..., k), thus realize the division to dictionary space.
In the present embodiment, described Subspace partition module 12 specifically for:
Assuming that target size is B=a × b, then set template size as a × b, generate the two dimension order of correspondence image size
Matrix D ∈ Rw×h, wherein, a, b are respectively target priori length and target priori width, and w, h are respectively infrared small target image
Length and width, w is divided exactly a, h and is divided exactly b；
The row stacking value that element is correspondence image pixel position in two dimension sequential matrix D, according to the elder generation of target size
Test size two dimension sequential matrix D is divided, obtain D={B1,B2,...,Bi,...Bk, wherein
When image length w aliquant target priori length a, during picture traverse h aliquant target priori width b, then
It is set as target being completely covered by template size according to practical situation, also can be divided exactly by the length of image and width
Value；
The each sub-block divided in described sequential matrix D is all converted into column vector, with often according to the mode of row stacking
Element in one column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=1,
2 ..., k), i.e. divide dictionary subspace, Φi=Φ (Vec (Bi)), i=1,2 ..., k.
Atom collection more new module 13, is used for obtaining redundancy vector rt-1At each the dictionary subspace Φ dividediIn throwing
ShadowAnd calculate PiEnergyFind out the dictionary subspace that ceiling capacity is corresponding simultaneously
Index subscriptAccording to calculated index subscript λtFind out the dictionary subspace of correspondenceAnd utilize
Atom collection A is updated, i.e. by dictionary subspaceWherein, AtIt is the current atom collection updated, At-1On being
The atom collection that an iteration updates.After updating atom collection A each time, by λtCorresponding dictionary subspaceSet to 0 to guarantee not
Repeat value.
Sparse coefficient more new module 14, the atom collection A after utilizing current iteration to updatetSparse coefficient is carried out minimum
Two take advantage of estimation, obtain signal at atom collection AtIn sparse coefficient componentAnd update redundancy vector rt
=y-Axt。
Output module 15, for according to the renewal redundancy vector r obtainedtIt is iterated stop technology, as redundancy vector rt
Norm more than preset constant ε time, then jump to described atom collection and update step and carry out next iteration, otherwise terminate iteration also
The final estimated value of output signal sparse coefficient.
The compressed sensing image recovery system 10 of a kind of objective attribute target attribute auxiliary that the present invention provides, in Small object picture signal
In the case of degree of rarefication the unknown, the auxiliary information that can reflect target characteristic is incorporated in the division of subspace, thus accurately has
The dictionary subspace that the selection of effect is mated the most, it is achieved efficiently and quickly reconstructing of Small object picture signal.With prior art
Compare, present invention have an advantage that (1) present invention is independent of the degree of rarefication of signal, eliminate the reliance on dependency between value and select dictionary
Subspace；(2) present invention can be effectively realized the reconstruct of target, improves reconstruction quality；(3) present invention can effectively reduce
Iterations and decrease operand, be greatly improved the speed of reconstruct.
The present invention has in the case of being independent of signal degree of rarefication, it is possible to efficiently and fast realize Small object image letter
Number reconstruct advantage, data volume, carrying cost and transmission bandwidth can be substantially reduced in actual applications, time-consuming.
The present invention is widely used, it is adaptable to the Small object image procossing of multiple applications, such as the infrared figure of military field
Picture, the CT image of medical domain and the Mura image etc. of field of industrial production.
It should be noted that in above-described embodiment, included unit is to carry out dividing according to function logic,
But it is not limited to above-mentioned division, as long as being capable of corresponding function；It addition, the specific name of each functional unit is also
Only to facilitate mutually distinguish, it is not limited to protection scope of the present invention.
It addition, one of ordinary skill in the art will appreciate that all or part of step realizing in the various embodiments described above method
The program that can be by completes to instruct relevant hardware, and corresponding program can be stored in an embodied on computer readable storage and be situated between
In matter, described storage medium, such as ROM/RAM, disk or CD etc..
The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all essences in the present invention
Any amendment, equivalent and the improvement etc. made within god and principle, should be included within the scope of the present invention.
Claims (4)
1. the compressed sensing image recovery method of an objective attribute target attribute auxiliary, it is characterised in that described method includes:
Initialization step: by redundancy vector rt-1, atom collection A and these parameters of iterations t initialize, wherein, rt-1
Being initialized as decrement direction finding amount y of input, A is initialized as empty set, and t is initialized as 1；
Subspace partition step: generate and image sequential matrix of the same size D ∈ Rw×h, w and h is image dimension, described suitable
Sequence matrix D preserves the number of 1 to w × h, according to the target size preset, described sequential matrix D is divided into some two dimensions simultaneously
Sub-block, and each sub-block divided in described sequential matrix D is all converted into column vector, with each according to the mode of row stacking
Element in individual column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=1,
2 ..., k), thus realize the division to dictionary space；
Atom collection updates step: obtain redundancy vector rt-1At each the dictionary subspace Φ dividediIn projectionAnd calculate PiEnergyFind out the dictionary subspace that ceiling capacity is corresponding simultaneously
Index subscriptAccording to calculated index subscript λtFind out the dictionary subspace of correspondenceAnd utilize
Atom collection A is updated, i.e. by dictionary subspaceWherein, AtIt is the current atom collection updated, At-1On being
The atom collection that an iteration updates.After updating atom collection A each time, by λtCorresponding dictionary subspaceSet to 0 to guarantee not
Repeat value；
Sparse coefficient updates step: utilize the atom collection A after current iteration renewaltSparse coefficient is carried out least-squares estimation, asks
Go out signal at atom collection AtIn sparse coefficient componentAnd update redundancy vector rt=y-Axt；
Output step: according to the renewal redundancy vector r obtainedtIt is iterated stop technology, as redundancy vector rtNorm be more than
During preset constant ε, then jump to described atom collection renewal step and carry out next iteration, otherwise terminate iteration and output signal is dilute
The final estimated value of sparse coefficient.
2. the compressed sensing image recovery method of objective attribute target attribute auxiliary as claimed in claim 1, it is characterised in that described son is empty
Between partiting step specifically include:
Assuming that target size is B=a × b, then set template size as a × b, generate the two-dimentional sequential matrix of correspondence image size
D∈Rw×h, wherein, a, b are respectively target priori length and target priori width, and w, h are respectively the length of infrared small target image
With width, w is divided exactly a, h and is divided exactly b；
The row stacking value that element is correspondence image pixel position in two dimension sequential matrix D, the priori according to target size is big
Little to two dimension sequential matrix D divide, obtain D={B1,B2,...,Bi,...Bk, wherein
When image length w aliquant target priori length a, during picture traverse h aliquant target priori width b, then basis
Template size is set as target being completely covered by practical situation, the value that also can be divided exactly by the length of image and width；
The each sub-block divided in described sequential matrix D is all converted into column vector, with each according to the mode of row stacking
Element in column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=1,
2 ..., k), i.e. divide dictionary subspace, Φi=Φ (Vec (Bi)), i=1,2 ..., k.
3. the compressed sensing image recovery system of an objective attribute target attribute auxiliary, it is characterised in that described system includes:
Initialization module, for by redundancy vector rt-1, atom collection A and these parameters of iterations t initialize, wherein,
rt-1Being initialized as decrement direction finding amount y of input, A is initialized as empty set, and t is initialized as 1；
Subspace partition module, for generating and image sequential matrix of the same size D ∈ Rw×h, w and h is image dimension, in institute
State the number preserving 1 to w × h in sequential matrix D, according to the target size preset, described sequential matrix D is divided into some simultaneously
Two dimension sub-block, and each sub-block divided in described sequential matrix D is all converted into column vector according to the mode of row stacking, with
Element in each column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=
1,2 ..., k), thus realize the division to dictionary space；
Atom collection more new module, is used for obtaining redundancy vector rt-1At each the dictionary subspace Φ dividediIn projectionAnd calculate PiEnergyFind out the dictionary subspace that ceiling capacity is corresponding simultaneously
Index subscriptAccording to calculated index subscript λtFind out the dictionary subspace of correspondenceAnd utilize
Atom collection A is updated, i.e. by dictionary subspaceWherein, AtIt is the current atom collection updated, At-1On being
The atom collection that an iteration updates.After updating atom collection A each time, by λtCorresponding dictionary subspaceSet to 0 to guarantee not
Repeat value；
Sparse coefficient more new module, the atom collection A after utilizing current iteration to updatetSparse coefficient is carried out least square estimate
Meter, obtains signal at atom collection AtIn sparse coefficient componentAnd update redundancy vector rt=y-
Axt；
Output module, for according to the renewal redundancy vector r obtainedtIt is iterated stop technology, as redundancy vector rtNorm
During more than preset constant ε, then jump to described atom collection renewal step and carry out next iteration, otherwise terminate iteration and export letter
The final estimated value of number sparse coefficient.
4. the compressed sensing image recovery system of objective attribute target attribute auxiliary as claimed in claim 3, it is characterised in that described son is empty
Between divide module specifically for:
Assuming that target size is B=a × b, then set template size as a × b, generate the two-dimentional sequential matrix of correspondence image size
D∈Rw×h, wherein, a, b are respectively target priori length and target priori width, and w, h are respectively the length of infrared small target image
With width, w is divided exactly a, h and is divided exactly b；
The row stacking value that element is correspondence image pixel position in two dimension sequential matrix D, the priori according to target size is big
Little to two dimension sequential matrix D divide, obtain D={B1,B2,...,Bi,...Bk, wherein
When image length w aliquant target priori length a, during picture traverse h aliquant target priori width b, then basis
Template size is set as target being completely covered by practical situation, the value that also can be divided exactly by the length of image and width；
The each sub-block divided in described sequential matrix D is all converted into column vector, with each according to the mode of row stacking
Element in column vector is the vector composition dictionary subspace Φ that correspondence picked out from the Φ of dictionary space in indexi(i=1,
2 ..., k), i.e. divide dictionary subspace, Φi=Φ (Vec (Bi)), i=1,2 ..., k.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610647440.5A CN106296611B (en) | 2016-08-09 | 2016-08-09 | A kind of compressed sensing image recovery method and its system using objective attribute target attribute auxiliary |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610647440.5A CN106296611B (en) | 2016-08-09 | 2016-08-09 | A kind of compressed sensing image recovery method and its system using objective attribute target attribute auxiliary |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106296611A true CN106296611A (en) | 2017-01-04 |
| CN106296611B CN106296611B (en) | 2019-04-16 |
Family
ID=57667399
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610647440.5A Expired - Fee Related CN106296611B (en) | 2016-08-09 | 2016-08-09 | A kind of compressed sensing image recovery method and its system using objective attribute target attribute auxiliary |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106296611B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018027584A1 (en) * | 2016-08-09 | 2018-02-15 | 深圳大学 | Method and system for restoring image using target attribute assisted compression perception |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103198500A (en) * | 2013-04-03 | 2013-07-10 | 西安电子科技大学 | Compressed sensing image reconstruction method based on principal component analysis (PCA) redundant dictionary and direction information |
| CN105811992A (en) * | 2016-03-01 | 2016-07-27 | 深圳大学 | Compressed sensing method and system capable of separating sparse signals |
-
2016
- 2016-08-09 CN CN201610647440.5A patent/CN106296611B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103198500A (en) * | 2013-04-03 | 2013-07-10 | 西安电子科技大学 | Compressed sensing image reconstruction method based on principal component analysis (PCA) redundant dictionary and direction information |
| CN105811992A (en) * | 2016-03-01 | 2016-07-27 | 深圳大学 | Compressed sensing method and system capable of separating sparse signals |
Non-Patent Citations (2)
| Title |
|---|
| LIANG ZHONGYIN等: "Sub-sampled IFFT based compressive sampling", 《2015 IEEE REGION 10 CONFERENCE》 * |
| 沈燕飞等: "基于秩极小化的压缩感知图像恢复算法", 《电子学报》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018027584A1 (en) * | 2016-08-09 | 2018-02-15 | 深圳大学 | Method and system for restoring image using target attribute assisted compression perception |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106296611B (en) | 2019-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11238602B2 (en) | Method for estimating high-quality depth maps based on depth prediction and enhancement subnetworks | |
| US10706593B2 (en) | Method and system for image reconstruction using target attribute assisted compressive sensing | |
| US20190303731A1 (en) | Target detection method and device, computing device and readable storage medium | |
| Fang et al. | Pyramid scene parsing network in 3D: Improving semantic segmentation of point clouds with multi-scale contextual information | |
| CN106408524A (en) | Two-dimensional image-assisted depth image enhancement method | |
| CN105869167A (en) | High-resolution depth map acquisition method based on active and passive fusion | |
| CN103745498B (en) | A kind of method for rapidly positioning based on image | |
| CN105388476B (en) | A kind of chromatography SAR imaging methods based on joint sparse model | |
| CN111582091B (en) | Pedestrian recognition method based on multi-branch convolutional neural network | |
| CN111383741A (en) | Method, device and equipment for establishing medical imaging model and storage medium | |
| Chu et al. | Multi-energy CT reconstruction based on low rank and sparsity with the split-bregman method (MLRSS) | |
| CN107610219A (en) | The thick densification method of Pixel-level point cloud that geometry clue perceives in a kind of three-dimensional scenic reconstruct | |
| CN109146792A (en) | Chip image super resolution ratio reconstruction method based on deep learning | |
| CN103780267A (en) | Measurement matrix design method based on LDPC matrix | |
| CN106296611A (en) | The compressed sensing image recovery method of a kind of objective attribute target attribute auxiliary and system thereof | |
| CN106910246A (en) | Speckle three-D imaging method and device that space-time is combined | |
| CN104616304A (en) | Self-adapting support weight stereo matching method based on field programmable gate array (FPGA) | |
| CN112529098A (en) | Dense multi-scale target detection system and method | |
| CN107644393A (en) | A kind of Parallel Implementation method of the abundance algorithm for estimating based on GPU | |
| CN107330912A (en) | A kind of target tracking method of rarefaction representation based on multi-feature fusion | |
| CN106652023B (en) | A kind of method and system of the extensive unordered quick exercise recovery structure of image | |
| CN114529519B (en) | Image compressed sensing reconstruction method and system based on multi-scale depth cavity residual error network | |
| US20150332447A1 (en) | Method and apparatus for generating spanning tree, method and apparatus for stereo matching, method and apparatus for up-sampling, and method and apparatus for generating reference pixel | |
| Ma et al. | Reduced-reference stereoscopic image quality assessment based on entropy of gradient primitives | |
| Chaux et al. | A parallel proximal splitting method for disparity estimation from multicomponent images under illumination variation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190416 |