CN105825529B - It is a kind of based on non local and low-rank decomposition method for compressing high spectrum image - Google Patents
It is a kind of based on non local and low-rank decomposition method for compressing high spectrum image Download PDFInfo
- Publication number
- CN105825529B CN105825529B CN201610157079.8A CN201610157079A CN105825529B CN 105825529 B CN105825529 B CN 105825529B CN 201610157079 A CN201610157079 A CN 201610157079A CN 105825529 B CN105825529 B CN 105825529B
- Authority
- CN
- China
- Prior art keywords
- rank
- high spectrum
- spectrum image
- tensor
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10036—Multispectral image; Hyperspectral image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
Abstract
The invention discloses a kind of based on non local and low-rank decomposition method for compressing high spectrum image, piecemeal is carried out to high spectrum image using non local similitude, obtain the identical EO-1 hyperion fritter of size, later according to similitude, similar piece will be gathered for one kind, the 4 orders evidence of of a sort piece of composition can be by low-rank decomposition, because the size of data after decomposing is much smaller than initial data size, to realize the compression to high-spectral data, and the spatial structure characteristic and spectral composition feature of high-spectral data can be made full use of.
Description
Technical field
The invention belongs to computer image processing technology field, it is related to a kind of compression of images and reconstructing method, and in particular to
It is a kind of based on non local and low-rank decomposition method for compressing high spectrum image.
Background technique
High spectrum image is in it includes multiple wave bands, and for gray level image and RGB image, EO-1 hyperion contains more
Spectral information, therefore the precision of image procossing can be greatly improved.Although as the maturation and cost of high light spectrum image-forming technology
Reduction, high spectrum image is more and more used, but high spectrum image still has some restrictive conditions.
1) multiple spectrum of high spectrum image provide more useful informations, while also providing more redundancies, greatly
The time complexity and space complexity of image procossing are improved greatly.
2) high spectrum image is often taken as multiple gray level image processing in previous application, and each spectrum can be used as
One width gray level image, this treatment process can lose spectral composition information.
3) it not only needs individually to handle single spectrum when some algorithm process high spectrum images, also by each spectrum institute table
The 2 dimension images shown pull into the vector of 1 dimension, and this processing mode is not only lost the spectral composition information of image, also destroys list
The spatial structural form of width gray level image.
Therefore, it is necessary to a kind of spatial structural form that can retain high spectrum image and spectral composition information and protecting
Stay the algorithm that redundancy is abandoned on the basis of high spectrum image key useful information.
Summary of the invention
In order to solve the above-mentioned technical problems, the present invention provides the spatial structural forms that one kind can retain high spectrum image
With spectral composition information, and can be abandoned on the basis of retaining high spectrum image key useful information redundancy based on non-office
The method for compressing high spectrum image in portion and low-rank decomposition.
The technical scheme adopted by the invention is that: it is a kind of based on non local and low-rank decomposition Compression of hyperspectral images side
Method, which comprises the following steps:
Step 1: it is long to high spectrum image and it is wide be not that 8 image of multiple is cut, be cut to it is long and it is wide be all 8
Multiple, spectrum number are constant;
Step 2: the high spectrum image after cutting out being standardized, makes each of which pixel value between 0 and 255;
Step 3: piecemeal being carried out to standardized high spectrum image, each piece of space size is 8*8, then presses similitude
Block is clustered, each class constitutes a 4 rank tensors;
Step 4: low-rank decomposition being carried out to each class respectively, obtains the sparse tensor of 14 rank and the dictionary square of 3 low-ranks
Battle array;
Step 5: finally obtained tensor sum dictionary matrix is exactly compressed all data;
Step 6: verifying compression effectiveness is compared according to compressed data reconstruction high spectrum image with original image.
Preferably, clustering by similitude to block described in step 3, specific implementation process includes following sub-step
It is rapid:
Step 3.1: the window for being 8*8 with a size is mobile with step-length 8 on the space of high spectrum image, obtains space
Size is the EO-1 hyperion block that 8*8 and spectrum number remain unchanged.There is no space covering between each adjacent block;
Step 3.2: classification number (obtaining classification number divided by 100 with block number in this patent) being asked according to block number, uses kmeans+
+ algorithm clusters all pieces, so that similar piece is in same class;
Step 3.3: since each piece of dimension is identical, so of a sort piece may be constructed a 4 rank tensors.Different classes
Because block number difference is the 4th rank of corresponding 4 rank tensor is of different sizes.
Preferably, carrying out low-rank decomposition respectively to each class described in step 4, specific implementation process includes following son
Step:
Step 4.1: providing the target dimension of coefficient tensor;
Step 4.2: figure gram is carried out to 4 rank tensors according to target dimension and is decomposed, it is after decomposition the result is that 14 rank tensor sum 4
A matrix;
The 4th matrix multiple of step 4.3:4 rank tensor sum the result is that the coefficient of requirement, other 3 matrixes are to require
Dictionary.
Preferably, being by coefficient tensor and word according to compressed data reconstruction high spectrum image described in step 6
Allusion quotation matrix multiple obtains tensor block, and tensor block is reconfigured to the image after reconstruct can be obtained by the sequence of piecemeal.
The invention has the benefit that
(1) present invention introduces tensor, tensor can directly indicate the hyperspectral image data of 3 ranks;Pressure of the present invention
Compression algorithm directly handles tensor data, thus can retaining space structural information and spectral composition information simultaneously, can solve
State problem.
(2) present invention introduces non local similitude, non local similitude is usually used in image denoising;Natural image is come
Say often there are many fritters much like, it is believed that these fritters can be substituted for each other, and be calculated in compression of images of the present invention
In method, the high spectrum image of tensor representation is divided into the equal tensor block (space size is defaulted as 8*8) of many space sizes, often
The spectrum number of a tensor block be it is identical, this can guarantee that algorithm can retain spectral composition information, while fritter can retain
Spatial structural form.Then similar piece is gathered for one kind, the 4 rank tensors that every one kind forms can be decomposed by dictionary learning
The dictionary and coefficient of low-dimensional, to realize data compression.
(3) present invention introduces low-rank decomposition, the 4 rank tensors for being clustered by similar block and being formed will be decomposed into a dictionary tensor
With 3 coefficient matrixes, since the block for including in each class is similar, it can be considered that obtained coefficient matrix is also equal
's.
(4) present invention is realized preferably retains the key message useful to subsequent processing while compressing image.
Detailed description of the invention
Fig. 1: for the compression algorithm piecemeal and sorting procedure of the embodiment of the present invention.
Fig. 2: for the compression algorithm low-rank decomposition step of the embodiment of the present invention.
Specific embodiment
Understand for the ease of those of ordinary skill in the art and implement the present invention, with reference to the accompanying drawings and embodiments to this hair
It is bright to be described in further detail, it should be understood that implementation example described herein is merely to illustrate and explain the present invention, not
For limiting the present invention.
It is provided by the invention a kind of based on non local and low-rank decomposition method for compressing high spectrum image, including following step
It is rapid:
Step 1: it is long to high spectrum image and it is wide be not that 8 image of multiple is cut, be cut to it is long and it is wide be all 8
Multiple, spectrum number are constant;
Step 2: the high spectrum image after cutting out being standardized, makes each of which pixel value between 0 and 255;
Step 3: piecemeal being carried out to standardized high spectrum image, each piece of space size is 8*8, then presses similitude
Block is clustered, each class constitutes a 4 rank tensors;
Referring to Fig.1, high spectrum image tensor representation is divided into size with not covering it to 3 rank tensors of a higher-dimension
Identical Xiao Zhang's gauge block;Then, gather content in these tensor blocks is similar for one kind, each class may be constructed 4 ranks
Tensor.Then, 1 original 3 rank tensor become the lesser 4 rank tensor of several pixel dimensions.
Step 4: low-rank decomposition being carried out to each class respectively, obtains the sparse tensor of 14 rank and the dictionary square of 3 low-ranks
Battle array;
See Fig. 2, according to dictionary learning, 4 rank tensors made of each cluster can be broken into 1 coefficient tensor
With 3 dictionary matrix DsH, DW, DHProduct.Since all tensor blocks in each class are similar, and can mutually replace
Generation, thus decompose obtained coefficient tensor be all it is identical, equally, dictionary matrix is also identical.From Figure 2 it can be seen that low-rank point
Element in tensor sum matrix after solution not for 0 is seldom, and total data volume is far smaller than the data volume of original high spectrum image,
To realize the purpose of compression of images.
Step 5: finally obtained tensor sum dictionary matrix is exactly compressed all data;
Step 6: verifying compression effectiveness is compared according to compressed data reconstruction high spectrum image with original image.
In order to verify the efficiency of compression, need compressed data reverting to original image size, recovery process is to lead to
It crosses coefficient tensor and dictionary matrix multiple obtains tensor block, tensor block is reconfigured by the sequence of piecemeal after reconstruct can be obtained
Image.
It is the realization step of Hyperspectral image compression algorithm of the present invention above.Pass through tensor, non local similitude
With the introducing of low-rank decomposition, the advantage of high spectrum image can be made full use of, retains weight while realizing significantly compressed data
Want information.
There are also following points for attention when specific implementation:
Firstly, since the previously mentioned block size by high spectrum image piecemeal, defaulted in the algorithm is 8*8, and
Without covering between block and block, thus the space size of original high spectrum image must satisfy it is long and it is wide be all 8 multiple, otherwise can
There is piecemeal mistake.Therefore often image is cut before piecemeal, the edge of the data of multiple requirement will be unsatisfactory for
Cutting makes its satisfaction.
Secondly as the difference of data acquisition equipment, often difference is very big for the pixel value of collected data, it is likely that meeting
Image compression effectiveness is influenced to be standardized the image data of clipped mistake before piecemeal in order to avoid this problem,
Even if all pixels value is in 0-255 within the scope of this, then carries out next processing.This two step will as the present invention relates to
Image compression algorithm pre-treatment step.
It should be understood that the part that this specification does not elaborate belongs to the prior art.
It should be understood that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this
The limitation of invention patent protection range, those skilled in the art under the inspiration of the present invention, are not departing from power of the present invention
Benefit requires to make replacement or deformation under protected ambit, fall within the scope of protection of the present invention, this hair
It is bright range is claimed to be determined by the appended claims.
Claims (2)
1. a kind of based on non local and low-rank decomposition method for compressing high spectrum image, which comprises the following steps:
Step 1: it is long to high spectrum image and it is wide be not that 8 image of multiple is cut, be cut to it is long and it is wide be all 8 times
Number, spectrum number are constant;
Step 2: the high spectrum image after cutting out being standardized, makes each of which pixel value between 0 and 255;
Step 3: piecemeal being carried out to standardized high spectrum image, each piece of space size is 8*8, then by similitude to block
It is clustered, each class constitutes a 4 rank tensors;
Described to cluster by similitude to block, specific implementation process includes following sub-step:
Step 3.1: the window for being 8*8 with a size is mobile with step-length 8 on the space of high spectrum image, obtains space size
For 8*8 and EO-1 hyperion block that spectrum number remains unchanged;There is no space covering between each adjacent block;
Step 3.2: classification number being asked according to block number, all pieces are clustered with kmeans++ algorithm, so that similar piece is in same
Class;Since each piece of dimension is identical, so of a sort piece can constitute a 4 rank tensors, different classes is because block number is different
So the 4th rank of corresponding 4 rank tensor is of different sizes;
Step 4: low-rank decomposition being carried out to each class respectively, obtains the sparse tensor of 14 rank and the dictionary matrix of 3 low-ranks;
Described to carry out low-rank decomposition respectively to each class, specific implementation process includes following sub-step:
Step 4.1: providing the target dimension of coefficient tensor;
Step 4.2: figure gram is carried out to 4 rank tensors according to target dimension and is decomposed, it is after decomposition the result is that 14 rank tensor sum, 4 squares
Battle array;4 matrix multiples of rank tensor sum the 4th the result is that the coefficient of requirement, other 3 matrixes are desired dictionaries;
Step 5: finally obtained tensor sum dictionary matrix is exactly compressed all data;
Step 6: verifying compression effectiveness is compared according to compressed data reconstruction high spectrum image with original image.
2. according to claim 1 based on non local and low-rank decomposition method for compressing high spectrum image, it is characterised in that:
It according to compressed data reconstruction high spectrum image described in step 6, is opened by coefficient tensor and dictionary matrix multiple
Tensor block is reconfigured the image after reconstruct can be obtained by gauge block by the sequence of piecemeal.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610157079.8A CN105825529B (en) | 2016-03-18 | 2016-03-18 | It is a kind of based on non local and low-rank decomposition method for compressing high spectrum image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610157079.8A CN105825529B (en) | 2016-03-18 | 2016-03-18 | It is a kind of based on non local and low-rank decomposition method for compressing high spectrum image |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105825529A CN105825529A (en) | 2016-08-03 |
CN105825529B true CN105825529B (en) | 2019-02-26 |
Family
ID=56523553
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610157079.8A Active CN105825529B (en) | 2016-03-18 | 2016-03-18 | It is a kind of based on non local and low-rank decomposition method for compressing high spectrum image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105825529B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106651820B (en) * | 2016-09-23 | 2019-06-21 | 西安电子科技大学 | Remote sensing image fusion method based on sparse tensor neighbour insertion |
CN107067380B (en) * | 2017-03-28 | 2020-04-28 | 天津大学 | High-resolution image reconstruction method based on low-rank tensor and hierarchical dictionary learning |
CN107247688A (en) * | 2017-06-02 | 2017-10-13 | 上海德衡数据科技有限公司 | A kind of multichannel data compressed sensing cube framework |
CN107944556B (en) * | 2017-12-12 | 2020-09-08 | 电子科技大学 | Deep neural network compression method based on block item tensor decomposition |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103020912A (en) * | 2012-12-18 | 2013-04-03 | 武汉大学 | Remote sensing image restoration method combining wave-band clustering with sparse representation |
CN104933684A (en) * | 2015-06-12 | 2015-09-23 | 北京工业大学 | Light field reconstruction method |
CN105160623A (en) * | 2015-08-17 | 2015-12-16 | 河南科技学院 | Unsupervised hyperspectral data dimension reduction method based on block low-rank tensor model |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB0914982D0 (en) * | 2009-08-27 | 2009-09-30 | Univ East Anglia | Methods and apparatus for generating accented image data |
-
2016
- 2016-03-18 CN CN201610157079.8A patent/CN105825529B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103020912A (en) * | 2012-12-18 | 2013-04-03 | 武汉大学 | Remote sensing image restoration method combining wave-band clustering with sparse representation |
CN104933684A (en) * | 2015-06-12 | 2015-09-23 | 北京工业大学 | Light field reconstruction method |
CN105160623A (en) * | 2015-08-17 | 2015-12-16 | 河南科技学院 | Unsupervised hyperspectral data dimension reduction method based on block low-rank tensor model |
Also Published As
Publication number | Publication date |
---|---|
CN105825529A (en) | 2016-08-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105825529B (en) | It is a kind of based on non local and low-rank decomposition method for compressing high spectrum image | |
Yuan et al. | Hyperspectral band selection by multitask sparsity pursuit | |
Dony | Karhunen-loeve transform | |
Agarwal et al. | Content based image retrieval using color edge detection and discrete wavelet transform | |
WO2022022001A1 (en) | Method for compressing style transfer network, and style transfer method, apparatus and system | |
EP0934662A1 (en) | Vector quantisation codebook generation method | |
CN111324791B (en) | Multi-view data subspace clustering method | |
Phien et al. | Efficient tensor completion: Low-rank tensor train | |
Dong et al. | Feature extraction through contourlet subband clustering for texture classification | |
CN107800437B (en) | Data compression method and device | |
CN106803105B (en) | Image classification method based on sparse representation dictionary learning | |
Wang et al. | Neighborhood preserving non-negative tensor factorization for image representation | |
Aniello et al. | Characterizing the dynamical semigroups that do not decrease a quantum entropy | |
Khashman et al. | Neural networks arbitration for optimum DCT image compression | |
Daga et al. | Image compression using harmony search algorithm | |
Ali et al. | Image Denoising with Color Scheme by Using Autoencoders | |
CN106909944B (en) | Face picture clustering method | |
Yumin et al. | Image retrieval based on multiple features using wavelet | |
Krishna et al. | Biorthogonal Wavelets for Multiresolution Image Compression | |
Vaish et al. | A new Image compression technique using principal component analysis and Huffman coding | |
Li et al. | Nonconvex multi-view subspace clustering via simultaneously learning the representation tensor and affinity matrix | |
Qi et al. | A DT-CWT-based infrared-visible image fusion method for smart city | |
Benedetto et al. | Wavelet packets for multi-and hyper-spectral imagery | |
Abood | Composite techniques based color image compression | |
Ahmed et al. | WSN application based on image compression using AHAAR wavelet transform |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |