CN108765280A - A kind of high spectrum image spatial resolution enhancement method - Google Patents
A kind of high spectrum image spatial resolution enhancement method Download PDFInfo
- Publication number
- CN108765280A CN108765280A CN201810274921.5A CN201810274921A CN108765280A CN 108765280 A CN108765280 A CN 108765280A CN 201810274921 A CN201810274921 A CN 201810274921A CN 108765280 A CN108765280 A CN 108765280A
- Authority
- CN
- China
- Prior art keywords
- image
- spectrum
- dictionary
- matrix
- resolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 88
- 238000001228 spectrum Methods 0.000 title claims abstract description 75
- 239000011159 matrix material Substances 0.000 claims abstract description 58
- 230000003595 spectral effect Effects 0.000 claims abstract description 26
- 230000001360 synchronised effect Effects 0.000 claims abstract description 7
- 238000000605 extraction Methods 0.000 claims abstract description 5
- 238000000985 reflectance spectrum Methods 0.000 claims abstract description 5
- 238000012549 training Methods 0.000 claims description 13
- 239000013598 vector Substances 0.000 claims description 11
- 230000009466 transformation Effects 0.000 claims description 7
- 238000009825 accumulation Methods 0.000 claims description 4
- 230000003472 neutralizing effect Effects 0.000 claims description 3
- 238000002310 reflectometry Methods 0.000 abstract description 3
- 238000004321 preservation Methods 0.000 abstract description 2
- 238000007796 conventional method Methods 0.000 abstract 1
- 239000000284 extract Substances 0.000 abstract 1
- 230000000694 effects Effects 0.000 description 15
- 238000011156 evaluation Methods 0.000 description 8
- 238000003384 imaging method Methods 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000000354 decomposition reaction Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 239000000463 material Substances 0.000 description 3
- 238000004445 quantitative analysis Methods 0.000 description 3
- 241000208340 Araliaceae Species 0.000 description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 235000008434 ginseng Nutrition 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000013139 quantization Methods 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 230000001186 cumulative effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 210000003739 neck Anatomy 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 230000010287 polarization Effects 0.000 description 1
- 238000000513 principal component analysis Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000004611 spectroscopical analysis Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- 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
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of high spectrum image super-resolution Enhancement Method, this method extracts the reflectance spectrum of different scenes in image first, by compressed sensing dictionary learning algorithm obtain strong sparsity, weak coherence spectrum dictionary；Secondly using sparsity, nonnegativity and the space structure self-similarity of signal in EO-1 hyperion scene, the low spectrum picture of high spatial resolution is obtained from the scene reflectivity spectrum resolution of extraction, solves to obtain sparse coding matrix by synchronous orthogonal matching pursuit algorithm；Last combined spectral dictionary and sparse coding matrix obtain target image.Since space and the spectral information of image is used in combination, analogue data and truthful data the experimental results showed that, the method of the present invention is compared to conventional method and matrix disassembling method, high spectrum image detailed information and texture structure can effectively be rebuild, effectively improve wave band average peak signal to noise ratio, wave band average structure similarity and spectrum angle map, and preferably spectral preservation information.
Description
Technical field
The present invention is under the jurisdiction of photonics image detection and image processing field, is related to a kind of new EO-1 hyperion polarization image oversubscription
Resolution Enhancement Method.
Background technology
The important component of high light spectrum image-forming technology comprehensive earth observation since being the 1980s, by imaging technique
It is combined together with spectral technique, including abundant space, radiation and spectral information, continuous, collection of illustrative plates the spies with spectrum
Property.High spectrum image has been widely used for the necks such as geologic prospect, marine monitoring, battle reconnaissance due to good spectral characteristic
Domain.In addition in medical imaging, the spatially resolved spectroscopy imaging obtained by high light spectrum image-forming technology is provided about tissue life
The diagnostic message of Neo-Confucianism, morphology and composition.The spatial resolution of remote sensing images is weigh target in hyperspectral remotely sensed image quality one
A important indicator reflects the ground minimum target size that can be seen on image, determines in practical application on a surface target
Detect recognition capability.It is influenced by image-forming condition and imaging circumstances, the image spatial resolution of acquisition is relatively low, causes largely to mix
Pixel substantially reduces subsequent detection and recognition performance, and then influences many applications in military and civilian field, therefore improves
The spatial resolution of high spectrum image is of great significance.To solve the problems, such as this, there is researcher to propose to sense using high-resolution
The simple solution of device, but it is a further reduction the density for reaching sensor photon, it is infeasible in many application scenarios.By
It is to promote effective hand of image spatial resolution using the image super-resolution technology based on soft method in the limitation of hardware device
Section.The property of super-resolution (Super Resolution, SR) problem is insufficient, it has been suggested that a variety of regularization methods solve this
Problem includes the method based on interpolation, the method based on multiple image, the single-frame images super-resolution side based on sample learning
Method.These methods obtain preferable reconstruction effect on gray level image or coloured image, but are directly applied to high spectrum image
As a result unsatisfactory.
High spectrum image SR problems, should not only improve spatial-domain information, but also need spectral preservation information.Classical base
In the high spectrum image SR methods of convex set projection, the information for merging multiple bands improves spatial resolution, and will observe scene
Frequency spectrum be redeveloped into the combination of a small amount of frequency spectrum basic function.In order to estimate kinematic parameter, there is scholar to propose based on the more of maximum a posteriori
Frame image SR methods reduce calculated load using principal component analysis and rebuild high-definition picture.However, based on multiple image
SR methods need accurate registration process.In order to overcome this difficulty, it is based on compressed sensing (Compressive in recent years
Sensing, CS) method obtain enough attention, can learn from high-resolution training image, the high-resolution rebuild
The high frequency detail of rate image.Different for each band noise intensity of high spectrum image, there is noise dirt in spatial domain and spectral domain
The problem of dye, has scholar to propose the sparse representation method based on grouping Three-dimensional DCT dictionary；In order to enhance bloom
The spatial resolution of spectrogram picture is had scholar to be merged high spectrum image with full-colour image using the method for matrix decomposition, there is proposition
Spectrum solution mixes and sparse coding method, is merged with RGB image for high spectrum image, and also scholar is mixed using coupling spectrum solution
The method of conjunction merges multispectral and high-spectrum remote sensing, improves the spatial resolution of high spectrum image.The above method is certain
The spatial resolution of image is improved in degree, but is introduced spectrum domain information without effective method and improved reconstruction effect
Fruit, there are spectral information problem of dtmf distortion DTMF.
Compared with Hyperspectral imager, although the overall quantization of low spectral resolution imaging system scene radiation is eliminated greatly
Partial spectral information, but can preferably retain the spatial structural form of scene image, it is obtained using the imaging system
High-definition picture helps to improve the spatial resolution of high spectrum image.
Invention content
Based on this, the present invention proposes a kind of high spectrum image spatial resolution enhancement method, includes following steps：
Step 1: extracting the reflectance spectrum of different scenes in image, obtained by compressed sensing dictionary learning algorithm strong dilute
Dredge property, weak coherence spectrum dictionary Φ；
Step 2: using the sparsity of signal, nonnegativity and space structure self-similarity in EO-1 hyperion scene, from extraction
Scene reflectivity spectrum resolution obtain the low spectrum picture of high spatial resolution, solved by synchronous orthogonal matching pursuit algorithm
To sparse coding matrix B；
Step 3: combining the spectrum dictionary Φ and the sparse coding matrix B obtains target image S.
Spectrum dictionary Φ in the step 1 is calculated by the following formula：
Wherein, XhFor high-resolution training set, h indicates high-resolution, all column vectors of training image blocks is connected into
Matrix Xh, Xh={ x1,...,xi,...xmn},x iIndicate that the column vector that pretreated image block is formed, mn indicate training set
Quantity, m, n distinguish representation space and tie up size, A ∈ Rd×mnIndicate all rarefaction representation coefficient αiMatrix, V0It is sparse constraint ginseng
Number, V1It is coherence's threshold value, W is observing matrix, and d indicates dictionary atom number.
Sparse coding matrix B in the step 2 is calculated by the following formula:
Wherein, ε indicates that error term, P indicate three-dimensional original input signal,It is P two dimensional forms, passes through connection
The two-dimensional matrix form that pixel in P is formed, T ∈ Rl× L is transformation matrix,It is to be obtained after dictionary φ is converted,βPiRepresenting matrix BpI-th row, M, N representation spaces tie up size, l indicate Spectral dimension, M > > m, N > > n, L >
> l.
When calculating sparse coding matrix B by formula (2), i interative computation, i=i+1 are carried out.Step is described as follows：
Step 2.1 calculates each dictionary atomThe accumulation correlation b of residual error approximate with current image blockj, i.e.,
Step 2.2,
Step 2.3,
Step 2.4,
Step 2.5, update residual error：
If step 2.6, | | Ri||F>σ||Ri-1||F, then algorithm stopping, otherwise continuing next iteration.
Wherein, further include initialization step before carrying out the ith iteration：
I=0；
Initial neutralizing：B0=0；
Initial residual error：
Collection is drawn in initialization：Row-supp { B }={ 1≤t≤d, βt≠ 0 }, representing matrix B
In non-zero row radix, βtIt is the t rows of matrix B.
Target image S=φ B are obtained by the above method.
The advantageous effects of the present invention：
1. preferably improving the acutance at edge and having restored more image detail informations, such as above target image
The marginal information of object of reference；
2. preferably rebuild the contour edge and detailed information of target image, improve target image edge sharpness and
Details around target image；
3. having better super-resolution rebuilding effect.
Description of the drawings
Fig. 1 is that " campus " image difference SR methods rebuild design sketch
Fig. 1 (a) is high-resolution reference picture
Fig. 1 (b) is the reconstruction image of bicubic interpolation method
Fig. 1 (c) is the reconstruction image of MF methods
Fig. 1 (d) is the reconstruction image of the method for the present invention
Fig. 2 is that " truck " scale model difference SR methods rebuild effect
Fig. 2 (a) is high-resolution reference picture
Fig. 2 (b) is the reconstruction image of bicubic interpolation method
Fig. 2 (c) is the reconstruction image of MF methods
Fig. 2 (d) is the reconstruction image of the method for the present invention
Fig. 3 is that the spectrum of different SR methods reconstruction images compares
Fig. 3 (a) is " campus " image
Fig. 3 (b) is " truck " scale model image
Relationship between Fig. 4 APSNR, ASSIM and SAM and image block size
APSNR, ASSIM and SAM index of Fig. 5 distinct methods reconstruction image in the case where different resolution promotes multiplying power
Fig. 6 is the high spectrum image spatial resolution enhancement method flow diagram of the present invention
Specific implementation mode
The application is described in further detail below in conjunction with the accompanying drawings, it is necessary to it is indicated herein to be, implement in detail below
Mode is served only for that the application is further detailed, and should not be understood as the limitation to the application protection domain, the field
Technical staff can make some nonessential modifications and adaptations according to above-mentioned application content to the application.
As shown in fig. 6, in the present invention, it is proposed that a kind of high spectrum image spatial resolution enhancement method, this method packet
Following steps are included：
Step 1: extracting the reflectance spectrum of different scenes in image, obtained by compressed sensing dictionary learning algorithm strong dilute
Dredge property, weak coherence spectrum dictionary Φ；
The target of high spectrum image super-resolution is from the low resolution high spectrum image Y obtainedhWith corresponding height
Restore to obtain high-resolution high spectrum image S, wherein Y in image in different resolution Yh∈Rm×n×L, Y ∈ RM×N×l, S ∈ RM×N×L,
M, m, N and n representation space tie up size, and L and l indicate Spectral dimension.Due to M > > m, N > > n, L > > l so that the problem is not
It is suitable fixed, consider YhIt is as follows respectively as the Linear Mapping of target image S with Y：
Yh∈ψh(S),Y∈ψ(S)
Wherein ψhIndicate RM×N×L→Rm×n×L, ψ expressions RM×N×L→RM×N×l。
By the sparsity prior information of high spectrum image it is found that usually only comprising a small number of different in high spectrum image scene
Material, and usually include only considerably less different spectrum compared with whole image, in each pixel.Therefore YhIn pixel yh
It can be indicated by the atom of linear combination spectrum dictionary Φ, wherein yh∈RL, can indicate y with following matrix formh：
yh≈φα
In above formula, φ ∈ RL×dIndicate the reflective vector of different materials in image scene, α ∈ RdIt is coefficient vector.Work as pixel
yhWhen the scene of expression also includes the pixel corresponding regions y of Y, wherein y ∈ Rl, then can be with approximate representation y：
y≈(Tφ)β (3)
In formula (3), T ∈ Rl×LIt is a transformation matrix, β ∈ RdIt is coefficient vector, wherein T is by Hyperspectral imager
Spectrum quantization matrix associated with high-resolution low spectrum imaging system, can be obtained using the contiguity between matrix：
y≈(TΦ)β≈Ts (4)
Wherein s ∈ RL, s indicates the pixel in high-resolution target image S.Formula (4) show if dictionary Φ it is known that if
The high-resolution high spectrum image S of sparse coding matrix Combined estimator appropriate can be passed through.
In the present invention, Φ is known as dictionary, T ∈ Rl×LFor transformation matrix,For the dictionary after transformation.Dictionary Φ's is every
One row are referred to as dictionary atom, and coefficient matrix (such as B) is referred to as sparse coding matrix.
The selection of dictionary plays very important effect in rarefaction representation field.Sparsity based on compressive sensing theory and
Incoherence, by by the dictionary learning method of compressed sensing obtain one strong sparsity, weak coherence spectrum dictionary.The word
Allusion quotation learning method learns dictionary from the LR images of acquisition itself, because in local or non local region, the LR of different-waveband
There are a large amount of similar image blocks in image；And correlative study the result shows that, made using pretreated low-resolution image
For training sample, the dictionary of study is than using the dictionary effect that high-definition picture is trained more preferable.
Dictionary learning method based on compressed sensing handles low resolution high spectrum image by linear interpolation method first
Yh, then train to obtain dictionary using pretreated image block.In addition, in order to generate high-resolution training set Xh, training is schemed
As all column vectors of block connect into matrix Xh, Xh={ x1,...,xi,...xmn, wherein xiIndicate pretreated image block
The column vector of formation, mn indicate the quantity of training image blocks.In order to meet the constraints of CS theories, dictionary learning task simultaneously
It indicates as follows：
In formula (1), A ∈ Rd×mnIndicate all rarefaction representation coefficient αiMatrix, V0It is sparse constraint parameter, V1It is relevant
Property threshold value, W is observing matrix, d indicate dictionary atom number, d can with value be 256~2048 between.Based on compressed sensing
Dictionary learning method makes full use of the strong sparsity of CS theories and weak coherence, makes consistent between study dictionary and perception matrix
Property becomes smaller；In addition, the dictionary of our calligraphy learning is not the dictionary atom containing fixed quantity, can be reduced according to relevant threshold
The dimension of dictionary accelerates the speed of sparse decomposition.By the above-mentioned dictionary learning method based on compressed sensing, solves formula (1) and obtain
To dictionary Φ.
Step 2: using the sparsity of signal, nonnegativity and space structure self-similarity in EO-1 hyperion scene, from extraction
Scene reflectivity spectrum resolution obtain the low spectrum picture of high spatial resolution, solved by synchronous orthogonal matching pursuit algorithm
To sparse coding matrix B；
Target image S in order to obtain needs solution to obtain suitable sparse coding matrix B according to formula (4), and the present invention passes through
Learn obtained dictionary Φ and high-resolution image Y and two important prior informations to solve sparse coding matrix B.One
It is in high-definition picture, neighbouring pixel may indicate identical material in the scene, can be by one group of identical dictionary
Atom indicates well；Second is that the element in sparse coding matrix B is non-negative, the ratio system in spectral signal source in scene is indicated
Number.For this purpose, handling image Y by small disjoint spatial image block, sparse coding matrix B is calculated.WithTable
Show each image block, and sparse Approximation Problem is synchronized by what is constrained below solving, obtains corresponding sparse coding matrix
In formula (2), ε indicates that error term, P indicate three-dimensional original input signal,It is the two dimensional form of P, passes through
Connect the two-dimensional matrix form that the pixel in P is formed, T ∈ Rl×LIt is transformation matrix,βPiRepresenting matrix BpI-th row,
M, N representation space tie up size, and l indicates Spectral dimension, M > > m, N > > n, L > > l.When the signal of input is that single row are former
The period of the day from 11 p.m. to 1 a.m, the orthogonal matching pursuit algorithm that standard may be used solve, but in formula (2)Contain MPNPA letter
Number row atom, and there is nonnegativity restriction, therefore formula (2) is that constraint synchronizes sparse Approximation Problem.In order to solve formula
(2), which is expanded on the basis of orthogonal matching pursuit algorithm, and each iteration can simultaneous selection dictionaryIn it is multiple
Row atom, and solution space is constrained to nonnegative matrix.
The algorithm is by selecting index setCorrespondingDictionary atomSeek input signalApproximation
Value, and it is eachContribute to the approximation of whole image block.In ith iteration, each dictionary atom and image are calculated first
The accumulation correlation of the current approximate residual error of block, image block itself are used as initial residual error；Then identification has cumulative maximum correlation
D dictionary atom, these atoms are added to setIn the subspace of index, wherein initial index collectionFor sky；Then
Above-mentioned subspace is used for the non-negative least squqre approximation of image block, and updates residual error Ri.If updated residual error is more than
Last iteration residual error, then algorithm stopping.
This algorithm is using the image block in the approximate iteration every time of nonnegative least, rather than the least square of standard is close
Seemingly, nonnegativity restrictions can be applied to sparse coding matrix.Sparse coding matrix B is obtained by solving formula (2), together with dictionary Φ
Obtain target image S.
Assuming that the high spectrum image Y of low resolutionhWith high-resolution high spectrum image Xh, h=1,2 ... and L }, wherein L
Indicate that the quantity of band in image, high spectrum image resolution enhancement algorithm of the invention are described as follows.The oversubscription of the present invention
Resolution algorithm is as follows：
Input：The Φ that 1 dictionary learning of algorithm obtains, dictionary is obtained by transformation
1：Initialization
1.1：I=0；
1.2：Initial neutralizing：B0=0；
1.3：Initial residual error：
1.4：Collection is drawn in initialization：Row-supp { B }={ 1≤t≤d, βt≠ 0 } it, indicates
The radix of non-zero row, β in matrix BtIt is the t rows of matrix B.
2：Formula (2) is solved by synchronous orthogonal matching pursuit algorithm：Iteration item i=i+1
2.1：Calculate each dictionary atomThe accumulation correlation b of residual error approximate with current image blockj, i.e.,
2.2、
2.3、
2.4、
2.5, residual error is updated：
If 2.6, | | Ri||F>σ||Ri-1||FThen stop, otherwise continuing next iteration.
Output：S=Φ B
Step 3: combining the spectrum dictionary Φ and the sparse coding matrix B obtains target image S, i.e., according to above-mentioned
Method obtains S=Φ B.
In order to examine effectiveness of the invention, using the picture number of the image and actual acquisition concentrated including common data
According to carrying out experimental verification, and by the reconstructed results of this method and classical bicubic interpolation method, matrix decomposition (Matrix
Factorization, MF) method is compared.In order to assess the reconstruction effect of the present invention, experimental result is from subjective and objective two
A aspect is compared analysis.It is subjective mainly to compare going for reconstructed results from visual effect for single-range reconstruction image
The reconstruction situation of the information such as fuzzy and grain details, objectively from Y-PSNR (Peak Signal-to-Noise
Ratio, PSNR), structural similarity (Structural Similarity, SSIM) carry out image reconstruction quality compare.In addition,
The present invention is flat using wave band average peak signal to noise ratio (Average Peak Signal-to-Noise Ratio, APSNR), wave band
Equal structural similarity (Average Structural Similarity, ASSIM) and spectral modeling map (Spectral
Angel Mapper, SAM) quantitative analysis is carried out to the quality of reconstruction image.
It is that point aperture polarizes hyperspectral imager, experiment operation ring with up-to-date style that the present invention, which acquires the equipment used in experimental image,
Border is：Lenovo ideapad700,Intel Core i5-6300HQ,[email protected],8GB RAM,MATLAB
R2014a.Present invention experiment uses " campus " high spectrum image and reality of common data sets (AVIRIS Dataset)
Scale model " truck " high spectrum image of acquisition.Wherein, the spectral band range of " campus " image is that 400nm is arrived
1200nm, spectral resolution 20nm；" truck " experimental image spectral band range acquired is 400nm to 700nm, spectrum
Resolution ratio is 10nm.High-resolution reference picture is carried out to Gaussian Blur and down-sampled processing respectively in experiment, obtains waiting locating
The low spatial resolution image of reason.The size of Gaussian Blur core is 8 × 8, and the down-sampled factor of row and column is 3.Dictionary learning
Training stage, it is 4 × 4, V to take N=10000 training image blocks, image block size1Value is 1.6, initialization dictionary Φ0's
Atomicity is d=100；In synchronous orthogonal matching pursuit algorithm, each iteration chooses D=25 dictionary atom, residual error ginseng simultaneously
Number ε=0.9.Bicubic interpolation method, MF methods and the method for the present invention is respectively adopted, super-resolution is carried out to low-resolution image
Rate is rebuild, and is compared to the reconstruction effect of distinct methods.
Fig. 1 compares " campus " image in the reconstruction effect of different SR algorithms, and as shown in the figure is 600nm wave bands
Reconstruction image.Fig. 1 (a) is high-resolution reference picture, and as shown in Fig. 1 (b), the reconstruction image of bicubic interpolation method integrally compares
It is relatively fuzzy；Fig. 1 (c) is the reconstruction image of MF methods, and compared with bicubic interpolation, MF methods have restored many details, improve
The edge sharpness of image and whole clarity；Fig. 1 (d) is the reconstruction image of the method for the present invention, compared with MF methods, the present invention
Method preferably improves the acutance at edge and has restored more image detail informations, for example, image above road edge
Information.
Fig. 2 compares the effect that acquired " truck " scale model image is rebuild in different SR methods, show 540nm
The reconstruction image of wave band.Fig. 2 (a) is high-resolution reference picture, as shown in Fig. 2 (b), the reconstruction image of bicubic interpolation method
Compare fuzzy, many detailed information can not be differentiated；Fig. 2 (c) is the reconstruction image of MF methods, compared with bicubic interpolation method,
MF methods have restored many image details, such as the clarity enhancing of " truck " ambient conditions；Fig. 2 (d) is the method for the present invention
Reconstruction image, compared with MF methods, the method for the present invention has preferably rebuild the contour edge and detailed information of " truck ", improves
The details of image edge acuity and " truck " image peripheral.
For further objective evaluation SR method for reconstructing, reconstruction the effect utilization PSNR and SSIM of different SR methods are compared
Analysis.It is rebuild under multiplying power at 3 times, evaluation result is as shown in table 1.The experimental results showed that the objective evaluation index of the method for the present invention
Be superior to other two methods, the average value of PSNR and SSIM than bicubic interpolation method be respectively increased 2.68dB and
0.1373,1.35dB and 0.0507 has been respectively increased than MF method.
The PSNR (dB) and SSIM evaluation indexes of 1. distinct methods of table compare
In addition, the present invention use wave band average peak signal to noise ratio (APSNR), wave band average structure similarity (ASSIM) with
And spectrum angle map (SAM) carries out quantitative analysis evaluation to the reconstruction quality of image.If Y indicates EO-1 hyperion reference picture, can
To be indicated with three-dimensional matrice, Y=(yi,j,q)∈RM×N×Q, M and N are respectively the sizes of image Y both horizontally and vertically, and Q is wave
The number of section, the image Y of the q wave bands formation of Yq, wherein q=1,2 ... Q }；Indicate the high spectrum image rebuild,ForQ band images formed vector.APSNR, ASSIM and SAM are defined respectively as：
T in formula (7)qFor YqPeak value, m isMean square error；tqFor the SSIM values of q wave bands；Y (i, j) and
Pixel respectively after image original pixels and reconstruction.
The value of APSNR and ASSIM is bigger, shows that the quality of EO-1 hyperion reconstruction image is higher.SAM is by original image and again
The absolute angle reflection spectrum distortion between two spectral vectors that each pixel of image is constituted is built, and to whole image
It carries out average computation and measures SAM.For ideal reconstruction image, the value of SAM should be 0.Table 2 and table 3 have been respectively compared " big
APSNR, ASSIM and SAM value of school garden " image and " truck " scale model image reconstruction image under different SR algorithms, slightly
The numerical value of body label indicates there is optimal effectiveness under corresponding evaluation index.The experimental results showed that the method for the present invention has preferably
Super-resolution rebuilding effect.
The APSNR (dB), ASSIM, SAM evaluation index comparing result of table 2 " campus " algorithms of different
The APSNR (dB), ASSIM, SAM evaluation index comparing result of table 3 " truck " image algorithms of different
Fig. 3 is to be compared using the spectrum of different SR methods reconstruction images, wherein figure (a) is specific in " campus " image
The curve of spectrum of pixel, totally 41 spectral bands；Figure (b) is the curve of spectrum of certain picture elements in " truck " scale model image,
Totally 31 spectral bands.Compare from the curve of spectrum in Fig. 3 as it can be seen that between the image and reference picture that the method for the present invention is rebuild
SPECTRAL DIVERSITY is minimum.The result shows that the method for the present invention has better spectrum fidelity.
The relationship that effect is rebuild to refine analysis parameter with the method for the present invention, is carried out using " truck " scale model image
It tests, the relationship between the parameters such as quantitative analysis reconstructed image quality and image block size, increase resolution multiplying power.Fig. 4 (a), 4
(b), 4 (c) is respectively the relationship between APSNR, ASSIM, SAM and image block size, as can be seen from the figure selects 4 × 4
It is more preferable that image block rebuilds effect.
Fig. 5 (a), (b), (c) are respectively that APSNR, ASSIM and SAM index of three kinds of algorithm reconstruction images and resolution ratio carry
Rise the relationship between multiplying power.As can be seen from the figure, with the increase of increase resolution multiplying power, the matter of three kinds of algorithm reconstruction images
Amount is gradually reduced.Compared with other two kinds of algorithms, the method for the present invention reconstructed image quality has when different resolution promotes multiplying power
There is better reconstruction quality, wherein APSNR and ASSIM indexs are than other two methods biggers, SAM index smallers.
Several embodiments of the invention above described embodiment only expresses, the description thereof is more specific and detailed, but simultaneously
Cannot the limitation to the scope of the claims of the present invention therefore be interpreted as.It should be pointed out that for those of ordinary skill in the art
For, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to the guarantor of the present invention
Protect range.Therefore, the protection domain of patent of the present invention should be determined by the appended claims.
Claims (5)
1. a kind of high spectrum image spatial resolution enhancement method, which is characterized in that this method step is described as follows：
Step 1: extraction image in different scenes reflectance spectrum, by compressed sensing dictionary learning algorithm obtain strong sparsity,
The spectrum dictionary Φ of weak coherence；
Step 2: using the sparsity of signal, nonnegativity and space structure self-similarity in EO-1 hyperion scene, from the field of extraction
Scape reflectance spectrum parses to obtain the low spectrum picture of high spatial resolution, solves to obtain by synchronous orthogonal matching pursuit algorithm dilute
Dredge encoder matrix B；
Step 3: combined spectral dictionary Φ and sparse coding matrix B obtain target image S.
2. high spectrum image spatial resolution enhancement method according to claim 1, which is characterized in that in the step 1
Spectrum dictionary Φ be calculated by the following formula：
Wherein, XhFor high-resolution training set, h indicates high-resolution, all column vectors of training image blocks is connected into matrix
Xh, Xh={ x1,...,xi,...xmn},xiIndicate that the column vector that pretreated image block is formed, mn indicate the number of training set
Amount, m, n distinguish representation space and tie up size, A ∈ Rd×mnIndicate all rarefaction representation coefficient αiMatrix, V0It is sparse constraint parameter,
V1It is coherence's threshold value, W is observing matrix, and d indicates dictionary atom number.
3. high spectrum image spatial resolution enhancement method according to claim 1, which is characterized in that in the step 2
Sparse coding matrix B be calculated by the following formula：
Wherein, ε indicates that error term, P indicate three-dimensional original input signal,It is P two dimensional forms, by connecting in P
The two-dimensional matrix form that pixel is formed, T ∈ Rl×LIt is transformation matrix,It is to be obtained after dictionary φ is converted, Table
Show matrix BpI-th row, M, N representation spaces tie up size, l indicate Spectral dimension, M > > m, N > > n, L > > l.
4. high spectrum image spatial resolution enhancement method according to claim 3, which is characterized in that described to pass through formula
(2) when calculating sparse coding matrix B, i interative computation, i=i+1 are carried out, step is described as follows：
Step 2.1 calculates each dictionary atomThe accumulation correlation b of residual error approximate with current image blockj, i.e.,
Step 2.2,
Step 2.3,
Step 2.4,
Step 2.5, update residual error：
If step 2.6, | | Ri||F>σ||Ri-1||F, then algorithm stopping, otherwise continuing next iteration.
Wherein, further include initialization step before carrying out the ith iteration：
I=0；
Initial neutralizing：B0=0；
Initial residual error：
Collection is drawn in initialization：Row-supp { B }={ 1≤t≤d, βt≠ 0 }, non-in representing matrix B
The radix of 0 row, βtIt is the t rows of matrix B.
5. high spectrum image spatial resolution enhancement method according to claim 1, which is characterized in that described by joint light
It is S=φ B that spectrum dictionary and sparse coding matrix, which obtain target image S,.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810274921.5A CN108765280A (en) | 2018-03-30 | 2018-03-30 | A kind of high spectrum image spatial resolution enhancement method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810274921.5A CN108765280A (en) | 2018-03-30 | 2018-03-30 | A kind of high spectrum image spatial resolution enhancement method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108765280A true CN108765280A (en) | 2018-11-06 |
Family
ID=63980890
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810274921.5A Pending CN108765280A (en) | 2018-03-30 | 2018-03-30 | A kind of high spectrum image spatial resolution enhancement method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108765280A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109816620A (en) * | 2019-01-31 | 2019-05-28 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN109887050A (en) * | 2019-02-26 | 2019-06-14 | 哈尔滨工程大学 | A kind of code aperture spectrum imaging method based on self-adapting dictionary study |
| CN110852950A (en) * | 2019-11-08 | 2020-02-28 | 中国科学院微小卫星创新研究院 | Hyperspectral image super-resolution reconstruction method based on sparse representation and image fusion |
| CN112488221A (en) * | 2020-12-07 | 2021-03-12 | 电子科技大学 | Road pavement abnormity detection method based on dynamic refreshing positive sample image library |
| CN112766102A (en) * | 2021-01-07 | 2021-05-07 | 武汉大学 | Unsupervised hyperspectral video target tracking method based on space-spectrum feature fusion |
| CN112927149A (en) * | 2021-02-18 | 2021-06-08 | 北京印刷学院 | Spatial resolution enhancement method and device for hyperspectral image and electronic equipment |
| CN112967350A (en) * | 2021-03-08 | 2021-06-15 | 哈尔滨工业大学 | Hyperspectral remote sensing image eigen decomposition method and system based on sparse image coding |
| CN113670440A (en) * | 2021-07-21 | 2021-11-19 | 杭州电子科技大学 | Compressed spectrum imaging method based on adaptive dictionary |
| CN114708281A (en) * | 2022-04-15 | 2022-07-05 | 桂林电子科技大学 | Image compressed sensing reconstruction method based on self-adaptive non-local feature fusion network |
| CN117571128A (en) * | 2024-01-16 | 2024-02-20 | 长春理工大学 | High-resolution polarized spectrum image imaging method and system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103400389A (en) * | 2013-08-12 | 2013-11-20 | 电子科技大学 | High resolution remote sensing image segmentation method |
| CN105469360A (en) * | 2015-12-25 | 2016-04-06 | 西北工业大学 | Non local joint sparse representation based hyperspectral image super-resolution reconstruction method |
-
2018
- 2018-03-30 CN CN201810274921.5A patent/CN108765280A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103400389A (en) * | 2013-08-12 | 2013-11-20 | 电子科技大学 | High resolution remote sensing image segmentation method |
| CN105469360A (en) * | 2015-12-25 | 2016-04-06 | 西北工业大学 | Non local joint sparse representation based hyperspectral image super-resolution reconstruction method |
Non-Patent Citations (3)
| Title |
|---|
| ZAHRA HASHEMI NEZHAD 等: "Fusion of Hyperspectral and Multispectral Images Using Spectral Unmixing and Sparse Coding", 《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》 * |
| 李进明: "基于稀疏表示的图像超分辨率重建方法研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
| 许蒙恩 等: "空间光谱联合稀疏表示的高光谱图像超分辨率方法", 《激光与光电子学进展》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109816620B (en) * | 2019-01-31 | 2021-01-05 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN109816620A (en) * | 2019-01-31 | 2019-05-28 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN109887050A (en) * | 2019-02-26 | 2019-06-14 | 哈尔滨工程大学 | A kind of code aperture spectrum imaging method based on self-adapting dictionary study |
| CN110852950B (en) * | 2019-11-08 | 2023-04-07 | 中国科学院微小卫星创新研究院 | Hyperspectral image super-resolution reconstruction method based on sparse representation and image fusion |
| CN110852950A (en) * | 2019-11-08 | 2020-02-28 | 中国科学院微小卫星创新研究院 | Hyperspectral image super-resolution reconstruction method based on sparse representation and image fusion |
| CN112488221A (en) * | 2020-12-07 | 2021-03-12 | 电子科技大学 | Road pavement abnormity detection method based on dynamic refreshing positive sample image library |
| CN112766102A (en) * | 2021-01-07 | 2021-05-07 | 武汉大学 | Unsupervised hyperspectral video target tracking method based on space-spectrum feature fusion |
| CN112766102B (en) * | 2021-01-07 | 2024-04-26 | 武汉大学 | Unsupervised hyperspectral video target tracking method based on spatial spectrum feature fusion |
| CN112927149A (en) * | 2021-02-18 | 2021-06-08 | 北京印刷学院 | Spatial resolution enhancement method and device for hyperspectral image and electronic equipment |
| CN112967350A (en) * | 2021-03-08 | 2021-06-15 | 哈尔滨工业大学 | Hyperspectral remote sensing image eigen decomposition method and system based on sparse image coding |
| CN113670440A (en) * | 2021-07-21 | 2021-11-19 | 杭州电子科技大学 | Compressed spectrum imaging method based on adaptive dictionary |
| CN113670440B (en) * | 2021-07-21 | 2023-11-10 | 杭州电子科技大学 | Compression spectrum imaging method based on self-adaptive dictionary |
| CN114708281B (en) * | 2022-04-15 | 2024-03-22 | 桂林电子科技大学 | Image compressed sensing reconstruction method based on self-adaptive non-local feature fusion network |
| CN114708281A (en) * | 2022-04-15 | 2022-07-05 | 桂林电子科技大学 | Image compressed sensing reconstruction method based on self-adaptive non-local feature fusion network |
| CN117571128A (en) * | 2024-01-16 | 2024-02-20 | 长春理工大学 | High-resolution polarized spectrum image imaging method and system |
| CN117571128B (en) * | 2024-01-16 | 2024-03-29 | 长春理工大学 | High-resolution polarized spectrum image imaging method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108765280A (en) | A kind of high spectrum image spatial resolution enhancement method | |
| Shao et al. | Remote sensing image fusion with deep convolutional neural network | |
| CN108573276B (en) | Change detection method based on high-resolution remote sensing image | |
| Ghahremani et al. | A compressed-sensing-based pan-sharpening method for spectral distortion reduction | |
| Chen et al. | SIRF: Simultaneous satellite image registration and fusion in a unified framework | |
| Xu et al. | Image inpainting by patch propagation using patch sparsity | |
| CN104867124B (en) | Multispectral and panchromatic image fusion method based on the sparse Non-negative Matrix Factorization of antithesis | |
| CN104008538B (en) | Based on single image super-resolution method | |
| CN114119444B (en) | Multi-source remote sensing image fusion method based on deep neural network | |
| CN104574336B (en) | Super-resolution image reconstruction system based on adaptive sub- mould dictionary selection | |
| CN110232653A (en) | The quick light-duty intensive residual error network of super-resolution rebuilding | |
| CN109636769A (en) | EO-1 hyperion and Multispectral Image Fusion Methods based on the intensive residual error network of two-way | |
| Yilmaz et al. | A theoretical and practical survey of image fusion methods for multispectral pansharpening | |
| CN110136060B (en) | Image super-resolution reconstruction method based on shallow dense connection network | |
| CN111340696B (en) | Convolutional neural network image super-resolution reconstruction method fused with bionic visual mechanism | |
| CN104021523B (en) | A kind of method of the image super-resolution amplification based on marginal classification | |
| CN108288256A (en) | Multispectral mosaic image restoration method | |
| Lin et al. | Deep unsupervised learning for image super-resolution with generative adversarial network | |
| Piella et al. | Multiresolution image fusion guided by a multimodal segmentation | |
| Zou et al. | Bayesian dictionary learning for hyperspectral image super resolution in mixed Poisson–Gaussian noise | |
| CN115511767A (en) | Self-supervised learning multi-modal image fusion method and application thereof | |
| Zhong et al. | Attention_FPNet: Two-branch remote sensing image pansharpening network based on attention feature fusion | |
| CN116934592A (en) | Image stitching method, system, equipment and medium based on deep learning | |
| Ye et al. | An unsupervised SAR and optical image fusion network based on structure-texture decomposition | |
| Liu et al. | Image super-resolution via hierarchical and collaborative sparse representation |
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 | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20181219 Address after: 230000 No. 555 Wangjiangxi Road, Hefei High-tech Zone, Anhui Province Applicant after: ANHUI XINHUA University Address before: 230000 Room 703, Building 71, 451 Huangshan Road, Hefei City, Anhui Province Applicant before: Xu Guoming |
|
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181106 |