CN104732504A - Image fusion method based on compressed sensing and WBCT - Google Patents
Image fusion method based on compressed sensing and WBCT Download PDFInfo
- Publication number
- CN104732504A CN104732504A CN201510037279.5A CN201510037279A CN104732504A CN 104732504 A CN104732504 A CN 104732504A CN 201510037279 A CN201510037279 A CN 201510037279A CN 104732504 A CN104732504 A CN 104732504A
- Authority
- CN
- China
- Prior art keywords
- wbct
- coefficient
- source images
- matrix
- image
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to the field of digital image processing, and provides a method capable of effective showing and processing images and other higher spatial data based on compressed sensing and WBCT. The method comprises the following steps that firstly, small-wave based Contourlet transformation is carried out on a to-be-fused image; secondly, a measurement matrix is formed, and measuring is carried out on LH, HL and HH high frequency sub-bands; thirdly, pixel point leveled evaluating is carried out on low frequency components of two source images by the rule adopting the Weighted Average (WA); fourthly, fusion is carried out on measured high frequency components of the two source images by using the principal component analysis; fifthly, reconstitution calculation is carried out on the fused high frequency components by using StOMP; sixthly, inverse transformation is carried out by using WBCT. The method is mainly applied to digital image processing.
Description
Technical field
The present invention relates to digital image processing field, Image Fusion field.Specifically, the image interfusion method based on compressed sensing and WBCT conversion is related to
Background technology
Image co-registration obtains limitation and the otherness defect of view data by compensate for single piece of information source to the fusion of different images information, complementation between comprehensive utilization different images and redundant information, obtain scene iamge description accurately more comprehensively, greatly enhance picture quality.Image co-registration based on wavelet transformation is current widely used a kind of image interfusion method.Wavelet transformation has good multiple dimensioned and time-frequency local characteristics.But traditional wavelet is used for only having limited direction when representing image, namely level, vertical and diagonal line three directions, can not represent the directional information in image more accurately; And small echo can only represent the some singularity of signal, well can not represent the anisotropy of image.
Image co-registration is applied in compressive sensing theory framework.If do not merged after Image Reconstruction, but directly the measured value few to each image merges by certain fusion method, again the data after fusion are carried out transmitting, storing, only need use the data of few fusion during reconstructed image, so just can obtain high-quality image.Greatly can improve the speed of calculating like this, the cost save transmission, storing, reconstruct.
Summary of the invention
For overcoming the deficiencies in the prior art, for the limitation of wavelet transformation, for the High dimensional space data such as effective expression and process image can be realized.For this reason, the technical scheme that the present invention takes is, based on the image interfusion method that compressed sensing and WBCT convert, comprises the steps:
The first step: the Contourlet (the direction multiscale analysis method of non-self-adapting) conversion (WBCT) based on small echo is carried out to image to be fused, obtain low frequency respectively and merge component and high frequency fusion component, the pyramid decomposition number of plies wherein used is three layers, the decomposition direction of every layer is 2k, and wherein k represents the number of plies of decomposition;
Second step: observing matrix M × N is used to observe to the original signal of N dimension the observation vector Y obtaining M dimension, M<<N, then optimization method high probability reconstruct from observed reading Y is utilized, construct a calculation matrix, respectively to LH, HL, HH high-frequency sub-band is measured; Obtain the measurement coefficient value matrix of these three subbands. and it is constant to retain low frequency sub-band coefficient LL, chooses Gauss with matrix as calculation matrix, carries out measurement value.LL: the low low-value that wavelet decomposition obtains; LH: the low high-value that wavelet decomposition obtains; HL: the height place value that wavelet decomposition obtains; HH: the high high-value that wavelet decomposition obtains;
3rd step: use the assessment adopting weighted mean (Weighted Average, the WA) rule based on window to carry out pixel rank to the low frequency component of two width source images, chooses interior each of matrix and puts corresponding numerical value, form new matrix;
4th step: use principal component analysis (PCA) to merge to the high fdrequency component after the measurement of two width source images, obtains the fusion matrix after measuring;
5th step: use two stage cultivation tracing algorithm (StOMP) to be reconstructed calculating to the high fdrequency component after merging, obtain the high fdrequency component being suitable for WBCT inverse transformation;
6th step: use WBCT inverse transformation, the high fdrequency component of all directions and the fusion results of low frequency component are combined, obtains fused image.
Employing based on weighted mean (Weighted Average, WA) the regular concrete steps of window is:
The first step: in source images, calculates energy centered by (m, n) point in surrounding window region or the variance tolerance S as this detailed information intensity
a(m, n), S
b(m, n), S
a(m, n) is the tolerance of A source images point detailed information intensity, S
b(m, n) is the tolerance of B source images point detailed information intensity, and m, n are point coordinate;
Second step, calculates c
a(m, n) and c
alocal, normalized cross-correlation coefficient M between (m, n)
aB, c
a(m, n) is A source images coefficient, c
b(m, n) is B source images coefficient;
3rd step, according to cross-correlation coefficient size, take different amalgamation modes:
Work as M
aBduring≤a, a is threshold value, a=0.85, illustrates that between source images coefficient, correlativity is lower, chooses the large coefficient of local variance reasonable for merging rear coefficients comparison, namely
Wherein c
f(m, n) is fused image coefficient,
Work as M
aBduring >a, illustrate that between coefficient, correlativity is larger, adopt average weighted method more reasonable,
c
F(m，n)＝w(m，n)c
A(m，n)+[E(m，n)-w(m，n)]c
B(m，n)
Wherein E (m, n) is unit matrix, and weight coefficient w (m, n) is determined by following formula:
Compressed sensing restructing algorithm is further refined as: what select during each match tracing is multiple matched atoms instead of single atom, decrease matching times, a matched filter can be formed under particular state, identify the coordinate that all amplitudes are greater than a special selected threshold value, the coordinate selected with these does least square, then deduct least square fitting, obtain a new surplus.
Compared with the prior art, technical characterstic of the present invention and effect:
Compressive sensing theory carries out suitable compression to data while image co-registration, decreases sampled data, saves storage space, contains again enough quantity of information simultaneously, carries out Precise fusion by suitable reconstruction algorithm to specific image or signal
Accompanying drawing explanation
Fig. 1 algorithm block diagram.
Fig. 2 WBCT tri-layers of decomposition chart.
Embodiment
For the limitation of wavelet transformation, represent in order to more effective and process the High dimensional space datas such as image, the present invention adopts the Contourlet (the direction multiscale analysis method of non-self-adapting) conversion (WBCT) based on small echo.The implementation procedure of WBCT also uses two-stage and decomposes, first, WBCT adopts wavelet transformation to realize multi-resolution decomposition, efficiently avoid the data redundancy that Laplace filter is introduced, secondly, the high-frequency sub-band travel direction using directional filter banks wavelet decomposition to be obtained decomposes.In the wavelet decomposition stage, WBCT have employed separable filter, and at DFB, (in anisotropic filter (the Directional Filter Banks) stage, have employed the tree-shaped directional filter banks of inseparable iteration be made up of fan-filter.The two-layer conversion of WBCT is all break-even conversion, and therefore WBCT is non-redundancy Transform.
Core of the present invention and object are the effective integrations based on carrying out two width images under compressed sensing framework, and then obtain the fused images of the better image detail of larger information entropy.Image fusion system based on compressed sensing and WBCT conversion will reduce hardware and the computational burden of emerging system greatly, and then accelerate the travelling speed of whole system.
The concrete steps merged at compressed sensing framework hypograph are as follows:
The first step: the contourlet transformation (WBCT) based on small echo is carried out to image to be fused, obtain low frequency respectively and merge component and high frequency fusion component, the pyramid decomposition number of plies wherein used is three layers, and the decomposition direction of every layer is 2k, and wherein k represents the number of plies of decomposition.；
Second step: observing matrix (also claiming calculation matrix) M × N (M<<N) is used to observe to the original signal of N dimension the observation vector Y obtaining M dimension, then can utilize optimization method high probability reconstruct from observed reading Y.Construct a calculation matrix, respectively LH, HL, HH high-frequency sub-band is measured, obtain the measurement coefficient value matrix of these three subbands. and it is constant to retain low frequency sub-band coefficient LL.Choose Gauss with matrix as calculation matrix, carry out measuring value (LL: the low low-value that wavelet decomposition obtains; LH: the low high-value that wavelet decomposition obtains; HL: the height place value that wavelet decomposition obtains; HH: the high high-value that wavelet decomposition obtains);
3rd step: use the assessment adopting weighted mean (Weighted Average, the WA) rule based on window to carry out pixel rank to the low frequency component of two width source images, chooses interior each of matrix and puts corresponding numerical value, form new matrix;
4th step: use principal component analysis (PCA) to merge to the high fdrequency component after the measurement of two width source images, obtains the fusion matrix after measuring;
5th step: use two stage cultivation tracing algorithm (StOMP) to be reconstructed calculating to the high fdrequency component after merging, obtain the high fdrequency component being suitable for WBCT inverse transformation;
6th step: use WBCT inverse transformation, the high fdrequency component of all directions and the fusion results of low frequency component are combined, obtains fused image.
Technology path:
1, image sparse
Image sparse is the basis that image carries out compressed sensing.The present invention mainly adopts the contourlet transformation (WBCT) based on small echo.See Fig. 2.
The specific implementation process of WBCT is: for the wavelet decomposition of any number of plies, using the high-frequency sub-band HL containing detailed information, LH and HH as the object of Directional Decomposition, applies the identical direction transformation of decomposed class to the high-frequency sub-band on same yardstick.In order to meet anisotropy rule, the maximum direction transformation of decomposed class can be done at the top of wavelet decomposition, the decomposed class of the direction transformation that then secondary high level reduced by half.Fig. 2 is the decomposing schematic representation of WBCT.
2, the choosing of calculation matrix
Select gaussian random matrix to be observing matrix, will meet RIP (limited equidistant character) characteristic by high probability very much, and make us can go out source images by Accurate Reconstruction completely.
3, the design of blending algorithm
Carry out above-mentioned sparse computing and compressed sensing measurement to image, we obtain low frequency component and high fdrequency component two parts of source images.Low frequency component has concentrated most of energy of source images, reflects the approximate of source images and average characteristics.And high fdrequency component is made up of three part components, be horizontal component, vertical component, diagonal components respectively.High fdrequency component mainly have expressed the information such as image border, brightness, region contour.Because the difference of information type contained by two kinds of components, so use the convergence strategy of different rows.
The design of 3.1 low frequency part blending algorithms
Adopt weighted mean (Weighted Average, the WA) rule based on window
The first step: in source images, calculates energy centered by (m, n) point in surrounding window region or the variance tolerance S as this detailed information intensity
a(m, n), S
b(m, n), S
a(m, n) is the tolerance of A source images point detailed information intensity, S
b(m, n) is the tolerance of B source images point detailed information intensity, and m, n are point coordinate;
Second step, calculates c
a(m, n) and c
alocal, normalized cross-correlation coefficient M between (m, n)
aB, c
a(m, n) is A source images coefficient, c
b(m, n) is B source images coefficient;
3rd step, according to cross-correlation coefficient size, take different amalgamation modes:
Work as M
aBduring≤a, a is threshold value, a=0.85, illustrates that between source images coefficient, correlativity is lower, chooses the large coefficient of local variance reasonable for merging rear coefficients comparison, namely
Wherein c
f(m, n) is fused image coefficient,
Work as M
aBduring >a, illustrate that between coefficient, correlativity is larger, adopt average weighted method more reasonable,
c
F(m，n)＝w(m，n)c
A(m，n)+[E(m，n)-w(m，n)]c
B(m，n)
Wherein E (m, n) is unit matrix, and weight coefficient w (m, n) is determined by following formula:
The design of 3.2 HFS blending algorithms
Three high fdrequency components can be obtained after image entered wavelet transformation.High fdrequency component mainly contains the information of pixel gray-value variation in image, have expressed the information such as image border, brightness, region contour, needs when wave band is more, the better spectral characteristic that must keep source images.Because principal component analysis (PCA) better can must reach requirement, thus adopt principal component analysis (PCA) as the fusion method of high fdrequency component.The method is the multidimensional linear transformation on the basis of statistical nature, by dimensionality reduction technology, makes the method that multiple component abbreviation to a few fundamental component goes.Its fundamental purpose is the dimension reducing data, and keeps feature maximum to variance contribution in data simultaneously.
4, compressed sensing restructing algorithm
The present invention adopts step to move orthogonal matching pursuit algorithm (Stagewise Orthogonal Matching Pursuit, StOMP), OMP algorithm is carried out simplification to a certain degree, reduces Its Sparse Decomposition precision, improve computing velocity.What select when the essence of this algorithm is each match tracing is multiple matched atoms instead of single atom, decreases matching times.Can form a matched filter under particular state, identify the coordinate that all amplitudes are greater than a special selected threshold value, the coordinate selected with these does least square, then deducts least square fitting, obtains a new surplus.
Claims (3)
1., based on the image interfusion method that compressed sensing and WBCT convert, it is characterized in that, comprise the steps:
The first step: the Contourlet (the direction multiscale analysis method of non-self-adapting) conversion (WBCT) based on small echo is carried out to image to be fused, obtain low frequency respectively and merge component and high frequency fusion component, the pyramid decomposition number of plies wherein used is three layers, the decomposition direction of every layer is 2k, and wherein k represents the number of plies of decomposition;
Second step: observing matrix M × N is used to observe to the original signal of N dimension the observation vector Y obtaining M dimension, M<<N, then optimization method high probability reconstruct from observed reading Y is utilized, construct a calculation matrix, respectively to LH, HL, HH high-frequency sub-band is measured; Obtain the measurement coefficient value matrix of these three subbands. and it is constant to retain low frequency sub-band coefficient LL, chooses Gauss with matrix as calculation matrix, carries out measurement value.LL: the low low-value that wavelet decomposition obtains; LH: the low high-value that wavelet decomposition obtains; HL: the height place value that wavelet decomposition obtains; HH: the high high-value that wavelet decomposition obtains;
3rd step: use the assessment adopting weighted mean (Weighted Average, the WA) rule based on window to carry out pixel rank to the low frequency component of two width source images, chooses interior each of matrix and puts corresponding numerical value, form new matrix;
4th step: use principal component analysis (PCA) to merge to the high fdrequency component after the measurement of two width source images, obtains the fusion matrix after measuring;
5th step: use two stage cultivation tracing algorithm (StOMP) to be reconstructed calculating to the high fdrequency component after merging, obtain the high fdrequency component being suitable for WBCT inverse transformation;
6th step: use WBCT inverse transformation, the high fdrequency component of all directions and the fusion results of low frequency component are combined, obtains fused image.
2. as claimed in claim 1 based on the image interfusion method of compressed sensing and WBCT conversion, it is characterized in that, adopt weighted mean (Weighted Average, WA) the regular concrete steps based on window to be:
The first step: in source images, calculates energy centered by (m, n) point in surrounding window region or the variance tolerance S as this detailed information intensity
a(m, n), S
b(m, n), S
a(m, n) is the tolerance of A source images point detailed information intensity, S
b(m, n) is the tolerance of B source images point detailed information intensity, and m, n are point coordinate;
Second step, calculates c
a(m, n) and c
alocal, normalized cross-correlation coefficient M between (m, n)
aB, c
a(m, n) is A source images coefficient, c
b(m, n) is B source images coefficient;
3rd step, according to cross-correlation coefficient size, take different amalgamation modes:
Work as M
aBduring≤a, a is threshold value, a=0.85, illustrates that between source images coefficient, correlativity is lower, chooses the large coefficient of local variance reasonable for merging rear coefficients comparison, namely
Wherein c
f(m, n) is fused image coefficient,
Work as M
aBduring >a, illustrate that between coefficient, correlativity is larger, adopt average weighted method more reasonable,
c
F(m，n)＝w(m，n)c
A(m，n)+[E(m，n)-w(m，n)]c
B(m，n)
Wherein E (m, n) is unit matrix, and weight coefficient w (m, n) is determined by following formula:
3. as claimed in claim 1 based on the image interfusion method that compressed sensing and WBCT convert, it is characterized in that, compressed sensing restructing algorithm is further refined as: what select during each match tracing is multiple matched atoms instead of single atom, decrease matching times, can form a matched filter under particular state, identify the coordinate that all amplitudes are greater than a special selected threshold value, the coordinate selected with these does least square, then deduct least square fitting, obtain a new surplus.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510037279.5A CN104732504A (en) | 2015-01-23 | 2015-01-23 | Image fusion method based on compressed sensing and WBCT |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510037279.5A CN104732504A (en) | 2015-01-23 | 2015-01-23 | Image fusion method based on compressed sensing and WBCT |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104732504A true CN104732504A (en) | 2015-06-24 |
Family
ID=53456375
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510037279.5A Pending CN104732504A (en) | 2015-01-23 | 2015-01-23 | Image fusion method based on compressed sensing and WBCT |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104732504A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105976346A (en) * | 2016-04-28 | 2016-09-28 | 电子科技大学 | Infrared and visible light image fusion method based on robust principal component sparse decomposition |
| CN107622482A (en) * | 2017-09-13 | 2018-01-23 | 电子科技大学 | A kind of image interfusion method based on leukorrhea micro-imaging |
| CN108830819A (en) * | 2018-05-23 | 2018-11-16 | 青柠优视科技（北京）有限公司 | A kind of image interfusion method and device of depth image and infrared image |
| CN109633760A (en) * | 2018-12-05 | 2019-04-16 | 北京大学 | A kind of underground fluid monitoring method based on the equivalent fusion of imaging of natural potential |
| CN114399448A (en) * | 2021-11-22 | 2022-04-26 | 中国科学院西安光学精密机械研究所 | Multi-polarization information gating fusion method based on non-subsampled shear wave transformation |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102254314A (en) * | 2011-07-17 | 2011-11-23 | 西安电子科技大学 | Visible-light/infrared image fusion method based on compressed sensing |
| CN103559696A (en) * | 2013-09-29 | 2014-02-05 | 南京邮电大学 | Image fusion method based on compressed sensing |
-
2015
- 2015-01-23 CN CN201510037279.5A patent/CN104732504A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102254314A (en) * | 2011-07-17 | 2011-11-23 | 西安电子科技大学 | Visible-light/infrared image fusion method based on compressed sensing |
| CN103559696A (en) * | 2013-09-29 | 2014-02-05 | 南京邮电大学 | Image fusion method based on compressed sensing |
Non-Patent Citations (5)
| Title |
|---|
| DAVID L. DONOHO 等: "Sparse Solution of Underdetermined Systems of Linear Equations by Stagewise Orthogonal Matching Pursuit", 《IEEE TRANSACTIONS ON INFORMATION THEORY》 * |
| 侯超 等: "基于小波-Contourlet变换的智能监控图像融合的增强方法", 《控制工程》 * |
| 周爱平 等: "基于二代 Curvelet变换与 MPCA的可见光与红外图像融合", 《计算机应用》 * |
| 朱石争: "小波域多聚焦图像融合算法的研究", 《中国硕士学位论文全文数据库 信息科技辑》 * |
| 肖易寒: "基于Contourlet变换的全景图像处理关键技术研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105976346A (en) * | 2016-04-28 | 2016-09-28 | 电子科技大学 | Infrared and visible light image fusion method based on robust principal component sparse decomposition |
| CN107622482A (en) * | 2017-09-13 | 2018-01-23 | 电子科技大学 | A kind of image interfusion method based on leukorrhea micro-imaging |
| CN108830819A (en) * | 2018-05-23 | 2018-11-16 | 青柠优视科技（北京）有限公司 | A kind of image interfusion method and device of depth image and infrared image |
| CN108830819B (en) * | 2018-05-23 | 2021-06-18 | 青柠优视科技（北京）有限公司 | Image fusion method and device for depth image and infrared image |
| CN109633760A (en) * | 2018-12-05 | 2019-04-16 | 北京大学 | A kind of underground fluid monitoring method based on the equivalent fusion of imaging of natural potential |
| CN114399448A (en) * | 2021-11-22 | 2022-04-26 | 中国科学院西安光学精密机械研究所 | Multi-polarization information gating fusion method based on non-subsampled shear wave transformation |
| CN114399448B (en) * | 2021-11-22 | 2023-04-11 | 中国科学院西安光学精密机械研究所 | Multi-polarization information gating fusion method based on non-subsampled shear wave transformation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104504740A (en) | Image fusion method of compressed sensing framework | |
| Yang et al. | Seismic data reconstruction via matrix completion | |
| CN100557633C (en) | Multisource image anastomosing method based on pectination ripple and Laplace tower decomposition | |
| CN104732504A (en) | Image fusion method based on compressed sensing and WBCT | |
| CN102142136B (en) | Neural network based sonar image super-resolution reconstruction method | |
| CN104217448B (en) | Magnetic resonance fast imaging method and system based on iterative characteristic amendment | |
| CN101303764A (en) | Method for self-adaption amalgamation of multi-sensor image based on non-lower sampling profile wave | |
| CN103247028A (en) | Multi-hypothesis prediction block compressed sensing image processing method | |
| CN105869126A (en) | Compression perception denoising method of high-spectral images | |
| CN103514600B (en) | A kind of infrared target fast robust tracking based on rarefaction representation | |
| Qiusheng et al. | Compressed sensing MRI based on the hybrid regularization by denoising and the epigraph projection | |
| CN106296583B (en) | Based on image block group sparse coding and the noisy high spectrum image ultra-resolution ratio reconstructing method that in pairs maps | |
| CN113421198B (en) | Hyperspectral image denoising method based on subspace non-local low-rank tensor decomposition | |
| CN108986074B (en) | Coal mine underground image processing method based on blocking sparse algorithm | |
| CN102196155B (en) | Self-adaptive coefficient shrinkage video denoising method based on Surfacelet transform (ST) | |
| CN103236041B (en) | A kind of image super-resolution rebuilding method based on contourlet transformation | |
| CN107133921A (en) | The image super-resolution rebuilding method and system being embedded in based on multi-level neighborhood | |
| CN109584190B (en) | Contrast image enhancement method based on wavelet transformation low-frequency histogram limitation | |
| CN107622479B (en) | Contour wave sub-band self-adaptive detail injection method for multi-spectral-band image panchromatic sharpening | |
| Liu et al. | SAR image denoising based on patch ordering in nonsubsample shearlet domain | |
| Indira et al. | Pixel based medical image fusion techniques using discrete wavelet transform and stationary wavelet transform | |
| CN102509268B (en) | Immune-clonal-selection-based nonsubsampled contourlet domain image denoising method | |
| CN103747268B (en) | A kind of layered self-adapting threshold video denoising method | |
| Balasubramanian et al. | MRI super-resolution using generative adversarial network and discrete wavelet transform | |
| Liu et al. | A novel noise reduction method for optical coherence tomography images |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20150624 |