CN107316306B - A kind of diameter radar image fast partition method based on Markov model - Google Patents
A kind of diameter radar image fast partition method based on Markov model Download PDFInfo
- Publication number
- CN107316306B CN107316306B CN201710485298.3A CN201710485298A CN107316306B CN 107316306 B CN107316306 B CN 107316306B CN 201710485298 A CN201710485298 A CN 201710485298A CN 107316306 B CN107316306 B CN 107316306B
- Authority
- CN
- China
- Prior art keywords
- pixel
- image
- segmentation
- node
- processed
- 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
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10044—Radar image
Abstract
The invention belongs to technical field of image segmentation, and in particular to a kind of diameter radar image fast partition method based on Markov model.For the present invention based on Markov model, the comprehensive study influence of super-pixel and neighborhood system to splitting speed proposes a kind of diameter radar image fast partition method.Super-pixel pretreatment is carried out to radar image first, and generates the neighborhood system of rationally description positional relationship, the segmentation result for then carrying out the image segmentation based on MRF model and post-processing to the end accordingly with this.Experiments have shown that this method is greatly improved image segmentation speed and farthest remains segmentation quality.
Description
Technical field
The invention belongs to technical field of image segmentation, and in particular to a kind of synthetic aperture radar based on Markov model
Image fast segmentation method.
Background technique
Synthetic aperture radar (Synthetic Aperture Radar, hereinafter referred to as SAR) can be under full weather conditions
Whole day provides high-definition picture.The main method of current SAR image segmentation is based on model-driven, by segmentation problem
It constructs mathematical model and realizes SAR image segmentation, tend to obtain ideal segmentation result.But with SAR imaging technique
Development, magnanimity, high-resolution radar image proposes harsh requirement to the efficiency of segmentation interpretation work, so how to mention
High SAR image splitting speed is the emphasis of current SAR Research on Method of Image Segmentation.
Image segmentation algorithm based on model-driven models segmentation problem based on the priori knowledge of image,
Usually optimization problem is converted by segmentation problem.The method for so improving splitting speed is just mainly Optimized Iterative algorithm and reduction
Two aspect of destination number to be processed.Optimized Iterative algorithm is started with from interative computation process, to algorithm itself or is entirely flowed
Cheng Jinhang simplifies, to achieve the purpose that acceleration, it is however generally that difficulty is larger, and applicability is not high, is easy to appear significant problem.Subtract
Processing target of waiting a little while,please quantity is the most straightforward approach that algorithm accelerates, and not change algorithm core, realization is simple, at
For most common method in algorithm acceleration.After determining to reduce target object quantity to be processed to realize that algorithm accelerates, how
Effectively, reasonably realize that object reduction just becomes main problem.
The neighborhood system of image, that is, in image each neighborhood of pixels relationship set, it is however generally that show as grid
Change.The neighborhood system of gridding is also beneficial to the realization of great amount of images Processing Algorithm, is based especially on the image point of model-driven
Cut algorithm.Operation is carried out to image pixel and reduces algorithm object to be processed, it is however generally that is had and is extracted representative object, concludes image
The methods of prior information, that is to say, that how to realize that the information for describing original image using less information is the key that accelerate.
And the process, it is usually associated with breaking for original gridding system, and then influence the realization of algorithm.Reasonable neighborhood system
It establishes, while so that algorithm accelerates, possesses the segmentation effect of former algorithm to greatest extent.Conversely, failing to fully describe image
Information then makes segmentation effect serious situations such as accidentally dividing occur.Itself for algorithm, new neighborhood system how is established,
The neighborhood relationships for describing every an object are the problems that algorithm realizes part difficulty maximum.
Summary of the invention
The purpose of the invention is to overcome the above-mentioned shortcoming for SAR image segmentation, to reach SAR image segmentation
Speed faster, provides a kind of SAR image fast partition method based on Markov model.Markov model, also referred to as horse
Er Kefu random field models (Markov Random Field Model, hereinafter referred to as MRF model), refer to meet orthotropicity and
The random field models of Markov property, it combines Bayes theoretical, provides the knob that uncertainty description is contacted with priori knowledge
Band, and observed image is utilized, the objective function of segmentation problem is determined according to the optiaml ciriterion in statistical decision and estimation theory, is asked
Solution meets the maximum possible distribution of these conditions, to convert optimization problem for segmentation problem.MRF model can be by element
Spatial relationship combine closely, can sufficiently react the potential structure of image and the randomness of image；By strict
Mathematical derivation and realization, the available segmentation result to work well.For the SAR image partitioning algorithm based on MRF model
Accelerate, with the super-pixel algorithm based on Clustering, as the method for reducing object to be processed, and for super-pixel pretreatment knot
The segmentation that fruit generates new neighborhood system and combines MRF model realization last.
The first step utilizes simple linear iteration cluster (Simple Linear Iterative Clustering, below letter
Claiming SLIC) super-pixel algorithm pre-processes SAR image, and suitable segmentation step-length is selected, guarantees that pre-processed results do not occur
Less divided and serious over-segmentation phenomenon, this method are traditional technology, and the present invention directly uses its conventional method, herein just no longer
It repeats, method obtains segmentation step-length through this process.
SLIC algorithm used herein is realized fast using color similarity and space length relationship with local iteration's cluster
The image segmentation of speed.
Second step generates pseudo- gridding neighborhood system to SLIC pre-processed results.
Pseudo- gridding (Pseudo-meshing) just assumes that the node letter of all super-pixel block of SLIC pre-processed results
Breath, is arranged according to gridding.Firstly, according to image slices vegetarian refreshments information in super-pixel segmentation result, i.e., each pixel
Transverse and longitudinal coordinate rkAnd ckAnd the gray value g of pixelk, the nodal information of super-pixel block is described:
D (i, j)=max (| r (i)-r (j) |, | c (i)-c (j) |)
Wherein super-pixel block nodal information, including transverse and longitudinal coordinate r (i) and c (i) are indicated with mean value in block；For between node
Relationship, it is contemplated that processing speed reason, using maximum distance rather than Euclidean distance indicates euclidean distance between node pair D (i, j).It is obtaining
After each super-pixel block nodal information, according to gridding arrangement mode, the position of each super-pixel block node within a grid is calculated
Place obtains image to be processed.Since the image to be processed is much smaller compared to original image, therefore referred to as " small figure ".To institute
There is node location calculating to finish, the blank spot occurred in small figure is handled, i.e., by its 2 rank neighborhood all the points gray value mean value
Assign the blank spot.In addition, judging the repetition drop point situation in node location calculating process: if repeating two sections of drop point
The gray scale difference value of point is in an identification section, then it is assumed that the two nodes belong to same segmentation classification, otherwise on the contrary.Identify section with
Image grayscale maximum difference is determined with total pixel block number is divided.
Third step generates down-sampling neighborhood system to SLIC pre-processed results.
Similar with pseudo- gridding neighborhood system, down-sampling neighborhood system also assumes that the super-pixel in SLIC pre-processed results
Block node is arranged according to gridding.Using super-pixel block average area size and image entirety size as foundation, selection
Suitable sampling step length carries out down-sampling to SLIC pre-processed results, directly obtains the small figure to be processed of no blank spot.To SLIC
The nodal information omitted in pre-processed results carries out judgement processing, and missed point and its nearest central point is made to compare judgement:
Wherein pixel block size is Ss, step-length is set as J between cog region, and identification section gray scale is set as G, nearest central point
Distance is lm, the gray value of nearest central point is gm, the gray value for being somebody's turn to do " omission " point is gt.According to omit size of node and
Gray value and its at a distance from nearest node and gray scale difference value, along with the identification section entirely divided and step-length, make point
Cut the judgement of classification.
Two kinds of neighborhood systems are brought into the partitioning algorithm based on MRF model respectively and are split operation by the 4th step.
The MRF model used herein use limited mixed normal model (Finite Gauss Matured Model, with
Lower abbreviation FGMM) description Characteristic Field, Potts model description label field, Iterative conditional modes (Iterative
Condition Model, hereinafter referred to as ICM) algorithm is as partitioning algorithm.
FGMM model is the part for having similar features region to each of image, establishes mixed Gauss model.This side
Method can be regarded as based on regional area and based on the integrated processes of Gaussian function, applicability, in terms of accomplish
It is balanced.The value range of Potts model is multiple values.But as Ising model, which also provides each position only
There are two types of values.Assuming that certain point pixel is x in imagek, then the conditional probability in Potts model can indicate are as follows:
Wherein,It is the set of its neighborhood territory pixel of i pixel, α and β are corresponding potential parameters.The part of Potts model is general
Rate are as follows:
ICM algorithm mainly updates image pixel point by point achievees the purpose that image segmentation.Assuming that image y={ y1,
y2,…,ynEach pixel yiIt is independent from each other under conditions of given initial segmentation result x, and yiCondition point about x
Cloth only depends on the label x of the pixeli, i.e.,
f(yi| x)=f (yi|xi)
Therefore, y can be expressed as about the condition distribution of x
Then last segmentation result can be expressed as
5th step, post-processes segmentation result.Pair when small figure segmentation result is generated according to neighborhood system before
It should be related to, compare back SLIC pre-processed results, obtain the segmentation result in the case of two kinds of neighborhoods.
6th step is assessed the different segmentation situation of two kinds of neighborhood systems and is accepted or rejected.
To the assessment that segmentation result exercises supervision, use with Dice Ratio and SAMainly to assess parameter.Supervision assessment
It is that the true value figure by segmentation result and manually marked compares assessment.Wherein, Dice Ratio is the correctness by dividing
Precision and sensibility Recall comprehensive assessment.To judge the secondary segmentation result, following several parameters: kidney-Yang are introduced
Property (true positive is abbreviated as tp) i.e. positive region in true destination number, false positive (false positive, brief note
It is true in true negative (true negative, be abbreviated as tn) i.e. negative areas for the decoy quantity in fp) i.e. positive region
Destination number, the decoy quantity in false negative (false negative, be abbreviated as fn) i.e. negative areas.Precision as a result,
Precision and sensibility Recall is defined as:
Segmentation precision SAFor
Wherein, TkIndicate region RkTrue value.
Last segmentation result is chosen according to the assessment situation of aforementioned four assessment parameter.
The invention has the benefit that compared with the prior art, image segmentation speed is greatly improved in method of the invention
It spends and farthest remains segmentation quality.
Detailed description of the invention
Fig. 1 is gridding neighborhood system structure in MRF model；
Fig. 2 is SLIC algorithm principle figure；
Fig. 3 is Dice Ratio supervision assessment schematic diagram.
Specific embodiment
Summary is described in detail technical solution of the present invention, and details are not described herein.
Claims (1)
1. a kind of diameter radar image fast partition method based on Markov model, which is characterized in that including following
Step:
S1, SAR image is pre-processed:
SAR image is pre-processed using simple linear iteration cluster super-pixel algorithm, selects suitable segmentation step-length；
S2, according to the pre-processed results of step S1, generate pseudo- gridding neighborhood system:
The nodal information of all super-pixel block of S21, hypothesis pre-processed results, is to arrange according to gridding, to super-pixel block
Nodal information modeling it is as follows:
D (i, j)=max (| r (i)-r (j) |, | c (i)-c (j) |)
Wherein, super-pixel block nodal information, r are indicated with mean value in blockkAnd ckFor the transverse and longitudinal coordinate of each pixel, gk pixel
Gray value, D (i, j) indicates that euclidean distance between node pair, i, j are node serial number, and n is pixel value in block of pixels, and n (i) is the picture of node i
Pixel value in plain block, S are block of pixels area, and S (i) is the block of pixels area of node i；
S22, it after obtaining each super-pixel block nodal information, according to gridding arrangement mode, calculates each super-pixel block node and exists
Where position in grid, image to be processed is obtained；
S23, after obtaining all node locations according to step S22, at the blank spot occurred in the image to be processed of acquisition
Reason, i.e., assign 2 rank neighborhood all the points gray value mean value of blank spot to the blank spot；
S24, identification section is obtained according to image grayscale maximum difference and the total pixel block number of segmentation；
S3, according to the pre-processed results of step S1, generate down-sampling neighborhood system:
S31, using super-pixel block average area size and image entirety size as foundation, suitable sampling step length is selected, to SLIC
Pre-processed results carry out down-sampling, directly obtain the small figure to be processed of no blank spot；
S32, judgement processing is carried out to the nodal information omitted in the pre-processed results of step S1, makes missed point and its nearest center
Point compares judgement:
Wherein pixel block size is Ss, step-length is set as J between cog region, and identification section gray scale is set as G, the distance of nearest central point
For lm, the gray value of nearest central point is gm, the gray value of the missed point is gt；
S33, according to the step S32 omission size of node obtained and gray value and its at a distance from nearest node and gray scale
Difference makes the judgement of segmentation classification along with the identification section entirely divided and step-length；
S4, it two kinds of neighborhood systems is brought into respectively in the partitioning algorithm based on Markov model is split operation, it is described
Markov model uses the Characteristic Field of limited mixed normal model description, the label field of Potts model description, and uses and change
For condition pattern algorithm as partitioning algorithm, specifically:
Limited mixed normal provides that only there are two types of values for each position, it is assumed that certain point pixel is x in imagek, then Potts model
In conditional probability be expressed as:
Wherein, xiIt is the pixel of node i, xjIt is the pixel of node j,It is the set of node i neighborhood territory pixel, NiIt is the neighbour of node i
Domain, α and β are corresponding potential parameters；
The local probability of Potts model are as follows:
Assuming that image y={ y1,y2,…,ynEach pixel yiIt is mutually indepedent under conditions of given initial segmentation result x
, and yiThe label x of the pixel is only depended on about the condition distribution of xi, it may be assumed that
f(yi| x)=f (yi|xi)
Therefore, y can be indicated about the condition distribution of x are as follows:
Segmentation result indicates are as follows:
S5, segmentation result is post-processed:
Corresponding relationship when image segmentation result to be processed is generated according to neighborhood system in step S3, compares back step S1 and locates in advance
Reason is as a result, obtain the segmentation result in the case of two kinds of neighborhoods；
S6, the assessment different segmentation situation of two kinds of neighborhood systems simultaneously choose last segmentation result:
To the assessment that segmentation result exercises supervision, use with Dice Ratio and SATo assess parameter, wherein Dice Ratio is
From the correctness Precision divided and sensibility Recall comprehensive assessment, to judge the secondary segmentation result, introduce following
Several parameters: true positives, i.e., the true destination number in positive region are abbreviated as tp, false positive, i.e. decoy in positive region
Quantity, is abbreviated as fp, true negative, i.e., the true destination number in negative areas is abbreviated as tn, false negative, i.e., in negative areas
Decoy quantity, is abbreviated as fn；Precision Precision and sensibility Recall be just as a result, is defined as:
Segmentation precision SAAre as follows:
Wherein, TkIndicate region RkTrue value.
The assessment situation for assessing parameter according to the four of above-mentioned introducing chooses last segmentation result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710485298.3A CN107316306B (en) | 2017-06-23 | 2017-06-23 | A kind of diameter radar image fast partition method based on Markov model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710485298.3A CN107316306B (en) | 2017-06-23 | 2017-06-23 | A kind of diameter radar image fast partition method based on Markov model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107316306A CN107316306A (en) | 2017-11-03 |
| CN107316306B true CN107316306B (en) | 2019-07-05 |
Family
ID=60180778
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710485298.3A Active CN107316306B (en) | 2017-06-23 | 2017-06-23 | A kind of diameter radar image fast partition method based on Markov model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107316306B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107909597B (en) * | 2017-11-14 | 2021-05-14 | 西安建筑科技大学 | Multi-scale MRF model image segmentation method with edge preservation function |
| CN107895373B (en) * | 2017-11-14 | 2021-04-27 | 西安建筑科技大学 | Image segmentation method based on local region consistency manifold constraint MRF model |
| CN108562897B (en) * | 2018-01-26 | 2022-01-11 | 桂林电子科技大学 | Structure sparse imaging method and device of MIMO through-wall radar |
| CN111583266B (en) * | 2020-05-08 | 2021-09-24 | 清华大学 | Self-adaptive synthetic aperture radar image super-pixel segmentation method based on Fermat vector |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101350099A (en) * | 2008-09-12 | 2009-01-21 | 西安电子科技大学 | Method for segmenting image based on wavelet domain concealed Markov tree model |
| CN102402685A (en) * | 2011-11-10 | 2012-04-04 | 西安电子科技大学 | Method for segmenting three Markov field SAR image based on Gabor characteristic |
| CN103606164A (en) * | 2013-12-03 | 2014-02-26 | 西安电子科技大学 | SAR image segmentation method based on high-dimensional triple Markov field |
| CN103729843A (en) * | 2013-12-20 | 2014-04-16 | 电子科技大学 | Medical image segmentation method based on Markov |
| US9146941B2 (en) * | 2012-08-03 | 2015-09-29 | Adobe Systems Incorporated | Image tag pair graph for image annotation |
-
2017
- 2017-06-23 CN CN201710485298.3A patent/CN107316306B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101350099A (en) * | 2008-09-12 | 2009-01-21 | 西安电子科技大学 | Method for segmenting image based on wavelet domain concealed Markov tree model |
| CN102402685A (en) * | 2011-11-10 | 2012-04-04 | 西安电子科技大学 | Method for segmenting three Markov field SAR image based on Gabor characteristic |
| US9146941B2 (en) * | 2012-08-03 | 2015-09-29 | Adobe Systems Incorporated | Image tag pair graph for image annotation |
| CN103606164A (en) * | 2013-12-03 | 2014-02-26 | 西安电子科技大学 | SAR image segmentation method based on high-dimensional triple Markov field |
| CN103729843A (en) * | 2013-12-20 | 2014-04-16 | 电子科技大学 | Medical image segmentation method based on Markov |
Non-Patent Citations (2)
| Title |
|---|
| FAST TOPOLOGY PRESERVING POLSAR IMAGE SUPERPIXEL SEGMENTATION;Guo, Weiwei等;《2016 IEEE INTERNATIONAL GEOSCIENCE AND REMOTE SENSING SYMPOSIUM (IGARSS)》;20160715;第6894-6897页 * |
| 基于空域MRF模型的SAR图像分割方法研究;梁博;《中国优秀硕士学位论文全文数据库 信息科技辑》;20180215;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107316306A (en) | 2017-11-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107316306B (en) | A kind of diameter radar image fast partition method based on Markov model | |
| CN105741279B (en) | The quick image partition method for suppressing fuzzy clustering based on rough set | |
| CN108038445B (en) | SAR automatic target identification method based on multi-view deep learning framework | |
| CN108711141B (en) | Motion blurred image blind restoration method using improved generation type countermeasure network | |
| CN103810473B (en) | A kind of target identification method of human object based on HMM | |
| CN109035172B (en) | Non-local mean ultrasonic image denoising method based on deep learning | |
| CN111723780B (en) | Directional migration method and system of cross-domain data based on high-resolution remote sensing image | |
| CN111028327A (en) | Three-dimensional point cloud processing method, device and equipment | |
| CN101520894A (en) | Method for extracting significant object based on region significance | |
| CN110826428A (en) | Ship detection method in high-speed SAR image | |
| CN104992403B (en) | Hybrid operator image redirection method based on visual similarity measurement | |
| CN107680116A (en) | A kind of method for monitoring moving object in video sequences | |
| CN114022729A (en) | Heterogeneous image matching positioning method and system based on twin network and supervised training | |
| CN109299668A (en) | A kind of hyperspectral image classification method based on Active Learning and clustering | |
| CN112699717A (en) | SAR image generation method and generation device based on GAN network | |
| CN104156929A (en) | Infrared weak and small target background inhibiting method and device on basis of global filtering | |
| CN109300137A (en) | A kind of two type fuzzy clustering magnetic resonance brain image segmentation method of more curved surface estimation intervals | |
| Wu et al. | Joint convolutional neural network for small-scale ship classification in SAR images | |
| CN111091147B (en) | Image classification method, device and equipment | |
| CN107766792B (en) | Remote sensing image ship target identification method | |
| CN103871089B (en) | Image superpixel meshing method based on fusion | |
| CN115223056A (en) | Multi-scale feature enhancement-based optical remote sensing image ship target detection method | |
| CN109003247B (en) | Method for removing color image mixed noise | |
| Irie et al. | Robust learning for deep monocular depth estimation | |
| CN112215199A (en) | SAR image ship detection method based on multi-receptive-field and dense feature aggregation network |
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 |