CN108171656A - Adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation - Google Patents

Adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation Download PDF

Info

Publication number
CN108171656A
CN108171656A CN201810031187.XA CN201810031187A CN108171656A CN 108171656 A CN108171656 A CN 108171656A CN 201810031187 A CN201810031187 A CN 201810031187A CN 108171656 A CN108171656 A CN 108171656A
Authority
CN
China
Prior art keywords
characteristic image
image block
image
resolution
blocks
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
Application number
CN201810031187.XA
Other languages
Chinese (zh)
Inventor
侯彪
焦李成
周康
马晶晶
马文萍
王爽
白静
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN201810031187.XA priority Critical patent/CN108171656A/en
Publication of CN108171656A publication Critical patent/CN108171656A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformation in the plane of the image
    • G06T3/40—Scaling the whole image or part thereof
    • G06T3/4053—Super resolution, i.e. output image resolution higher than sensor resolution
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching
    • 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
    • G06T2207/10004—Still image; Photographic 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

Abstract

The invention discloses a kind of adaptive Global Dictionary remote sensing images ultra-resolution methods based on rarefaction representation, and it is excessively smooth mainly to solve the problems, such as that prior art Reconstructing High information loses big and edge.Its implementation is：Its minutia is extracted respectively to 100 panel height resolution images, obtains high-resolution minutia image；Minutia image is split, obtains characteristic image set of blocks Q；Image collection is divided into K classes and training obtains wordbook D；According to the low resolution image Y of input, low resolve minutiae characteristic image is obtained by high-pass filtering；Sub- dictionary corresponding to the most like image collection of selection resolve minutiae characteristic image low with this, calculates optimal sparse coefficientObtain approximate full resolution pricture X'；The global characteristics bound term weight W of low resolve minutiae characteristic image is calculated, and final full resolution pricture is obtained with X' and W.Present invention reconstruction high-resolution edge feature is clear, and the profile details information completely of Small object is detected and classified available for Remote Sensing Target.

Description

Adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation
Technical field
The invention belongs to technical field of image processing, particularly a kind of method for being related to remote sensing images super-resolution can be applied The detection and classification of target in remote sensing image.
Background technology
Remote sensing image, that is, visible remote sensing image refers to working sensor in the remote sensing acquired in visible light wave segment limit Image, it is seen that light remote sensing is to obtain the best band of high spatial resolution, mostly measures the side such as mapping applied to aerospace photogrammetry To.Can detect and identify that the application of object is very extensive by remote sensing technology and image processing techniques, military, civilian and Play an important roll in terms of scientific research.For example, enemy's military activity and military target are identified and found in military affairs；On civilian, Including resource investigation, geologic structure research detects pasture and forest fire etc..Remote sensing technology has important and widely should With, and higher quality, high-resolution remote sensing images have its important role to research use.But the transmission of remote sensing satellite, Recycling and the expense replaced are expensive, and limited by satellite bandwidth, tend not to obtain the image of super-resolution, together When compared with natural image, remote sensing images include complicated characteristic information, similar there are a large amount of point target, Small object feature Object, and remote sensing images super-resolution more focuses on the authenticity of actual object, this just needs to take corresponding technology from low point It distinguishes to rebuild in image and obtains high-resolution image, i.e., turn to be promoted to high-resolution by certain algorithm by the image of low resolution Rate.
High-definition picture has higher picture element density, more detailed information, finer and smoother image quality.Obtain high score Resolution image, most direct method are to use high-resolution camera, however, in actual application, due to manufacture craft And the considerations of engineering cost, many occasions all will not carry out adopting for picture signal using the camera of high-resolution, super-resolution Collection.Therefore, full resolution pricture is obtained with certain application demand by super-resolution technique.
For the above feature of remote sensing image, classical ultra-resolution method has：
Method based on free hand drawing image.Mainly include traditional various interpolation methods, autoregression method and iterative backward are thrown Image method etc..This kind of method is relatively simple, wherein interpolation method and autoregression method mainly according to unknown pixel around Pixel estimate pixel value, thus both methods speed is relatively fast, but will appear at picture edge characteristic Smoothing problasm.Iterative backward projection is then using the method for the fuzzy down-sampling of the reverse solution of least square method, due to this Method is an ill inverse problem, does not know optimal solution and is easily trapped into local optimum, although Many researchers add Bound term limits last solution, but again without the inner link for making full use of characteristics of image, it is impossible to obtain the height of high quality Resolution image.
Method based on study.This kind of method needs to be largely used to trained image set, including it is low resolution and it is high-resolution Image.It is contacted by learning certain features in image, the Reconstructing High from low resolution image mainly included complete The sparse representation method of dictionary, markov random file and deep learning frame etc..F.Bill, C.Liu et al. are proposed Markov random field image super-resolution methods, by having collected many low resolution images and corresponding full resolution pricture high frequency imaging Block, and using Markov random fields come Reconstructing High.Referring to F.Bill, and C.Liu, " Markov random fields for super-resolution and texture synthesis,”in Advances in Markov Random Fields for Vision&Image Processing,A.Blake,P.Kohli, C.Rother,MS,USA: MIT press,2011.Due to the similar image block of feature, original image may differ by very remote, the full resolution pricture rebuild In the presence of the problem of dtmf distortion DTMF of part.Subsequent poplar is built super, and John Wright et al. are in order to match corresponding full resolution pricture and low resolution figure Picture utilizes the application of rarefaction representation in the picture, it is proposed that the joint dictionary method based on sparse coding.Referring to J.Yang, J.Wright,T.Huang,and Y.Ma,“Image super-resolution via sparse representation,” IEEE Trans.Image Process.,vol.19,no.11, pp.2861–2873,Nov.2010.Subsequent Dong Wei lifes et al. Propose based on sparse coding self-adapting dictionary selection method, for different characteristic image block using different dictionaries into Row coding.Participate in W.Dong, L.Zhang, G.Shi, and X.Wu, " Image deblurring and super- resolution by adaptive sparse domain selection and adaptive regularization,” IEEE Trans.Image Process.,vol.20,no.7,pp. 1838–1857,Jul.2011.With the hair of deep learning Exhibition, by the frame application to Image Super-resolution of deep learning, D.Liu, Z.W.Wang et al. propose sparse Many researchers The method of priori depth network.Referring to D. Liu, Z.W.Wang, B.H.Wen, J.C.Yang, W.Han, and T.S.Huang, “Robust single image super-resolution via deep networks with sparse prior,” IEEE Trans.Image Process.,vol. 25,no.7,pp.3194–3207.Jul.2016.For the reconstruction more prepared The method that full resolution pricture, Z.Y.Wang, Y. Yang et al. combine method and study based on single width, it is proposed that combination learning Ultra-resolution method.Referring to Z.Y.Wang, Y.Yang, Z.Wang, S.Chang, and J.Yanget, " Learning super-resolution jointly from external and internal examples,”IEEE Trans.Image Process.,vol.24,no.11, pp.4359–4371,Nov.2015。
Since remote sensing images are with greater need for the integrality, the clarity of marginal information, Small object for feature of paying attention to detail, multiple target Can identification and whole vision verisimilitude.And the local feature in the above method excessively one image block of concern, do not fill Divide the inner link for considering image and feature of overall importance, lead to the loss of minutia information and global contact details, so as to Make the full resolution pricture of reconstruction that can generate some and cross smoothing problasm, the distortion of ringing and whole vision.
Invention content
It is an object of the invention to be directed to the deficiency of above-mentioned prior art, propose a kind of adaptive complete based on rarefaction representation Office's dictionary remote sensing images ultra-resolution method to rebuild High spatial resolution remote sensing using sparse coding and global restriction item, ensures figure As the integrality of local message and global information, the quality of Reconstructing High is improved
Realizing the technical solution of the object of the invention is：Image local spy is fully excavated using self-adapting dictionary and rarefaction representation Sign；The global contact extracted between image block is used as priori, and the details such as inner link for taking full advantage of original image Information so as to fulfill the remote sensing images ultra-resolution method of adaptive Global Dictionary, is implemented as follows：
(1) to 100 panel height resolution images, its minutia is extracted by high-pass filtering respectively, obtains high-resolution details spy Levy image；
(2) minutia image is divided into the characteristic image block of N number of 7 × 7 size, obtains characteristic image set of blocks Q= {p1,p2,...,pn,...,pN, wherein piFor n-th of characteristic image block, n is the integer from 1 to N, and N is equal to 100000；
(3) image collection is divided into K class P={ s using K-means methods1,s2,...,sh,...,sK, wherein shFor h Category feature image block set, h are the integer from 1 to K, and K is the batch total in image collection, and is trained using K-SVD algorithms Obtain wordbook D={ D1,D2,...,Dh,...,DK, wherein DhFor h-th of sub- dictionary；
(4) approximate full resolution pricture is obtained：
The low resolution image Y of (4a) input, extracts its minutia by high-pass filtering, obtains low resolve minutiae characteristic pattern Picture, and the K class image collections that the low resolve minutiae characteristic image is obtained with (3) match, in wordbook D, selection is with being somebody's turn to do Sub- dictionary corresponding to the most like image collection of low resolve minutiae characteristic image, calculates optimal sparse coefficient
Wherein, α is sparse coefficient variable to be calculated, and e represents constraint term coefficient, and L is down-sampling matrix, and Β represents fuzzy Operator, | | | |2Represent two norms, | α |1Represent a norm of α；
(4b) is according to optimal sparse coefficientApproximate full resolution pricture is calculated：
(5) the global characteristics bound term weight W of low resolve minutiae characteristic image is calculated:
Low resolve minutiae characteristic image is divided into the characteristic image set of blocks Y={ y that T size is 7 × 7 by (5a)1, y2,..,yi,...,yT, wherein yiFor ith feature image block, for i from 1 to T, T is the image block of low resolve minutiae characteristic image Sum；
(5b) defines the distance between any two characteristic image blockWherein piFor ith feature figure As block, pjFor j-th of characteristic image block, j is from 1 to T, and j is not equal to i；
(5c) establishes the KD trees of a characteristic image block using k nearest neighbor algorithm according to characteristic image set of blocks Y；
(5d) is each image block y using the KD trees of established characteristic image blockiObtain M most like images Block collectionWhereinFor yiM-th of similar characteristic image block, m is from 1 to M；
(5e) calculates the weight coefficient shared by each image block:WhereinThe the one m special Levy weight coefficient shared by image block, ZmIt is normalization factor,For data compatibility formula, Middle exp () is the exponential function that natural constant e is bottom, and σ is balance parameters；
(5f) is according to weight coefficientThe global characteristics bound term weight matrix W of M rows T row is obtained, wherein m rows i-th The value of row is：
(6) approximation full resolution pricture X' and global characteristics bound term weight matrix W is utilized, it is public by solving following optimization Formula obtains final full resolution pricture：
Wherein, λ is bound term parameter,Represent final full resolution pricture, | | | |2Represent two norms,Represent two Square of norm；X represents full resolution pricture variable.
The present invention has the following advantages compared with prior art：
1st, the present invention is due to the use of self-adapting dictionary selection, the suitable word of image selection that can be larger to feature difference Allusion quotation ensure that the robustness and adaptability of this method；
2nd, dictionary selection and sparse coding are combined by the present invention, can accurately position image local feature information, are dropped Low optimization error, makes Reconstructing High more life-like；
3rd, for the present invention due to adding global restriction item, the contact between characteristic image block is even closer, is easy to small mesh Mark, the reconstruction of multiple target feature avoid minutia from losing, and edge feature crosses smoothing problasm；
Simulation result shows the method for the present invention compared with bilinear interpolation method, adaptive sparse area selecting method, rarefaction representation Method and non local autoregression method can significantly more efficient Reconstructing Highs.
Description of the drawings
Fig. 1 is the flow diagram of the present invention；
Fig. 2 is the super-resolution result figure that the present invention uses a width simple feature remote sensing images；
Fig. 3 is the super-resolution result figure that the present invention uses a width multiple small targets complex characteristic remote sensing images；
Specific embodiment
With reference to Fig. 1, specific implementation step of the invention is as follows：
Step 1: to 100 panel height resolution images, its minutia is extracted by high-pass filter respectively, obtains high-resolution Minutia image.
The high-pass filter is Gauss high-pass filter, high-pass filter size and image it is equal in magnitude, variance is 60, minutia is extracted to each sub-picture, to prepare for training dictionary below.
Step 2: according to high-resolution minutia image row and it is wide choose 7 rows, 7 row successively, by minutia image point The characteristic image block of N number of 7 × 7 size is segmented into, obtains characteristic image set of blocks Q={ p1,p2,...,pn,...,pN, wherein piFor N-th of characteristic image block, n are the integer from 1 to N, and N is equal to 100000.
Step 3: image collection is divided into K classes P using K-means methods.
(3a) does not repeat choosing method using equiprobability, at random in characteristic image set of blocks Q={ p1,p2,..., pn,...,pNIn take point set O={ o centered on K characteristic image block1,o2,...,oh,...,oK, i.e., all central points it Between different, wherein ohFor h category feature image block sets shCentral point, h is integer from 1 to K, and K is batch total；
(3b) calculates each characteristic image block p in characteristic image set of blocks Q respectivelynTo the distance of this K central point：
Wherein, pnFor n-th of characteristic image block,Represent square of two norms；
(3c) finds out distance feature image block p from the distance calculatednH-th nearest of central point oh, and pnIt moves Move set shIn, it then obtains characteristic image set of blocks and is classified as P={ s1,s2,...,sh,...,sK, shFor h classes spy Levy image block set；
(3d) is again in any sort characteristic image set of blocks shThe new central point o of middle selectionh, obtain new center point set O= {o1,o2,...,oh,...,oK}；
(3e) repeats (3b)-(3d), until center point set does not change, and according to final center point set, obtains Final characteristic image set of blocks is classified as P={ s1,s2,...,sh,...,sK}。
Step 4: according to the classification P of characteristic image set of blocks, P is trained using K-SVD algorithms, obtains wordbook D。
(4a) utilizes the obtained classification P={ s of characteristic image set of blocks1,s2,...,sh,...,sK, pass through following public affairs Formula calculates the optimal dictionary D of h category feature image block setsh：
Wherein, shFor h category feature image block sets, ahFor sparse coefficient variable, ζ represents constraint term coefficient,Table Show square of two norms, | |1Represent a norm；
(4b) is according to the optimal dictionary D per category feature image block seth, obtain wordbook D={ D1,D2,..., Dh,...,DK}。
Step 5: according to the low resolution image Y of input, its approximate full resolution pricture is obtained.
The low resolution image Y of (5a) input, extracts its minutia by high-pass filter, obtains low resolve minutiae feature Image：
In order to select and the most similar sub- dictionary D of low resolve minutiae characteristic imageh, need the low resolve minutiae characteristic pattern As being matched one by one with the K classes image collection that step 3 obtains, in wordbook D, resolve minutiae characteristic image low with this is selected most Sub- dictionary corresponding to similar image collection, and pass through following formula, calculate optimal sparse coefficient
Wherein, α is sparse coefficient variable to be calculated, and ε represents constraint term coefficient, and L is down-sampling matrix, and Β represents fuzzy Operator, | | | |2Represent two norms, | α |1Represent a norm of α；
The high-pass filter is Gauss high-pass filter, high-pass filter size and image it is equal in magnitude, variance is 60, its minutia is extracted to the low resolution image Y of input, is prepared to obtain approximate full resolution pricture below；
Step 6: calculate the global characteristics bound term weight W of low resolve minutiae characteristic image.
Low resolve minutiae characteristic image is divided into the characteristic image set of blocks Y={ y that T size is 7 × 7 by (6a)1, y2,..,yi,...,yT, wherein yiFor ith feature image block, for i from 1 to T, T is the image block of low resolve minutiae characteristic image Sum；
(6b) defines the distance between any two characteristic image blockWherein piFor ith feature image Block, pjFor j-th of characteristic image block, j is from 1 to T, and j is not equal to i；
(6c) establishes the KD trees of a characteristic image block using k nearest neighbor algorithm according to characteristic image set of blocks Y：
(6c1) is by characteristic image set of blocks Y={ y1,y2,..,yi,...,yTIn each characteristic image block yiIt is converted into Dimension is 49 vector vi, obtain the vector set V={ v of characteristic image block1,v2,..,vi,...,vT, wherein viIt is i-th The vector of characteristic image block；
(6c2) calculates the variance δ of each dimensionr, obtain variance set delta={ δ of characteristic image block dimension1,δ2,.., δr,...,δ49, wherein δrFor the variance of r-th of dimension, r is from 1 to 49；
(6c3) selects a dimension r of variance maximum, chooses the median in the dimension, and V is divided into two subtrees Vleft,Vright；
(6c4) is to first subtree Vleft(5c2) and (5c3) are repeated, until two subtree V being newly divided intoleft_1With Vleft_2Size is all 1；
(6c5) is to second subtree Vright(5c2) and (5c3) are repeated, until two subtree V being newly divided intoright_1 And Vright_2Size is all 1；
All subtrees that (6c6) is obtained according to (6c4) and (6c5) obtain the KD trees of final characteristic image block；
(6d) is each image block y using the KD trees of established characteristic image blockiObtain M most like images Block collectionWhereinFor yiM-th of similar characteristic image block, m is from 1 to M；
(6e) calculates the weight coefficient shared by each image block:WhereinM-th of feature Weight coefficient shared by image block, ZmIt is normalization factor,For data compatibility formula, exp () is the index letter that natural constant e is bottom, and σ is balance parameters, value dI, jVariance；
(6f) is according to weight coefficientThe global characteristics bound term weight matrix W of M rows T row is obtained, wherein m rows i-th The value of row is：
Step 7: approximation full resolution pricture X' and global characteristics bound term weight matrix W is utilized, by solving following optimization Formula obtains final full resolution pricture：
Wherein, λ is bound term parameter,Represent final full resolution pricture, X' is approximate full resolution pricture, and W is the overall situation Feature constraint item weight matrix, | | | |2Represent two norms,Represent square of two norms, X represents full resolution pricture variable.
The effect of the present invention can be further illustrated by following emulation：
1. emulation mode
Using the method for the present invention and bilinear interpolation method, adaptive sparse area selecting method, sparse representation method, non-office Autoregression method in portion's carries out super-resolution experiment to two width remote sensing images respectively, and from the integrality of minutia；Marginal information Clarity；Small object, multiple target can identification；The verisimilitude of whole vision evaluates the respective performance of these methods, uses peak It is worth the experimental result that signal-to-noise ratio (PSNR) property measured comparison each method rebuilds high resolution remote sensing figure.
2. emulation content and result
Emulation 1, the emulation experiment of simple feature remote sensing images.
With this method and bilinear interpolation method, adaptive sparse area selecting method, sparse representation method, it is non local from Homing method remote sensing images of differentiating low to simple feature carry out Super-resolution Reconstruction, and the results are shown in Figure 2.Wherein：
Fig. 2 (a) is low resolution remote sensing images, and the Characteristics of The Remote Sensing Images is more dull；
Fig. 2 (b) is that Super-resolution Reconstruction obtains as a result, by Fig. 2 (b) finally to be carried out with bilinear interpolation method to Fig. 2 (a) The PSNR for going out its Super-resolution Reconstruction result is 35.521DB；
Fig. 2 (c) is that Super-resolution Reconstruction obtains as a result, by Fig. 2 to be carried out with adaptive sparse area selecting method to Fig. 2 (a) (c) PSNR that can calculate its Super-resolution Reconstruction result is 35.53DB；
Fig. 2 (d) is that Super-resolution Reconstruction obtains as a result, by Fig. 2 (d) to be carried out with non local autoregression method to Fig. 2 (a) The PSNR that its Super-resolution Reconstruction result can be calculated is 35.976DB；
Fig. 2 (e) is carries out as a result, can be calculated by Fig. 2 (e) that Super-resolution Reconstruction obtains with sparse representation method to Fig. 2 (a) The PSNR of its Super-resolution Reconstruction result is 36.449DB；
Fig. 2 (f) is carries out as a result, can calculate its oversubscription by Fig. 2 (f) that Super-resolution Reconstruction obtains with this method to Fig. 2 (a) The PSNR for distinguishing reconstructed results is 37.015DB.
From Figure 2 it can be seen that the edge clear of full resolution pricture that the present invention is rebuild is thin, information completely is saved.In contrast, two-wire Property interpolation method, adaptive sparse area selecting method Super-resolution Reconstruction result loses a large amount of detailed information.Non local autoregression side Some distortions of the visual performances such as method, the color of sparse representation method Super-resolution Reconstruction result entirety, and the index for passing through PSNR More intuitively show that this method can obtain the Super-resolution Reconstruction result of high quality.
Emulation 2, the emulation experiment of multiple small targets complex characteristic remote sensing images.
With this method and bilinear interpolation method, adaptive sparse area selecting method, sparse representation method, it is non local from To multiple small targets, complicated low resolution remote sensing images carry out Super-resolution Reconstruction to homing method, and effect compares as shown in Fig. 3.Wherein:
Fig. 3 (a) is the complicated low resolution remote sensing images of multiple small targets；
Fig. 3 (b) is that bilinear interpolation method carries out Fig. 3 (a) result that Super-resolution Reconstruction obtains, and can be calculated by Fig. 3 (b) The PSNR of its Super-resolution Reconstruction result is 31.144DB；
Fig. 3 (c) be adaptive sparse area selecting method to Fig. 3 (a) carry out Super-resolution Reconstruction obtain as a result, by Fig. 3 (c) The PSNR that its Super-resolution Reconstruction result can be calculated is 31.897DB；
Fig. 3 (d) is that non local autoregression method carries out Fig. 3 (a) result that Super-resolution Reconstruction obtains, by Fig. 3 (d) finally The PSNR for going out its Super-resolution Reconstruction result is 32.062DB；
Fig. 3 (e) is that sparse representation method carries out Fig. 3 (a) result that Super-resolution Reconstruction obtains, it can be calculated by Fig. 3 (e) The PSNR of Super-resolution Reconstruction result is 32.686DB；
Fig. 3 (f) be the present invention to Fig. 3 (a) carry out Super-resolution Reconstruction obtain as a result, its super-resolution can be calculated by Fig. 3 (f) The PSNR of reconstructed results is 33.451DB.
As seen from Figure 3, smooth phenomenon, ripple, wisp was not present in the edge feature of Super-resolution Reconstruction result of the present invention Profiles are waited also than more visible.And bilinear interpolation method, adaptive sparse area selecting method, sparse representation method Super-resolution Reconstruction As a result all there are it is a large amount of excessively smooth the phenomenon that, the Super-resolution Reconstruction result of non local autoregression method then loses many detail edges Characteristic information, while also indicate that the present invention other methods that compare can obtain preferable Super-resolution Reconstruction knot by PSNR indexs Fruit.

Claims (4)

1. a kind of adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation, including：
(1) to 100 panel height resolution images, its minutia is extracted by high-pass filtering respectively, obtains high-resolution minutia figure Picture；
(2) minutia image is divided into the characteristic image block of N number of 7 × 7 size, obtains characteristic image set of blocks Q={ p1, p2,...,pn,...,pN, wherein piFor n-th of characteristic image block, n is the integer from 1 to N, and N is equal to 100000；
(3) image collection is divided into K class P={ s using K-means methods1,s2,...,sh,...,sK, wherein shFor h classes spy Image block set is levied, h is the integer from 1 to K, and K is the batch total in image collection, and trains to obtain using K-SVD algorithms Wordbook D={ D1,D2,...,Dh,...,DK, wherein DhFor h-th of sub- dictionary；
(4) approximate full resolution pricture is obtained：
The low resolution image Y of (4a) input, extracts its minutia by high-pass filtering, obtains low resolve minutiae characteristic image, And the K class image collections for obtaining the low resolve minutiae characteristic image with (3) match, and in wordbook D, select low with this point It distinguishes the sub- dictionary corresponding to the most like image collection of minutia image, calculates optimal sparse coefficient
Wherein, α is sparse coefficient variable to be calculated, and e represents constraint term coefficient, and L is down-sampling matrix, and Β represents fuzzy and calculates Son, | | | |2Represent two norms, | α |1Represent a norm of α；
(4b) is according to optimal sparse coefficientApproximate full resolution pricture is calculated：
(5) the global characteristics bound term weight W of low resolve minutiae characteristic image is calculated:
Low resolve minutiae characteristic image is divided into the characteristic image set of blocks Y={ y that T size is 7 × 7 by (5a)1,y2,.., yi,...,yT, wherein yiFor ith feature image block, for i from 1 to T, T is the image block sum of low resolve minutiae characteristic image；
(5b) defines the distance between any two characteristic image blockWherein piFor ith feature image block, pjFor j-th of characteristic image block, j is from 1 to T, and j is not equal to i；
(5c) establishes the KD trees of a characteristic image block using k nearest neighbor algorithm according to characteristic image set of blocks Y；
(5d) is each image block y using the KD trees of established characteristic image blockiObtain M most like image block collectionWhereinFor yiM-th of similar characteristic image block, m is from 1 to M；
(5e) calculates the weight coefficient shared by each image block:WhereinM-th of characteristic image Weight coefficient shared by block, ZmIt is normalization factor,For data compatibility formula, wherein exp () is the exponential function that natural constant e is bottom, and σ is balance parameters；
(5f) is according to weight coefficientThe global characteristics bound term weight matrix W of M rows T row is obtained, wherein the value that m rows i-th arrange For：
(6) approximation full resolution pricture X' and global characteristics bound term weight matrix W is utilized, by solving following optimization formula, is obtained To final full resolution pricture：
Wherein, λ is bound term parameter,Represent final full resolution pricture, | | | |2Represent two norms,Represent two norms Square；X represents full resolution pricture variable.
2. according to the method described in claim 1, image collection is wherein divided into K classes P using K-means methods in step (3) ={ s1,s2,...,sh,...,sK, it carries out as follows：
(3a) is at random in characteristic image set of blocks Q={ p1,p2,...,pn,...,pNIn take point centered on K characteristic image block Collect O={ o1,o2,...,oh,...,oK, wherein ohFor h category feature image block sets shCentral point, h is from 1 to the whole of K Number, K is batch total；
(3b) calculates each characteristic image block p in characteristic image set of blocks Q respectivelynTo the distance of this K central point；
(3c) finds out closest characteristic image block p from the distance calculatednWith h-th of central point oh, and pnIt is mobile To set shIn, it then obtains characteristic image set of blocks and is classified as P={ s1,s2,...,sh,...,sK, shFor h category feature figures As set of blocks；
(3d) is again in shThe new central point o of middle selectionh, obtain new center point set O={ o1,o2,...,oh,...,oK}；
(3e) repeats (3b)-(3d), until center point set does not change, obtains the classification of final characteristic image set of blocks For P={ s1,s2,...,sh,...,sK}。
3. it according to the method described in claim 1, wherein trains to obtain wordbook D={ D using K-SVD algorithms in step (3)1, D2,...,Dh,...,DK, it carries out as follows：
(3f) utilizes the obtained classification P={ s of characteristic image set of blocks1,s2,...,sh,...,sK, calculate h category feature figures As the optimal dictionary D of set of blocksh：
Wherein, shFor h category feature image block sets, αhFor sparse coefficient variable, ζ represents constraint term coefficient,Represent two models Several squares, | |1Represent a norm；
(3g) is according to optimal dictionary Dh, obtain wordbook D={ D1,D2,...,Dh,...,DK}。
4. according to the method described in claim 1, k nearest neighbor algorithm is wherein used according to characteristic image set of blocks Y in step (5c) The KD trees of a characteristic image block are established, are carried out as follows：
(5c1) is by characteristic image set of blocks Y={ y1,y2,..,yi,...,yTIn each characteristic image block yiIt is converted into dimension For 49 vector vi, obtain the vector set V={ v of characteristic image block1,v2,..,vi,...,vT, wherein viFor ith feature The vector of image block；
(5c2) calculates the variance δ of each dimensionr, obtain variance set delta={ δ of characteristic image block dimension1,δ2,.., δr,...,δ49, wherein δrFor the variance of r-th of dimension, r is from 1 to 49；
(5c3) selects a dimension r of variance maximum, chooses the median in the dimension, and V is divided into two subtree Vleft, Vright；
(5c4) is to first subtree Vleft(5c2) and (5c3) are repeated, until two subtree V being newly divided intoleft_1And Vleft_2 Size is all 1；
(5c5) is to second subtree Vright(5c2) and (5c3) are repeated, until two subtree V being newly divided intoright_1With Vright_2Size is all 1；
All subtrees that (5c6) is obtained according to (5c4) and (5c5) obtain the KD trees of final characteristic image block.
CN201810031187.XA 2018-01-12 2018-01-12 Adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation Pending CN108171656A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810031187.XA CN108171656A (en) 2018-01-12 2018-01-12 Adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810031187.XA CN108171656A (en) 2018-01-12 2018-01-12 Adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation

Publications (1)

Publication Number Publication Date
CN108171656A true CN108171656A (en) 2018-06-15

Family

ID=62514645

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810031187.XA Pending CN108171656A (en) 2018-01-12 2018-01-12 Adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation

Country Status (1)

Country Link
CN (1) CN108171656A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109035142A (en) * 2018-07-16 2018-12-18 西安交通大学 A kind of satellite image ultra-resolution method fighting network integration Aerial Images priori
CN110503631A (en) * 2019-07-24 2019-11-26 山东师范大学 A kind of method for detecting change of remote sensing image
CN111612691A (en) * 2020-04-17 2020-09-01 重庆大学 Image super-resolution processing improvement method based on sparse representation
CN111814814A (en) * 2019-04-10 2020-10-23 四川大学 Single-stage target detection method based on image super-resolution network
CN112419333A (en) * 2020-11-17 2021-02-26 武汉大学 Remote sensing image self-adaptive feature selection segmentation method and system

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102156875A (en) * 2011-03-25 2011-08-17 西安电子科技大学 Image super-resolution reconstruction method based on multitask KSVD (K singular value decomposition) dictionary learning
CN103093445A (en) * 2013-01-17 2013-05-08 西安电子科技大学 Unified feature space image super-resolution reconstruction method based on joint sparse constraint
CN107067367A (en) * 2016-09-08 2017-08-18 南京工程学院 A kind of Image Super-resolution Reconstruction processing method

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102156875A (en) * 2011-03-25 2011-08-17 西安电子科技大学 Image super-resolution reconstruction method based on multitask KSVD (K singular value decomposition) dictionary learning
CN103093445A (en) * 2013-01-17 2013-05-08 西安电子科技大学 Unified feature space image super-resolution reconstruction method based on joint sparse constraint
CN107067367A (en) * 2016-09-08 2017-08-18 南京工程学院 A kind of Image Super-resolution Reconstruction processing method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
BIAO HOU 等: "Adaptive Super-Resolution for Remote Sensing Images Based on Sparse Representation With Global Joint Dictionary Model", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》 *
WEISHENG DONG 等: "Image Deblurring and Super-Resolution by Adaptive Sparse Domain Selection and Adaptive Regularization", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 *
佚名: "K近邻算法实现之KD树", 《BLOG.SINA.COM.CN/S/BLOG_81963F3C0102Y6GH.HTML》 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109035142A (en) * 2018-07-16 2018-12-18 西安交通大学 A kind of satellite image ultra-resolution method fighting network integration Aerial Images priori
CN109035142B (en) * 2018-07-16 2020-06-19 西安交通大学 Satellite image super-resolution method combining countermeasure network with aerial image prior
CN111814814A (en) * 2019-04-10 2020-10-23 四川大学 Single-stage target detection method based on image super-resolution network
CN111814814B (en) * 2019-04-10 2022-04-12 四川大学 Single-stage target detection method based on image super-resolution network
CN110503631A (en) * 2019-07-24 2019-11-26 山东师范大学 A kind of method for detecting change of remote sensing image
CN111612691A (en) * 2020-04-17 2020-09-01 重庆大学 Image super-resolution processing improvement method based on sparse representation
CN112419333A (en) * 2020-11-17 2021-02-26 武汉大学 Remote sensing image self-adaptive feature selection segmentation method and system
CN112419333B (en) * 2020-11-17 2022-04-29 武汉大学 Remote sensing image self-adaptive feature selection segmentation method and system

Similar Documents

Publication Publication Date Title
CN110191299B (en) Multi-frame interpolation method based on convolutional neural network
Shao et al. Remote sensing image fusion with deep convolutional neural network
CN108171656A (en) Adaptive Global Dictionary remote sensing images ultra-resolution method based on rarefaction representation
CN112634137B (en) Hyperspectral and panchromatic image fusion method for extracting multiscale spatial spectrum features based on AE
CN108596108B (en) Aerial remote sensing image change detection method based on triple semantic relation learning
CN106056155B (en) Superpixel segmentation method based on boundary information fusion
CN105740799B (en) Classification of hyperspectral remote sensing image method and system based on the selection of three-dimensional Gabor characteristic
Zhang et al. CCR: Clustering and collaborative representation for fast single image super-resolution
CN108549891A (en) Multi-scale diffusion well-marked target detection method based on background Yu target priori
CN109360156A (en) Single image rain removing method based on the image block for generating confrontation network
CN110084108A (en) Pedestrian re-identification system and method based on GAN neural network
CN105718882B (en) A kind of resolution ratio self-adaptive feature extraction and the pedestrian's recognition methods again merged
CN110473142B (en) Single image super-resolution reconstruction method based on deep learning
CN106846322B (en) The SAR image segmentation method learnt based on curve wave filter and convolutional coding structure
CN108510504A (en) Image partition method and device
Asokan et al. Machine learning based image processing techniques for satellite image analysis-a survey
CN105513033B (en) A kind of super resolution ratio reconstruction method that non local joint sparse indicates
CN103824272A (en) Face super-resolution reconstruction method based on K-neighboring re-recognition
CN111223049B (en) Remote sensing image variation fusion method based on structure-texture decomposition
Beaulieu et al. Deep image-to-image transfer applied to resolution enhancement of sentinel-2 images
CN106960182A (en) A kind of pedestrian integrated based on multiple features recognition methods again
CN108288256A (en) A kind of multispectral mosaic image restored method
CN111680579B (en) Remote sensing image classification method for self-adaptive weight multi-view measurement learning
CN104408697A (en) Image super-resolution reconstruction method based on genetic algorithm and regular prior model
CN116977674A (en) Image matching method, related device, storage medium and program product

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20180615

WD01 Invention patent application deemed withdrawn after publication