CN107240138B - Panchromatic remote sensing image compression method based on sample binary tree dictionary learning - Google Patents
Panchromatic remote sensing image compression method based on sample binary tree dictionary learning Download PDFInfo
- Publication number
- CN107240138B CN107240138B CN201710378830.1A CN201710378830A CN107240138B CN 107240138 B CN107240138 B CN 107240138B CN 201710378830 A CN201710378830 A CN 201710378830A CN 107240138 B CN107240138 B CN 107240138B
- Authority
- CN
- China
- Prior art keywords
- sample
- layer
- binary tree
- child nodes
- complex samples
- 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
- G06T9/40—Tree coding, e.g. quadtree, octree
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
- H04N19/124—Quantisation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/169—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding
- H04N19/17—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding the unit being an image region, e.g. an object
- H04N19/176—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding the unit being an image region, e.g. an object the region being a block, e.g. a macroblock
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/85—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using pre-processing or post-processing specially adapted for video compression
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/90—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using coding techniques not provided for in groups H04N19/10-H04N19/85, e.g. fractals
- H04N19/96—Tree coding, e.g. quad-tree coding
Abstract
The invention proposes a kind of panchromatic remote sensing image compression methods based on sample binary tree dictionary learning, for solving the technical issues of complicated texture information present in existing panchromatic remote sensing image compression method cannot access effective expression.Realize step are as follows: by the evaluation function of image pattern complexity, distinguish simple sample and complex samples in image；Training image and image to be compressed are pre-processed respectively to obtain sample set Y and T；Training sample binary tree and test sample binary tree are established by sample set Y and T respectively, complete the division of different complexity samples；The sample of different complexities is used to train the dictionary of different scale in training sample y-bend leaf nodes；And the sample set of different complexities sparse coding under corresponding dictionary in test sample y-bend leaf nodes, obtain coefficient matrix；Coefficient matrix obtains binary code stream by quantization encoding.The PSNR index and subjective visual evaluation of compression reconfiguration image of the present invention are high.
Description
Technical field
The invention belongs to signal processing technology fields, are related to a kind of remote sensing image compression method, and in particular to one kind is based on
The panchromatic remote sensing image compression method of sample binary tree dictionary learning can be used for the panchromatic as storage and biography of high spatial resolution
It is defeated, while not damaging the availability of data but.
Background technique
Remote sensing images are divided into panchromatic remote sensing images and multi-spectral remote sensing image according to Spectral dimension, multi-spectral remote sensing image
Spatial resolution is low, and spectral resolution is high, possesses multiple spectral coverages, and panchromatic remote sensing images spatial resolution is high, spectral resolution
Low, only one spectral coverage, in particular with the development of remote sensing technology in recent years, panchromatic remote sensing images resolution ratio is significantly mentioned
Height, even up to sub-meter grade resolution ratio.The high panchromatic remote sensing images of spatial resolution possess good visual effect, but empty
Between resolution ratio improve while bring panchromatic remote sensing image data amount increase the problem of, so panchromatic Remote Sensing Image Compression possesses
Higher researching value.
The appearance of dictionary learning is so that specific expression of the signal under dictionary has sparsity, to improve sparse volume
The effect of code, and sparse coding model is widely applied theoretical model a kind of in recent years, the purpose of the model is excessively complete
Search for one group of atom in dictionary, and the expression signal that the linear combination of this group of atom can be sparse, exactly because this sparsity
Possibility is provided for panchromatic Remote Sensing Image Compression for sparse coding model.
After compression of images, can recover original image completely can be divided into nothing for panchromatic remote sensing image compression method
The panchromatic remote sensing image compression method of damage and the panchromatic remote sensing image compression method damaged.
The redundancy of panchromatic remote sensing images scene is utilized in lossless panchromatic remote sensing image compression method, and by efficient
Coding techniques accurately reconstructs original image, can compress to big multiplying power is reached with the high panchromatic remote sensing images of spatially uniform
Purpose, without important information is lost, however, lossless compressiong can not reach very big compression ratio, therefore nothing
Damaging compress technique, there is bottlenecks in the compression of big data quantity；
The compressed panchromatic remote sensing image data of the panchromatic remote sensing image compression method damaged has information loss, so
The panchromatic remote sensing image compression method that damages allows people to measure compression ratio with picture quality, main at present based on transformation
The panchromatic remote sensing image compression method damaged has JPEG2000, traditional compression method dictionary-based learning, wherein
JPEG2000 is the compression method based on DWT transformation, and DWT transformation is to immobilize in this dictionary using DWT dictionary
, do not ensure that expression of the sample under this dictionary is sparse.Traditional compression method dictionary-based learning, passes through
Repetitive exercise makes dictionary that can adapt to specific image, and image carries out sparse coding under trained dictionary, obtains sparse
The coefficient of coding obtains compressed binary code stream to its quantization encoding, solves the simple sample of specific texture information
Do not have the problem of sparsity under fixed dictionary, but single dictionary indicates that ability is limited, the sample of texture information complexity
The sparsity also not had under the dictionary learnt, so, JPEG2000 and traditional compression side dictionary-based learning
Method cannot effectively indicate the panchromatic remote sensing images of large scene, texture structure complexity.
Summary of the invention
It is an object of the invention to be directed to the deficiency of above-mentioned prior art, propose a kind of based on sample binary tree dictionary
The panchromatic remote sensing image compression method practised, for solving complicated line present in existing panchromatic remote sensing image compression method
The technical issues of reason information cannot access effective expression.
Technical thought of the invention is: by the evaluation function of image pattern complexity, distinguish in image simple sample and
Complex samples pre-process training image and image to be compressed to obtain sample set Y and T respectively, are distinguished by sample set Y and T
Training sample binary tree and test sample binary tree are established, the division of different complexity samples, training sample y-bend leaf are completed
The sample of different complexities is used to train the dictionary of different scale on node, and different complicated in test sample y-bend leaf nodes
The sample set of degree sparse coding under corresponding dictionary, obtains coefficient matrix, and coefficient matrix obtains binary code by quantization encoding
Stream.
According to above-mentioned technical thought, the technical solution for realizing that the object of the invention is taken includes the following steps:
(1) randomly selecting multiple sizes from the panchromatic training image of one or more isNon- smooth image block, and will
Each image block pulls into column vector, then these column vectors are lined up matrix, obtains sample set
Wherein, yiFor the column vector that i-th of image block pulls into, i=1,2 ..., N, N is sample size, and n is the dimension of column vector；
(2) m layers of leaf node number and the identical sample binary tree of the number of plies are established, and obtain multiple training sample sets:
2a) define image block complexity function BC (X):
Wherein, X is to include 4 subimage block xsImage block, s be subimage block xsSerial number, std is variance function；
2b) determine the left child nodes and right child nodes of sample y-bend root vertex: by image block complexity function BC
(X) it is used as cluster feature, the non-smooth image block in step (1) is clustered, the simple sample of the sample binary tree second layer is obtained
This Y1l, sample binary tree second layer complex samples Y1rWith the criterion B for determining simple sample and complex samples1, and by Y1lMake
For the left child nodes of sample y-bend root vertex, by Y1rRight child nodes as sample y-bend root vertex, wherein n1lFor this layer of left child nodes
Number of samples, n1rFor the number of samples of this layer of right child nodes；
2c) by sample binary tree second layer complex samples Y1rIn each image block be not cut into four pieces overlappingly, obtain sample
Complex samples Y ' after the second layer cutting of this binary tree1r；
2d) determine complex samples Y '1rThe left child nodes of place node and right child nodes: by image block complexity function
BC (X) is used as cluster feature, to the complex samples Y ' after sample binary tree second layer cutting1rIn image block clustered, obtain
To sample binary tree third layer simple sample Y2l, sample binary tree third layer complex samples Y2rWith sample simple in the third layer of tree
The criterion B of this and complex samples2, and by Y2lAs Y '1rThe left child nodes of place node, by Y2rAs Y '1rPlace section
The right child nodes of point, wherein
n2lFor the number of samples of this layer of left child nodes, n2rFor the number of samples of this layer of right child nodes；
2e) by sample binary tree third layer complex samples Y2rIn each image block be not cut into four pieces overlappingly, obtain sample
Complex samples Y ' after the third layer cutting of this binary tree2r；
2f) image block complexity function BC (X) is regard as cluster feature, according to top-down sequence, to sample y-bend
The image block in complex samples after tree third layer to m-1 layers of cutting is clustered, and obtains third layer to m-1 layers of complexity
The left child nodes and right child nodes of node where sample, and not overlappingly by image block each in each layer of complex samples
Four pieces are cut into, obtains the 4th layer to m layers of left child nodes, right child nodes and this layer of simple sample and complex samples
The child nodes that criterion, these child nodes and root node and the second layer and third layer determine, form m layers of leaf node
Number sample binary tree identical with the number of plies, obtains each leaf node training sample set Y of sample binary tree1l,Y2l,...,Y(m-1)l,
Y′(m-1)r, wherein Y1l,Y2l,...,Y(m-1)lThe respectively sample binary tree second layer is to m layers of left child nodes training sample
Collection, Y '(m-1)rFor m layers of right child nodes Y(m-1)rTraining sample set after cutting；
(3) leaf node training sample set Y each to sample binary tree1l,Y2l,...,Y(m-1)l,Y′(m-1)rWord is carried out respectively
Allusion quotation training, obtains m dictionary D1,D2…Dm；
(4) according to sequence extraction size is from top to bottom, from left to right from image to be compressedImage block, and will
Each image block pulls into column vector, then these column vectors are lined up matrix, obtains sample set
Wherein, tiFor the column vector that i-th of image block pulls into, i=1,2 ..., M, M is sample size, and n is the dimension of column vector in T；
(5) m layers of leaf node number and the identical test sample binary tree of the number of plies are established, and obtain multiple test sample collections:
Branch operation 5a) is carried out to test sample collection T: sample complex in test sample collection T is less than criterion B1
Simple sample T of the sample as the test sample binary tree second layer1l, and by simple sample T1lAs test sample y-bend
The left child nodes of root vertex, while sample complex in test sample collection T is greater than criterion B1Sample as survey
The complex samples T of sample this binary tree second layer1r, and by complex samples T1rThe right side as test sample y-bend root vertex
Child nodes；
5b) by the complex samples T of the test sample binary tree second layer1rIn each image block is nonoverlapping is divided into four pieces, obtain
Complex samples T ' to after sample binary tree second layer cutting1r；
5c) to the complex samples T ' after sample binary tree second layer cutting1rCarry out branch operation: by T1rMiddle sample is complicated
Degree is less than B2Sample as test sample binary tree third layer simple sample T2l, and by simple sample T2lAs complex samples
T′1rThe left child nodes of place node, by T1rMiddle sample complex is greater than B2Sample as test sample binary tree third layer
Complex samples T2r, and by complex samples T2rAs complex samples T '1rThe right child nodes of place node；
5d) by test sample binary tree third layer complex samples T2rIn each image block be not cut into four pieces overlappingly, obtain
Complex samples T ' to after test sample binary tree third layer cutting2r；
5e) according to top-down sequence, the left child section of node where determining third layer to m-1 layers of complex samples
Point and right child nodes, and image block each in each layer of complex samples is not cut into four pieces, the 4th layer to m overlappingly
The child nodes that the left child nodes and right child nodes and root node and the second layer and third layer of layer determine form m layers of leaf segment
Point number and the identical test sample binary tree of the number of plies, obtain each leaf node test sample collection T of test sample binary tree1l,
T2l,...,T(m-1)l,T′(m-1)r, wherein T1l,T2l,...,T(m-1)lRespectively the test sample binary tree second layer is left to m layers
Child nodes test sample collection, T '(m-1)rFor m layers of right child nodes T(m-1)rTest sample collection after cutting；
(6) to the test sample collection T under each leaf node of test sample binary tree established in step (5)1l,T2l,...,
T(m-1)l,T′(m-1)r, corresponded in step (3) and carry out sparse coding, i.e. sample T under dictionaryklIn dictionary DkSparse coding is carried out,
Wherein, k=1 ..., m-1, k are sample or dictionary serial number；And sample T(m-1)rIn dictionary DmLower carry out sparse coding, obtains sparse
Code coefficient X1l...Xkl...X(m-1)l、X(m-1)r；
(7) to sparse coding coefficient X1l...Xkl...X(m-1)l、X(m-1)rQuantization encoding is carried out, binary code stream is obtained.
The present invention compared with prior art, has the advantage that
The present invention distinguishes simple sample and complex samples, and propose by defining image block complexity function
Based on the dictionary learning method of sample binary tree, realizes division and study to different complexity samples, obtain multiple and different rulers
The dictionary of degree, the dictionary of this different scale can effectively indicate the complicated local grain information of panchromatic remote sensing images, and existing
There is technology to compare, improves PSNR index and subjective visual evaluation.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention；
Fig. 2 is the test image that emulation experiment of the present invention uses；
Fig. 3 is the present invention and KSVD algorithm, JPEG2000, the emulation effect of the compression reconfiguration test image when bpp is 0.3
Fruit comparison diagram；
Fig. 4 is the present invention and KSVD algorithm, JPEG2000, the compression reconfiguration test image simulated effect when bpp is 0.5
Comparison diagram；
Fig. 5 be the present invention with KSVD algorithm, JPEG2000 bpp be 0.7 when, compression reconfiguration test image simulated effect
Comparison diagram；
Fig. 6 is that the present invention and KSVD algorithm, JPEG2000 carry out compression emulation to test image, under obtained different bpp
PSNR line chart.
Specific implementation method
Below in conjunction with the drawings and specific embodiments, present invention is further described in detail.
Referring to Fig.1, the panchromatic remote sensing image compression method based on sample binary tree dictionary learning, includes the following steps:
Step 1 randomly selects 100000 sizes from the panchromatic training image of one or more as 16 × 16
Non- smooth image block, and each image block is pulled into column vector, then these column vectors are lined up into matrix, obtain sample setWherein, yiFor the column vector that i-th of image block pulls into, i=1,2 ...,
100000；
Step 2 establishes three layers of leaf node number and the identical sample binary tree of the number of plies, and obtains three training sample sets:
2a) define image block complexity function BC (X):
Wherein, X is to include 4 subimage block xsImage block, s be subimage block xsSerial number, std is variance function；
2b) determine the left child nodes and right child nodes of sample y-bend root vertex: by image block complexity function BC
(X) it is used as cluster feature, the sample in sample set Y is clustered using FCM algorithm, it is simple to obtain the sample binary tree second layer
Sample Y1l, sample binary tree second layer complex samples Y1rWith the criterion B for determining simple sample and complex samples1, and by Y1l
As the left child nodes of sample y-bend root vertex, by Y1rRight child nodes as sample y-bend root vertex, wherein n1lFor this layer of left child's section
The number of samples of point, n1rFor the number of samples of this layer of right child nodes, FCM algorithm steps are as follows:
2b1) element in subordinated-degree matrix U is initialized as 0.5；
2b2) fixed degree of membership U, updates cluster centre ci, more new formula are as follows:
Wherein, w is control algolithm parameter flexible, and w=2, ciFor ith cluster center, i=1,2, ajFor sample set Y
In j-th of sample sample complex, n is number of samples, uijFor the element that the i-th row jth of subordinated-degree matrix U arranges, a is indicatedj
Belong to cluster centre ciProbability, and meet
2b3) calculate cost function J (U, c1,...cc), if the value of cost function less than 0.001, stops algorithm, and
Subordinated-degree matrix U ' is obtained, and jumps to step 2b5), cost function J (U, c1,...cc) are as follows:
Wherein, dij 2=| | aj-ci| | for j-th of sample in sample set Y sample complex to ith cluster center Europe
Formula distance；
2b4) update subordinated-degree matrix U, jump to step 2b2), the i-th row in subordinated-degree matrix U, jth column element more
New formula are as follows:
The sample binary tree second layer simple sample Y 2b5) is obtained according to subordinated-degree matrix U '1l, the sample binary tree second layer it is multiple
Miscellaneous sample Y1rThe criterion B of simple sample and complex samples1；
2c) by sample binary tree second layer complex samples Y1rIn each image block be not cut into four pieces overlappingly, obtain sample
Complex samples Y ' after the second layer cutting of this binary tree1r；
2d) determine complex samples Y '1rThe left child nodes of place node and right child nodes: by image block complexity function BC
(X) it is used as cluster feature, using FCM algorithm to the complex samples Y ' after sample binary tree second layer cutting1rIn image block gathered
Class obtains sample binary tree third layer simple sample Y2l, sample binary tree third layer complex samples Y2rWith sample simple in the third layer of tree
The criterion B of this and complex samples2, and by Y2lAs Y '1rThe left child nodes of place node, by Y2rAs Y '1rPlace node
Right child nodes, wherein
n2lFor the number of samples of this layer of left child nodes, n2rFor the number of samples of this layer of right child nodes；
2e) by sample binary tree third layer complex samples Y2rIn each image block is nonoverlapping is cut into four pieces, obtain sample
Complex samples Y ' after the third layer cutting of this binary tree2r；
Step 3, to three leaf node training sample set Y of sample binary tree1l、Y2lWith Y '2rBe respectively adopted KSVD algorithm into
Row dictionary training obtains three dictionary D1、D2And D3, KSVD algorithm steps are as follows:
Dictionary D, the number of iterations m, sample H 3a) are initialized, and sets maximum number of iterations:
With 1 initialization the number of iterations m, maximum number of iterations is set as 25 times, and dictionary D is carried out initially using DCT dictionary
Change, the dictionary D after being initialized(1), with sample binary tree second layer simple sample Y1l, the simple sample of sample binary tree third layer
This Y2lWith the complex samples Y ' after sample binary tree third layer cutting2rIn a sample set sample set H is initialized,
Wherein,diFor i-th of dictionary atom, p is the number of dictionary atom, and i=1 ...,
P, k are the size of atom；
3b) use OMP algorithm to sample H in dictionary D(m)Lower carry out sparse coding, obtains coefficient matrixWherein, D(m)For the dictionary in the m times iteration, m=1 ..., 25, xiFor coefficient matrix
The i-th column coefficient in X, M are the number of samples in sample set H；
Coefficient matrix X 3c) is used, to dictionary D(m)In atom diIt is updated, obtains updated dictionary D '(m), more
New formula are as follows:
Wherein,For the i-th row coefficient of coefficient matrix X；
3d) dictionary in the m+1 times iteration is updated, i.e. D(m+1)=D '(m), and update the number of iterations, i.e. m=m+
1, if the number of iterations m≤25, jump to step 3b), otherwise, stops algorithm, obtain updated word in the 25th iteration
Allusion quotation D '(25), execute step 3e)；
If 3e) step 3a) in use sample binary tree second layer simple sample Y1lSample set Y is initialized, then is walked
Rapid 3d) obtained in updated dictionary D '(25)As trained dictionary D1If step 3a) in use sample binary tree third
Layer simple sample Y2lSample set Y is initialized, then step 3d) obtained in updated dictionary D '(25)As train
Dictionary D2If step 3a) in the complex samples Y ' after the third layer cutting of sample binary tree2rSample set Y is carried out initial
Change, then step 3d) obtained in updated dictionary D '(25)As trained dictionary D3。
Step 4, from image to be compressed according to from top to bottom, from left to right sequence extract size be 16 × 16 image block, and will
Each image block pulls into column vector, then these column vectors are lined up matrix, obtains sample set
Wherein, tiFor the column vector that i-th of image block pulls into, i=1,2 ..., M, M is sample size；
Step 5 establishes three layers of leaf node number and the identical test sample binary tree of the number of plies, and obtains three test samples
Collection:
Sample complex in test sample collection T 5a) is less than criterion B1Sample as test sample binary tree
Two layers of simple sample T1l, and by simple sample T1lAs the left child nodes of test sample y-bend root vertex, simultaneously will
Sample complex is greater than criterion B in test sample collection T1Complex samples of the sample as the test sample binary tree second layer
T1r, and by complex samples T1rRight child nodes as test sample y-bend root vertex；
5b) by the complex samples T of the test sample binary tree second layer1rIn each image block is nonoverlapping is divided into four pieces, obtain
Complex samples T ' to after sample binary tree second layer cutting1r；
5c) to the complex samples T ' after sample binary tree second layer cutting1rCarry out branch operation: by T1rMiddle sample is complicated
Degree is less than B2Sample as test sample binary tree third layer simple sample T2l, and by simple sample T2lAs complex samples
T′1rThe left child nodes of place node, by T1rMiddle sample complex is greater than B2Sample as test sample binary tree third layer
Complex samples T2r, and by complex samples T2rAs complex samples T '1rThe right child nodes of place node；
5d) by test sample binary tree third layer complex samples T2rIn each image block be not cut into four pieces overlappingly, obtain
Complex samples T ' to after test sample binary tree third layer cutting2r；
Test sample collection T under step 6, each leaf node of test sample binary tree1l、T2lWith T '3r, it is all made of OMP algorithm
Respectively in dictionary D1、D2And D3Lower carry out sparse coding, obtains sparse coding coefficient X1l、X2lAnd X3r；
Step 7, to sparse coding coefficient X1l、X2lAnd X3rQuantization encoding is carried out, realizes step are as follows:
7a) to sparse coding coefficient X1l、X2lAnd X3rIn element be rounded, the sparse coding coefficient X after being quantified1l、
X2lAnd X3r；
7b) to the sparse coding coefficient X after quantization1l、X2lAnd X3rIn the corresponding index of nonzero coefficient and nonzero coefficient
Huffman coding is carried out respectively, obtains binary code stream.
Below in conjunction with emulation experiment, technical effect of the invention is further illustrated:
1) experiment condition:
This experiment image data used is that the Pan of QuickBird satellite schemes, and QuickBird satellite is in October, 2001
It succeeds in sending up, is the important sources of the remotely-sensed datas such as geographical change, forest climate.Its sensor can shoot multi light spectrum hands
With all band image, and it is to provide the earliest commercial satellite of sub-meter grade resolution ratio, full wave spatial resolution can achieve
0.65m, the spatial resolution of spectral band can achieve 2.62 meters.What the Pan figure in this experiment was shot is city image, in total
5 width images, wherein 1 width image is used to do test image, resolution ratio is 2048 × 2048, as shown in Figure 2；4 width are used to do training
Image, wherein it is 4001 × 6001 that 3 width resolution ratio, which are 2048 × 2048,1 width resolution ratio,.
Experiment simulation environment: using software MATLAB 2014R as emulation tool, CPU IntelCorei3-2310M,
Dominant frequency is 2.10GHz, and memory 4G, operating system is 8.1 professional version of Windows.
2) emulation content and interpretation of result:
Emulation 1: at 0.3,0.5,0.7bpp, KSVD algorithm, JPEG2000 and the present invention is respectively adopted to test chart
As carrying out compression emulation, the effect difference reconstructed is as shown in figure 3, figure 4 and figure 5.
Fig. 3 (a) is KSVD algorithm, the image when bpp is 0.3, after compression reconfiguration is carried out to test image；
Fig. 3 (b) is JPEG2000, the image when bpp is 0.3, after compression reconfiguration is carried out to test image；
Fig. 3 (c) is the present invention, the image when bpp is 0.3, after compression reconfiguration is carried out to test image；
Fig. 4 (a) is KSVD algorithm, the image when bpp is 0.5, after compression reconfiguration is carried out to test image；
Fig. 4 (b) is JPEG2000, the image when bpp is 0.5, after compression reconfiguration is carried out to test image；
Fig. 4 (c) is the present invention, the image when bpp is 0.5, after compression reconfiguration is carried out to test image；
Fig. 5 (a) is KSVD algorithm, the image when bpp is 0.7, after compression reconfiguration is carried out to test image；
Fig. 5 (b) is JPEG2000, the image when bpp is 0.7, after compression reconfiguration is carried out to test image；
Fig. 5 (c) is the present invention, the image when bpp is 0.7, after compression reconfiguration is carried out to test image.
It can be seen that at identical bpp from the experimental result of Fig. 3, Fig. 4 and Fig. 5, the image line that JPEG2000 is reconstructed
Information fuzzy is managed, KSVD algorithm is since expression of the single dictionary to test sample is limited, so the image of reconstruct is also undesirable, and
And there are blocking artifact phenomenons, and the image that the present invention reconstructs is relatively clear on street and these texture informations of house, so with
JPEG2000 method and KSVD algorithm are compared, and the visual effect of reconstructed image of the present invention will be got well.
Emulation 2: at different bpp, KSVD algorithm, JPEG2000 and the present invention is respectively adopted, test image is carried out
The emulation experiment of compression, experimental result is as shown in table 1 and Fig. 6:
Table 1
Fig. 6 is that the present invention and KSVD algorithm, JPEG2000 carry out compression emulation to test image, under obtained different bpp
PSNR line chart；
Table 1 is that the present invention and KSVD algorithm, JPEG2000 carry out compression emulation to test image, under obtained different bpp
PSNR table.
From the point of view of the result of Fig. 6 and table 1, with the increase of bpp, the image of the present invention and JPEG2000 compression reconfiguration
The gap of PSNR is smaller and smaller, and the present invention and the PSNR gap of the image of KSVD compression reconfiguration are increasing, is keeping identical
Bpp under, the PSNR high of the image of the PSNR ratio JPEG2000 and KSVD compression reconfiguration of the image of compression reconfiguration of the present invention,
Illustrate that the effect of compression emulation of the present invention to test image will be got well, so the present invention is more superior.
Claims (4)
1. a kind of panchromatic remote sensing image compression method based on sample binary tree dictionary learning, includes the following steps:
(1) randomly selecting multiple sizes from the panchromatic training image of one or more isNon- smooth image block, and will be every
A image block pulls into column vector, then these column vectors are lined up matrix, obtains sample set
Wherein, yiFor the column vector that i-th of image block pulls into, i=1,2 ..., N, N is sample size, and n is the dimension of column vector；
(2) m layers of leaf node number and the identical sample binary tree of the number of plies are established, and obtain multiple training sample sets:
2a) define image block complexity function BC (X):
Wherein, X is to include 4 subimage block xsImage block, s be subimage block xsSerial number, std is variance function；
2b) determine the left child nodes and right child nodes of sample y-bend root vertex: by image block complexity function BC
(X) it is used as cluster feature, the non-smooth image block in step (1) is clustered, obtains the second layer letter of sample binary tree
Single sample Y1l, sample binary tree second layer complex samples Y1rWith the criterion B for determining simple sample and complex samples1, and
By Y1lAs the left child nodes of sample y-bend root vertex, by Y1rAs the right child nodes of sample y-bend root vertex,
In, n1lFor this layer of left child's section
The number of samples of point, n1rFor the number of samples of this layer of right child nodes；
2c) by sample binary tree second layer complex samples Y1rIn each image block be not cut into four pieces overlappingly, obtain sample two
Complex samples Y ' after fork tree second layer cutting1r；
2d) determine complex samples Y '1rThe left child nodes of place node and right child nodes: image block complexity function BC (X) is made
For cluster feature, to the complex samples Y ' after sample binary tree second layer cutting1rIn image block clustered, obtain sample two
Fork tree third layer simple sample Y2l, sample binary tree third layer complex samples Y2rWith simple sample and complexity in the third layer of tree
The criterion B of sample2, and by Y2lAs Y '1rThe left child nodes of place node, by Y2rAs Y '1rThe right child of place node
Child node, whereinn2l
For the number of samples of this layer of left child nodes, n2rFor the number of samples of this layer of right child nodes；
2e) by sample binary tree third layer complex samples Y2rIn each image block be not cut into four pieces overlappingly, obtain sample two
Complex samples Y ' after fork tree third layer cutting2r；
2f) image block complexity function BC (X) is regard as cluster feature, according to top-down sequence, to sample binary tree the
The image block in complex samples after three layers to m-1 layers cutting is clustered, and obtains third layer to m-1 layers of complex samples
The left child nodes of place node and right child nodes, and by image block each in each layer of complex samples not overlappingly cutting
At four pieces, the judgement of the 4th layer to m layers of left child nodes, right child nodes and this layer of simple sample and complex samples is obtained
The child nodes that standard, these child nodes and root node and the second layer and third layer determine, form m layer leaf node number with
The identical sample binary tree of the number of plies obtains each leaf node training sample set Y of sample binary tree1l,Y2l,…,Y(m-1)l,Y
′(m-1)r, wherein Y1l,Y2l,…,Y(m-1)lThe respectively sample binary tree second layer to m layers of left child nodes training sample set,
Y′(m-1)rFor m layers of right child nodes Y(m-1)rTraining sample set after cutting；
(3) leaf node training sample set Y each to sample binary tree1l,Y2l,…,Y(m-1)l,Y′(m-1)rDictionary training is carried out respectively,
Obtain m dictionary D1,D2…Dm；
(4) according to sequence extraction size is from top to bottom, from left to right from image to be compressedImage block, and will be every
A image block pulls into column vector, then these column vectors are lined up matrix, obtains sample setIts
In, tiFor the column vector that i-th of image block pulls into, i=1,2 ..., M, M is sample size, and n is the dimension of column vector in T；
(5) m layers of leaf node number and the identical test sample binary tree of the number of plies are established, and obtain multiple test sample collections:
Branch operation 5a) is carried out to test sample collection T: sample complex in test sample collection T is less than criterion B1Sample
Simple sample T as the test sample binary tree second layer1l, and by simple sample T1lAs test sample y-bend tree root section
The left child nodes of point, while sample complex in test sample collection T is greater than criterion B1Sample as test sample
The complex samples T of the binary tree second layer1r, and by complex samples T1rRight child as test sample y-bend root vertex saves
Point；
5b) by the complex samples T of the test sample binary tree second layer1rIn each image block is nonoverlapping is divided into four pieces, obtain sample
Complex samples T ' after the second layer cutting of this binary tree1r；
5c) to the complex samples T ' after sample binary tree second layer cutting1rCarry out branch operation: by T1rMiddle sample complex is less than
B2Sample as test sample binary tree third layer simple sample T2l, and by simple sample T2lAs complex samples T '1rInstitute
In the left child nodes of node, by T1rMiddle sample complex is greater than B2Sample as test sample binary tree third layer complexity sample
This T2r, and by complex samples T2rAs complex samples T '1rThe right child nodes of place node；
5d) by test sample binary tree third layer complex samples T2rIn each image block be not cut into four pieces overlappingly, surveyed
Complex samples T ' after sample this binary tree third layer cutting2r；
5e) according to top-down sequence, the left child nodes of node where determining third layer to m-1 layers of complex samples and
Right child nodes, and image block each in each layer of complex samples is not cut into four pieces overlappingly, the 4th layer to m layers
The child nodes that left child nodes and right child nodes and root node and the second layer and third layer determine form m layers of leaf node
The number test sample binary tree identical with the number of plies, obtains each leaf node test sample collection T of test sample binary tree1l,
T2l,…,T(m-1)l,T′(m-1)r, wherein T1l,T2l,…,T(m-1)lThe respectively test sample binary tree second layer is to m layers of left child
Child node test sample collection, T '(m-1)rFor m layers of right child nodes T(m-1)rTest sample collection after cutting；
(6) to the test sample collection T under each leaf node of test sample binary tree established in step (5)1l,T2l,...,
T(m-1)l,T′(m-1)r, corresponded in step (3) and carry out sparse coding, i.e. sample T under dictionaryklIn dictionary DkSparse coding is carried out,
Wherein, k=1 ..., m-1, k are sample or dictionary serial number；And sample T(m-1)rIn dictionary DmLower carry out sparse coding, obtains sparse
Code coefficient X1l…Xkl…X(m-1)l、X(m-1)r；
(7) to sparse coding coefficient X1l…Xkl…X(m-1)l、X(m-1)rQuantization encoding is carried out, binary code stream is obtained.
2. the panchromatic remote sensing image compression method according to claim 1 based on sample binary tree dictionary learning, step
2b), step 2d) and step 2f) described in cluster, be all made of FCM clustering algorithm.
3. the panchromatic remote sensing image compression method according to claim 1 based on sample binary tree dictionary learning, step (3)
Described in leaf node training sample set Y each to sample binary tree1l,Y2l,…,Y(m-1)l,Y′(m-1)rDictionary instruction is carried out respectively
Practice, using KSVD dictionary learning algorithm.
4. the panchromatic remote sensing image compression method according to claim 1 based on sample binary tree dictionary learning, step (7)
Described in sparse coding coefficient X1l…Xkl…X(m-1)l、X(m-1)rQuantization encoding is carried out, realizes step are as follows:
7a) to sparse coding coefficient X1l...Xkl…X(m-1)l、X(m-1)rIn element be rounded, the sparse coding system after being quantified
Number X1l...Xkl...X(m-1)l、X(m-1)r；
7b) to the sparse coding coefficient X after quantization1l...Xkl...X(m-1)l、X(m-1)rIn nonzero coefficient and nonzero coefficient pair
The index answered carries out huffman coding respectively, obtains binary code stream.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710378830.1A CN107240138B (en) | 2017-05-25 | 2017-05-25 | Panchromatic remote sensing image compression method based on sample binary tree dictionary learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710378830.1A CN107240138B (en) | 2017-05-25 | 2017-05-25 | Panchromatic remote sensing image compression method based on sample binary tree dictionary learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107240138A CN107240138A (en) | 2017-10-10 |
| CN107240138B true CN107240138B (en) | 2019-07-23 |
Family
ID=59985704
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710378830.1A Active CN107240138B (en) | 2017-05-25 | 2017-05-25 | Panchromatic remote sensing image compression method based on sample binary tree dictionary learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107240138B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110019400B (en) * | 2017-12-25 | 2021-01-12 | 深圳云天励飞技术有限公司 | Data storage method, electronic device and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101609563A (en) * | 2009-07-27 | 2009-12-23 | 浙江工商大学 | A kind of construction method of binary tree of 3 D model shape features |
| CN103458242A (en) * | 2013-07-02 | 2013-12-18 | 北京京北方信息技术有限公司 | Method for compressing and uncompressing image based on color classification and cluster |
| CN103489203A (en) * | 2013-01-31 | 2014-01-01 | 清华大学 | Image coding method and system based on dictionary learning |
| CN106407978A (en) * | 2016-09-24 | 2017-02-15 | 上海大学 | Unconstrained in-video salient object detection method combined with objectness degree |
-
2017
- 2017-05-25 CN CN201710378830.1A patent/CN107240138B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101609563A (en) * | 2009-07-27 | 2009-12-23 | 浙江工商大学 | A kind of construction method of binary tree of 3 D model shape features |
| CN103489203A (en) * | 2013-01-31 | 2014-01-01 | 清华大学 | Image coding method and system based on dictionary learning |
| CN103458242A (en) * | 2013-07-02 | 2013-12-18 | 北京京北方信息技术有限公司 | Method for compressing and uncompressing image based on color classification and cluster |
| CN106407978A (en) * | 2016-09-24 | 2017-02-15 | 上海大学 | Unconstrained in-video salient object detection method combined with objectness degree |
Non-Patent Citations (3)
| Title |
|---|
| SAR Image Segmentation Based on Hierarchical Visual Semantic and Adaptive Neighborhood Multinomial Latent Model;Fang Liu 等;《IEEE Geoscience and Remote Sensing Society 》;20160401;第4287 - 4301页 |
| 基于改进K-SVD字典学习的医学图像压缩算法;酉霞 等;《信息工程·计算机科学与技术》;20170502;第74-78页 |
| 超光谱遥感图像压缩技术研究;霍承富;《中国博士学位论文全书数据库（信息科技辑）》;20130115;第I140-48页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107240138A (en) | 2017-10-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ganguli et al. | GeoGAN: A conditional GAN with reconstruction and style loss to generate standard layer of maps from satellite images | |
| CN110334765B (en) | Remote sensing image classification method based on attention mechanism multi-scale deep learning | |
| CN107657217A (en) | The fusion method of infrared and visible light video based on moving object detection | |
| CN106446936B (en) | Empty spectrum joint data based on convolutional neural networks turn the Hyperspectral data classification method of waveform diagram | |
| CN108399362A (en) | A kind of rapid pedestrian detection method and device | |
| CN108876884B (en) | Hyperspectral image reconstruction method based on non-local tensor low-rank regularization | |
| CN104331913B (en) | Polarimetric SAR Image compression method based on sparse K SVD | |
| CN108491849A (en) | Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks | |
| CN105163121A (en) | Large-compression-ratio satellite remote sensing image compression method based on deep self-encoding network | |
| CN104754340A (en) | Reconnaissance image compression method for unmanned aerial vehicle | |
| CN105513102B (en) | EO-1 hyperion compressed sensing method for reconstructing based on non-local total variation and low-rank sparse | |
| CN111860233B (en) | SAR image complex building extraction method and system based on attention network selection | |
| CN107567591A (en) | Compression and decompression on the data of radio signal | |
| CN110516716A (en) | Non-reference picture quality appraisement method based on multiple-limb similarity network | |
| CN111598786A (en) | Hyperspectral image unmixing method based on deep denoising self-coding network | |
| Byju et al. | Remote-sensing image scene classification with deep neural networks in JPEG 2000 compressed domain | |
| CN109146785A (en) | A kind of image super-resolution method based on the sparse autocoder of improvement | |
| CN104408731A (en) | Region graph and statistic similarity coding-based SAR (synthetic aperture radar) image segmentation method | |
| CN107240138B (en) | Panchromatic remote sensing image compression method based on sample binary tree dictionary learning | |
| CN109819256B (en) | Video compression sensing method based on feature sensing | |
| CN111882512B (en) | Image fusion method, device and equipment based on deep learning and storage medium | |
| CN115527027A (en) | Remote sensing image ground object segmentation method based on multi-feature fusion mechanism | |
| CN109164444A (en) | A kind of natural landscape reconstructing method based on remotely-sensed data | |
| CN104917532B (en) | Faceform's compression method | |
| CN106993240A (en) | Many video summarization methods based on sparse coding |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |