CN104361614A - Polarization SAR image compression method based on multi-direction dictionary learning - Google Patents
Polarization SAR image compression method based on multi-direction dictionary learning Download PDFInfo
- Publication number
- CN104361614A CN104361614A CN201410682164.7A CN201410682164A CN104361614A CN 104361614 A CN104361614 A CN 104361614A CN 201410682164 A CN201410682164 A CN 201410682164A CN 104361614 A CN104361614 A CN 104361614A
- Authority
- CN
- China
- Prior art keywords
- matrix
- image
- coefficients
- frequency sub
- dimensional
- 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
Abstract
The invention discloses a polarization SAR image compression method based on multi-direction dictionary learning. The method mainly solves the problem that the quality of an image compressed and reconstructed through the existing technology is low. The method includes the implementation steps of firstly, inputting a set of polarization SAR four-channel images, and conducting asymmetrical three-dimensional wavelet transform; secondly, conducting sparse representation on coefficient matrixes, in different directions, of high-frequency sub-bands under different scales of all the channels after the discrete wavelet transform is conducted so as to obtain all sparse matrixes; thirdly, conducting quantizing and coding on the coefficient matrixes of all low-frequency sub-bands after the discrete wavelet transform is conducted to obtain low-frequency code streams; fourthly, conducting unified quantizing and coding on the sparse matrixes in different scales and different directions so as to obtain high-frequency code streams; fifthly, forming final code streams through the low-frequency code streams and the high-frequency code streams. By means of the method, redundancy between the channels can be effectively eliminated, the marginal information and contour information of images are better preserved, the quality of a compressed and reconstructed image is improved, and the method can be used for transmitting and storing polarization SAR images.
Description
Technical field
The invention belongs to technical field of image processing, further relate to polarimetric synthetic aperture radar SAR image compression method, can be used for transmission and the storage of Polarimetric SAR Image.
Background technology
Synthetic-aperture radar SAR, as the high-resolution microwave remote sensing instrument of one, has imaging characteristics that is round-the-clock, round-the-clock, containing a large amount of signal characteristics, is applied in remote sensing fields widely, as military surveillance, topographic mapping, target identification etc.Polarimetric synthetic aperture radar SAR provides more information than traditional synthetic-aperture radar, greatly strengthen the processing power for information.But corresponding data volume can increase at double.Owing to needing transmission real-time for jumbo data and storing, the compression of data is just seemed particularly necessary.
Traditional Polarimetric SAR Image compression method regards its four-way image as independently 4 SAR image, SAR image compression method is utilized to compress each channel image respectively, Polarimetric SAR Image self-characteristic can not be utilized fully, as the correlativity between passage, do not reach desirable compression effectiveness.In recent years, some scholars are had to take into account the characteristic of Polarimetric SAR Image self and be applied in compression, as a civilian superfine people the article pointed out at 3 D-SPIHT Compression of Multi-polarimetric SAR Images one, by the HH channel image of Polarimetric SAR Image, VV channel image, HV channel image integrally, carries out three-dimensional matrice conversion, and then the SPIHT coding of application enhancements carries out compression transmission.Although the method is considered and the effective redundancy removed between POLARIZATION CHANNEL, due in compressibility one timing, spiht algorithm effectively can not retain marginal information and the profile information of image, causes the reduction of reconstructed image quality.
Summary of the invention
The object of the invention is to overcome the deficiencies in the prior art, propose a kind of Polarimetric SAR Image compression method based on multi-direction dictionary learning.Effectively to remove the redundancy of redundancy between polarization SAR passage and each passage therein, reduce the loss of image important information, improve the quality of reconstructed image.
For the concrete steps realizing the object of the invention comprise as follows:
(1) input one group of polarization SAR four-way image to be compressed, and asymmetric three-dimensional dispersed data is carried out to this four-way image, obtain high-frequency sub-band matrix of coefficients under each channel low frequency sub-band coefficients matrix and different scale;
(2) by the level of high-frequency sub-band under each passage different scale, vertical, the matrix of coefficients use level in direction, diagonal angle three, the dictionary rarefaction representation respectively vertically, to angular direction, the sparse matrix of each passage different scale different directions is obtained;
(3) the low frequency sub-band matrix of coefficients after each channel discrete wavelet transformation is quantized, DPCM predictive coding is used to encode to its quantization parameter matrix, with Huffman coding, the remaining prediction residual of DPCM predictive coding is encoded again, obtain low frequency code stream;
(4) to the sparse matrix unified quantization of each passage different scale different directions, with Huffman coding, the coefficient amplitude value and coefficient positions index value that quantize sparse matrix are encoded, obtain high frequency code stream;
(5) form final code stream with low frequency code stream and high frequency code stream, and export.
The present invention compared with prior art has the following advantages:
First, the present invention carries out three-dimensional dispersed data owing to treating compression polarization SAR four-way image, tentatively eliminate between POLARIZATION CHANNEL and the redundancy of each channel interior, remaining redundancy after wavelet field utilizes dictionary learning algorithm to remove conversion afterwards, overcome existing method and only remove insufficient shortcoming by the method redundancy of conversion, further increase compressibility.
Second, the present invention carries out rarefaction representation to respective direction high-frequency sub-band matrix of coefficients due to the dictionary using multi-direction RLS dictionary learning method and obtain, overcome the shortcoming that prior art effectively can not retain image edge information and profile information, greatly enhance the quality of reconstructed image.
Accompanying drawing explanation
Fig. 1 is realization flow figure of the present invention;
Fig. 2 is the original graph of one group of polarization SAR four-way image to be compressed used in the present invention;
Fig. 3 is horizontal direction different scale high-frequency sub-band matrix of coefficients schematic diagram after two-dimensional discrete wavelet conversion;
Fig. 4 is that the present invention and existing method contrast the compression reconfiguration of the HH channel image of one group of Polarimetric SAR Image under 0.25bpp.
Embodiment
Below in conjunction with accompanying drawing, the present invention will be further described.
With reference to accompanying drawing 1, performing step of the present invention is as follows:
Step 1: input picture.
One group of polarization SAR four-way image to be compressed of input option.As shown in Figure 2, Fig. 2 is one group of Polarimetric SAR Image (NASA/JPL, 1988) in san francisco, usa area to the polarization SAR four-way image to be compressed used in example of the present invention, is designated as San Francisco in the present invention.Wherein, Fig. 2 (a) is HH channel image, Fig. 2 (b) is HV channel image, Fig. 2 (c) is VH channel image, Fig. 2 (d) is VV channel image, each width image size is 512*512, and image bit is 8 deeply, and form is BMP.
Step 2: carry out asymmetric three-dimensional dispersed data to the polarization SAR four-way image of input, obtains high-frequency sub-band matrix of coefficients under each channel low frequency sub-band coefficients matrix and different scale.
The first step, becomes three-dimensional matrice X (x, y, z) by four-way image array according to the sequential configuration of HH/HV/VH/VV four-way;
Second step, carries out to three-dimensional matrice X (x, y, z) one-dimensinal discrete small wave transformation that matrix directions yardstick is 1, obtains the three-dimensional matrice Y (x, y, P) after one-dimensinal discrete small wave transformation;
3rd step, carries out to the every one page of three-dimensional matrice Y (x, y, P) after conversion the two-dimensional discrete wavelet conversion that yardstick is L respectively, obtains high-frequency sub-band matrix of coefficients under each matrix page low frequency sub-band matrix of coefficients and different scale, gets L=3 in experiment.
Wherein, x, y, z represent row, column and the page of three-dimensional matrice X (x, y, z) respectively, and P represents the page of three-dimensional matrice Y (x, y, P).In experiment, P value is 0,1,2,3.
Step 3: to the level of high-frequency sub-band under each passage different scale, vertical, by the dictionary of correspondence direction, rarefaction representation is carried out to the matrix of coefficients of angular direction, obtain the sparse matrix of each passage different scale different directions.
(3a) dictionary of all directions is obtained according to the method for RLS dictionary learning:
(3a1) all directions sample set matrix needed for dictionary training is constructed:
(3a1_1) existing polarization SAR four-way image is selected, and four-way image array is become three-dimensional matrice A (x according to the sequential configuration of HH/HV/VH/VV four-way, y, z), to this three-dimensional matrice A (x, y, z) one-dimensinal discrete small wave transformation that matrix directions yardstick is 1 is carried out, obtain the three-dimensional matrice B (x, y, P) after one-dimensinal discrete small wave transformation.Wherein, x, y, z represent row, column and the page of three-dimensional matrice A (x, y, z) respectively, and P represents the page of three-dimensional matrice B (x, y, P).In experiment, P value is 0,1,2,3;
(3a1_2) two-dimensional discrete wavelet conversion that yardstick is L is carried out respectively to the every one page of the three-dimensional matrice B (x, y, P) after one-dimensinal discrete small wave transformation, obtain the matrix of coefficients of high-frequency sub-band different directions under each matrix page different scale; L=3 is got in experiment;
Note three-dimensional matrice B (x, y, P) the 1st page, the 2nd page, the 3rd page, the 4th page yardstick is that after the two-dimensional discrete wavelet conversion of 1, the matrix of coefficients of high-frequency sub-band horizontal direction is respectively C11, C12, C13, C14; Yardstick is that the matrix of coefficients of the high-frequency sub-band horizontal direction of 2 is respectively C21, C22, C23, C24; Yardstick is that the matrix of coefficients of the high-frequency sub-band horizontal direction of 3 is respectively C31, C32, C33, C34.
Such as, under 1st page of different scale, the matrix of coefficients of high-frequency sub-band horizontal direction is formed as shown in Figure 3, wherein C11 is that after two-dimensional discrete wavelet conversion, matrix of coefficients divides equally the upper right portion after 4 parts, C21 is after matrix of coefficients divides equally 4 parts, get the upper right portion that its upper left divides equally 4 parts again, C31 is after matrix of coefficients divides equally 4 parts, gets after its upper left divides equally 4 parts again, gets the upper right portion that its upper left divides equally 4 parts again.The formation of other matrix pages is identical with the 1st page.
The matrix of coefficients of each yardstick high-frequency sub-band horizontal direction, with identical with the matrix of coefficients constructive method of high-frequency sub-band horizontal direction to the matrix of coefficients of angular direction, only need be replaced with high-frequency sub-band vertical direction and the matrix of coefficients to angular direction by high-frequency sub-band vertical direction;
(3a1_3) matrix stack of all directions carried out matrix conversion successively and be linked in sequence, forming the sample set matrix in this direction.
Be described in detail for the make of horizontal direction sample matrix collection below, vertical identical with horizontal direction method with the building method of the sample matrix collection to angular direction.
To the matrix of coefficients piecemeal of each high-frequency sub-band horizontal direction, as C11, block size is 8*8, then to the block of each 8*8 according to from top to bottom, the method from left to right scanned forms the vector of a 64*1; Matrix-block after each conversion according to it at original matrix, as in C11 from top to bottom, sequential configuration from left to right becomes new matrix, is expressed as C11 '.The matrix of coefficients of the high-frequency sub-band horizontal direction after all conversions is combined according to the order of C11 ', C21 ', C31 ', C41 ', C12 ' C34 ', forms horizontal direction sample set matrix;
(3a2) respectively to each sample set matrix application RLS dictionary learning method, all directions dictionary is obtained:
(3a2_1) select initial dictionary D0, be made up of 441 vectors front in sample set matrix in experiment, initial C Matrix C
0, value and D
0unanimously, the forgetting factor λ in each iteration
iscope be 0≤λ
i≤ 1, wherein subscript i represents iterations;
(3a2_2) the training vector x that acquisition one is new from sample set matrix
i, obtain it at dictionary D
i-1under sparse coefficient w
i:
Wherein, s is w
ithe threshold value that middle nonzero value number is predetermined;
This sparse coefficient w
isolve the order recurrence matching pursuit algorithm using and proposed in the document " A fast orthogonal matching pursuit algorith " by people such as M.Gharavi-Alkhansari, make w
iand D
i-1for input vector and the input dictionary of algorithm, s is the threshold value of setting, final output sparse coefficient w
i;
(3a2_3) according to new training vector x
i, dictionary D
i-1with sparse coefficient w
i, error of calculation r:
r=x
i-D
i-1w
i;
(3a2_4) set
Calculate the vector of definition
Wherein
represent w
itransposition;
(3a2_5) dictionary D is upgraded according to the following formula
iwith C Matrix C
ifor:
D
i=D
i-1+αru
T,
(3a2_6) circulate (3a2_2)-(3a2-5) until subscript i meets the iterations of setting, obtain dictionary Di;
(3b) matrix of coefficients in the level of each passage different scale high-frequency sub-band, vertical, direction, diagonal angle three is in conjunction with the dictionary application order recurrence matching pursuit algorithm of correspondence direction, obtains the sparse matrix of each passage different scale different directions.
Step 4: the low frequency sub-band matrix of coefficients after each channel discrete wavelet transformation is quantized, DPCM predictive coding is used to encode to its quantization parameter matrix, with Huffman coding, the remaining prediction residual of DPCM predictive coding is encoded again, obtain low frequency code stream.
Step 5: to the sparse matrix unified quantization of each passage different scale different directions, encodes to the coefficient amplitude value and coefficient positions index value that quantize sparse matrix with Huffman coding, obtains high frequency code stream.
Step 6: form final output code flow with low frequency code stream and high frequency code stream.
Effect of the present invention further illustrates by following emulation.
1. emulation experiment condition:
Hardware test platform of the present invention is: processor is Inter Core 2Duo E6550, and dominant frequency is 2.33GHz, internal memory 2GB, and software platform is: Windows 7 Ultimate 32-bit operating system and Matlab R2010b.Input picture of the present invention is polarization SAR four-way image to be compressed, and each width image size is 512*512, and image bit is 8 deeply, and form is BMP.
The method that emulation uses is following three kinds:
1) civilian superfine people is opened at document " 3 D-SPIHT Compression of Multi-polarimetric SAR Images.Electronics and information journal, 2008 " the middle compression method to Polarimetric SAR Image proposed, is called for short 3D-SPIHT method.
2) a kind of compression method for image of proposing in document " JEPG2000; the next millennium compression standard for still images; 1999; 131-132 " of the people such as Maryline Charrier, be compression method the most frequently used in image processing field, be called for short JEPG2000 method.
3) the inventive method.
2. emulate content:
Apply method of the present invention and existing 3D-SPIHT method, the method of JEPG2000 carries out the coding transmission of different compressibility respectively to one group of San Francisco image, wherein restructuring graph that decoding and reconstituting obtains is carried out as shown in Figure 4 to the compressed bit stream under 0.25bpp, wherein Fig. 4 (a) is for the present invention is to the reconstructed image after the HH channel image compression transmission of San Francisco, Fig. 4 (b) is for prior art 3D-SPIHT is to the reconstructed image after the HH channel image compression transmission of San Francisco, Fig. 4 (c) is for prior art JEPG2000 is to the reconstructed image after the HH channel image compression transmission of San Francisco.
Described bpp refers to the figure place shared by each pixel of image, as 0.25bpp refers to that the compression of images original each pixel being taken 8 is that each pixel only takies 0.25.
3. analysis of simulation result:
As can be seen from the contrast of Fig. 4 tri-width restructuring graph, the reconstructed image of the inventive method, compared with the reconstructed image of prior art, better can keep the edge of image to be compressed and detailed information and the flatness of homogenous region is better.
In general, Y-PSNR PSNR is traditional image compression algorithm performance applications objective evaluation the most widely, is described the distortion of reconstructed image and original image by the norm measure of the error of calculation.Its computing formula is as follows:
PSNR=10×log
10(255
2/MSE),
Wherein, f is original image,
for reconstructed image, n
x, n
ybe respectively line number and the columns of image.
Adopt 3D-SPIHT and the JEPG2000 method of the inventive method and prior art to the PSNR value of San Francisco four-way compression of images as table 1, shown in table 2.
Table 1San Francisco four-way compression of images PSNR value
Table 2San Francisco four-way compression of images mean P SNR value
From table 1, the PSNR objective evaluation result of table 2 can be found out, under identical bpp, the inventive method is than existing 3D-SPIHT method, and JEPG2000 method image more effectively can protect original image information.
Claims (3)
1., based on a Polarimetric SAR Image compression method for multi-direction dictionary learning, comprise the steps:
(1) input one group of polarization SAR four-way image to be compressed, and asymmetric three-dimensional dispersed data is carried out to this four-way image, obtain high-frequency sub-band matrix of coefficients under each channel low frequency sub-band coefficients matrix and different scale;
(2) by the level of high-frequency sub-band under each passage different scale, vertical, the matrix of coefficients use level in direction, diagonal angle three, the dictionary rarefaction representation respectively vertically, to angular direction, the sparse matrix of each passage different scale different directions is obtained;
(3) the low frequency sub-band matrix of coefficients after each channel discrete wavelet transformation is quantized, DPCM predictive coding is used to encode to its quantization parameter matrix, with Huffman coding, the remaining prediction residual of DPCM predictive coding is encoded again, obtain low frequency code stream;
(4) to the sparse matrix unified quantization of each passage different scale different directions, with Huffman coding, the coefficient amplitude value and coefficient positions index value that quantize sparse matrix are encoded, obtain high frequency code stream;
(5) form final code stream with low frequency code stream and high frequency code stream, and export.
2. according to claim 1 based on the Polarimetric SAR Image compression method of multi-direction dictionary learning, it is characterized in that, described step carries out asymmetric three-dimensional dispersed data to four-way image in (1), in accordance with the following steps:
(2a) four-way image array is become three-dimensional matrice according to the sequential configuration of HH/HV/VH/VV four-way;
(2b) matrix directions one-dimensinal discrete small wave transformation is carried out to this three-dimensional matrice, obtain the three-dimensional matrice after one-dimensinal discrete small wave transformation;
(2c) two-dimensional discrete wavelet conversion that yardstick is L is carried out respectively to every one page of the three-dimensional matrice after above-mentioned conversion.
3. according to claim 1 based on the Polarimetric SAR Image compression method of multi-direction dictionary learning, it is characterized in that, the dictionary of all directions in described step (2), obtains as follows:
(3a) existing polarization SAR four-way image is selected, image array is become three-dimensional matrice according to HH/HV/VH/VV four-way sequential configuration, and asymmetric three-dimensional dispersed data is carried out to this three-dimensional matrice, obtain high-frequency sub-band matrix of coefficients under each passage different scale;
(3b) with the level of high-frequency sub-band under each passage different scale, the sample set vertically, to the matrix of coefficients structure respective direction dictionary training of angular direction, all directions dictionary is obtained according to the method for RLS dictionary learning.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410682164.7A CN104361614A (en) | 2014-11-21 | 2014-11-21 | Polarization SAR image compression method based on multi-direction dictionary learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410682164.7A CN104361614A (en) | 2014-11-21 | 2014-11-21 | Polarization SAR image compression method based on multi-direction dictionary learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN104361614A true CN104361614A (en) | 2015-02-18 |
Family
ID=52528872
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410682164.7A Pending CN104361614A (en) | 2014-11-21 | 2014-11-21 | Polarization SAR image compression method based on multi-direction dictionary learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104361614A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112995637A (en) * | 2021-03-10 | 2021-06-18 | 湘潭大学 | Multi-section medical image compression method based on three-dimensional discrete wavelet transform |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0914004A1 (en) * | 1997-10-29 | 1999-05-06 | Ntec Media GmbH | Coding system and method for lossless and lossy compression of still and motion images |
CN104065974A (en) * | 2014-06-30 | 2014-09-24 | 清华大学 | Image compression method and system |
-
2014
- 2014-11-21 CN CN201410682164.7A patent/CN104361614A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0914004A1 (en) * | 1997-10-29 | 1999-05-06 | Ntec Media GmbH | Coding system and method for lossless and lossy compression of still and motion images |
CN104065974A (en) * | 2014-06-30 | 2014-09-24 | 清华大学 | Image compression method and system |
Non-Patent Citations (3)
Title |
---|
XIN ZHAN, RONG ZHANG, DONG YIN, AND CHENGFU HUO: "《SAR Image Compression Using Multiscale Dictionary Learning and Sparse Representation》", 《IEEE GEOSCIENCE AND REMOTE SENSING LETTERS》 * |
张文超 王岩飞 潘志刚: "《基于三维矩阵变换的多极化SAR图像压缩》", 《电子学报》 * |
陈原: "《基于变换编码的SAR图像压缩技术研究》", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112995637A (en) * | 2021-03-10 | 2021-06-18 | 湘潭大学 | Multi-section medical image compression method based on three-dimensional discrete wavelet transform |
CN112995637B (en) * | 2021-03-10 | 2023-02-28 | 湘潭大学 | Multi-section medical image compression method based on three-dimensional discrete wavelet transform |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104331913B (en) | Polarimetric SAR Image compression method based on sparse K SVD | |
Zhan et al. | SAR image compression using multiscale dictionary learning and sparse representation | |
CN104199627B (en) | Gradable video encoding system based on multiple dimensioned online dictionary learning | |
Joshi et al. | Image and video compression: Fundamentals, Techniques, and Applications | |
Kumar et al. | A review: DWT-DCT technique and arithmetic-Huffman coding based image compression | |
CN103546759A (en) | Image compression coding method based on combination of wavelet packets and vector quantization | |
Meyer-Bäse et al. | Medical image compression using topology-preserving neural networks | |
Wang et al. | Hyperspectral image compression based on lapped transform and Tucker decomposition | |
Reddy et al. | Lossless compression of medical images for better diagnosis | |
CN106612435A (en) | Joint image compression method based on SVD-DWT-DCT | |
CN102918838B (en) | The coding method of a block of image sequence and reconstructing method | |
Dinu et al. | Performance analysis and comparison of medical image compression techniques | |
CN105354867A (en) | Hyperspectral image compression algorithm research of adaptive redundant dictionary compressed sensing | |
Joshua et al. | Comparison of DCT and DWT image compression | |
CN104270642B (en) | A kind of lossless compression method for high spectrum image chosen based on RKLT and principal component | |
Li et al. | Curvelet based image compression via core vector machine | |
CN104683818A (en) | Image compression method based on biorthogonal invariant set multi-wavelets | |
CN104361614A (en) | Polarization SAR image compression method based on multi-direction dictionary learning | |
CN103985096A (en) | Hyperspectral image regression prediction compression method based on off-line training | |
Jeong et al. | An overhead-free region-based JPEG framework for task-driven image compression | |
CN111770344A (en) | Laser spectrum image compression method and system based on deep learning network | |
CN102065291B (en) | Sparse representation model-based image decoding method | |
Al-Fayadh et al. | Singular value decomposition based classified vector quantization image compression method using discrete sine transform | |
Barret et al. | Lossy hyperspectral images coding with exogenous quasi optimal transforms | |
Pal et al. | A hybrid DCT-VQ based approach for efficient compression of color 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: 20150218 |