CN106803070A - A kind of port area Ship Target change detecting method based on remote sensing images - Google Patents

Abstract

The invention discloses a kind of port area Ship Target change detecting method based on remote sensing images, belong to object detection field；Two remote sensing images of same harbour difference phase are obtained first；Secondly, by two remote sensing image registrations of same port area different time collection；Then, line mask computing is entered to two images using template image, shields land area, obtain water area, to water area image sea color homogenization, the land area to shielding is filled using sea region average；Image after two width registration is poor, split to making the image after difference；Finally, the characteristics of image of section is extracted, false-alarm rejecting is carried out, the change naval vessel section in section is confirmed, that is, obtains Ship Target position and the corresponding information of two width Remote Sensing Imagery Changes；The method can quickly and accurately exclude influence of the Sea background of cloud and change to final result, reduce false alarm rate, and energy completely, accurately be split and detect the Ship Target of change, improve verification and measurement ratio.

Description

A kind of port area Ship Target change detecting method based on remote sensing images

Technical field

The invention belongs to technical field of image detection, and in particular to a kind of port area Ship Target change of remote sensing images Detection method.

Background technology

With aerospace technology, sensor technology, computer technology, mode identification technology related discipline swift and violent hair Exhibition, space remote sensing technology achieves made rapid progress, and corresponding technical support and reason are provided to Automatic Targets recognition capability By basis.Multitemporal Remote Sensing Images port area Ship Target change detection and analysis technology is the pass that China is badly in need of solving at this stage One of key technology.China territorial waters is wide, and marine safety is the pith in national defence, carries out the change detection of harbour Ship Target Research is significant：At civilian aspect, be conducive to the monitoring of the aspects such as berth allocation, naval vessel entering and leaving port to specific harbour And management, it is China's sea-freight monitoring management scheduling provided auxiliary information, increase the efficiency of port transport；In military aspect, Defend territorial integrity and safeguard that unification of the motherland is the significant task that China faces, concern harbour can in time be learned by the technology The situation of change of interior Ship Target, analyzes possibility military affairs behavior and the purpose of other side, to battlefield dynamic sensing and strategy in advance Being intended to analysis has vital meaning.

Initial people are changed detection to the Ship Target in harbour chart picture, and mainly sentencing figure person by specialty carries out hand Work is marked and drawed, and the different figure persons that sentence carry out detecting that the final result for drawing is often different for same image, even same Name sentences figure person, also always not completely the same to the plotting result of same image in different time, while with image acquiring way Increase, if the view data of magnanimity by manually marked and drawed cannot meet first wartime intelligence rapidly update requirement, It is then even more acid test for sentencing figure person, therefore realizes that automatic change detection is that the following technology develops with computer Inexorable trend.

Therefore, a kind of effective Multitemporal Remote Sensing Images port area Ship Target change detection techniques just become non- Often it is necessary.The method for mainly using pattern-recognition for solving the technological approaches that the problem is typically used at present is processed, Generally include to carry out artwork the steps such as image registration, extra large land segmentation, the poor, image segmentation of image work, false-alarm rejecting.But, by In remote sensing images than larger, cause image registration step computationally intensive, in order to speed up processing typically sacrifices operational precision, lead The presence of registration error is caused, the precision of latter results has been had influence on；Because remote sensing images content is complicated, gray scale is frequently not uniform Distribution, cause extra large land segmentation step often cannot entirely accurate will the two complete parttion out, these areas being partitioned into by mistake Domain can finally increase the false drop rate of system；Because shooting time is different, weather conditions are different in remote sensing images, Same Scene is not May be widely different with the gray value of the corresponding pixel of the remote sensing images of phase, it is likely to be obtained perhaps in the step for image is differed from Many falseness variation targets.

The content of the invention

In view of this, it is an object of the invention to provide a kind of remote sensing images port based on pattern-recognition Yu image processing techniques Mouth region domain Ship Target change detecting method, can solve the problem that current port area Ship Target change detecting method registration is inaccurate Really, the problems such as false alarm rate is high, obtains change Detection accuracy higher.

A kind of port area Ship Target change detecting method based on remote sensing images, comprises the following steps：

S1：For pay close attention to harbour, obtain the history remote sensing images at the harbour, prepare offline the port area based on same The distributed image primitive and mask image of one width remote sensing images；Wherein distributed image primitive is multiple non-conterminous artificial Truncated picture block, mask image refer to by people by hand the mode of smearing obtain cover land area, expose water area Port area image；

S2：Phasor carries out registration when different to port area, comprises the following steps：

S21：For the remote sensing images I of the online two width width difference phase for obtaining₁And I₂And the distributed image of step S1 Primitive R, extracts the Feature Descriptor with scale invariability, and be designated as D respectively₁、D₂And D_R；

S22：By Feature Descriptor D₁With D_RMatched, obtained affine matrix H1, by D₂With D_RMatched, imitated Penetrate matrix H 2；

S23：Remote sensing images I1 is carried out into resampling according to affine matrix H1, I is obtained₁', by remote sensing images I2 according to affine Matrix H 2 carries out resampling, obtains I₂′；

S3：Using the mask image of step S1 to the image I after step S2 registrations₁' and I₂' masking operations are carried out, obtain two Phasor I during width₁" and I₂″；

S4：The I obtained to S3₁" and I₂" water area carry out extra large color homogenization, and respectively obtain image I₁" ' and I₂″′；

S5：Phasor I during two width after sea is homogenized₁" ' and I₂" ' respective pixel value is poor, obtains two width phase difference Figure：

Wherein, P_IncreaseRepresent the pixel in the increase figure in two width difference diagrams, P_ReduceRepresent the reduction in two width difference diagrams Pixel in figure, P₁And P₂Respectively I₁" ' and I₂" ' in pixel；

S6：Using the image partition method based on maximum entropy, segmentation threshold k is determined, after obtaining binaryzation based on threshold value k Image, thus makees from step S5 region of variation to be split in the two images after difference, obtains the region of variation of two width figures；

S7：The region of variation of phasor carries out the void based on priori when the two width port areas for obtaining difference is extracted to S6 It is alert to reject；

S8：Phasor carries out the false-alarm rejecting based on SVM classifier when two width to port area in S7 are different, and method includes The following steps：

S81：Cut into slices on a collection of naval vessel of interception by hand using history remote sensing images, segmentation behaviour is carried out to the section of each Ship Target Make, extract feature, be trained in feeding support vector machines, obtain disaggregated model；

S82：According to remaining region of variation in S7 steps S21 image I₁And I₂Cut on the naval vessel of middle interception fixed size Piece, it is same to each section to extract feature in S81, classified in the disaggregated model for sending into support vector machines, obtain final Ship Recognition result.

Preferably, in the step S1, mask images include more than 30% land area.

Preferably, in the S4, the method for extra large color homogenization is：Count the image grey level histogram of two width figure water areas H1 and h2, calculates its gray average and compares, it is assumed that the gray average of h1 is some higher respectively, then using h2 as benchmark Nogata Figure, to I₂" do not operated, to I₁" water area carry out histogram criteria；Calculate I₂" water area extra large color it is equal Value, I is filled with the extra large color average₁" and I₂" region covered by mask obtains image I₁" ' and I₂″′。

Preferably, the specific method that threshold value k is obtained in the S6 is：Image gray levels are divided into C by selected threshold value k₁And C₂ Two parts, wherein C₁It is that gray level is the pixel of [0,1,2 ..., k], C₂It is that gray level is one group of pixel of [k+1 ..., L-1], It is calculated maximum between-cluster varianceGray value k：

Wherein, P₁K () is set C₁The probability of generation, P₂K () is set C₂The probability of generation, m₁(k) and m₂K () is set C₁And C₂The average gray of pixel, m_GIt is global average；

Preferably, the method that the false-alarm based on priori is rejected in the S7 is：

S71：Bianry image to being obtained by S6 uses UNICOM's zone marker；

S72：To the bianry image by the way of sliding block scanning, removal points are small less than given threshold or dispersion In the UNICOM region of given threshold；

S73：Size according to UNICOM region rejects impossible candidate region；

S74：Impossible candidate region is rejected according to the neighborhood characteristic around target, i.e.,：By variance around UNICOM region The regional determination that given threshold is less than with UNICOM region region variance difference is false-alarm；

S75：Using the area information of candidate target itself, impossible candidate region is rejected.

The present invention has the advantages that：

1), the present invention utilizes the method for registering images based on distributed image primitive, to the distributed image for pre-saving Primitive extracts the matching characteristic point of image, instead of the original method that matching characteristic point is extracted to whole figure, accelerates images match Speed, meanwhile, distributed primitive selects more characteristic image-region, and the Feature Descriptor of extraction is unique relatively strong, subtracts The possibility of few error hiding, improves registration accuracy；

2), the present invention carries out extra large land segmentation using the offline template for preparing by hand, although increased artificial degree of participation, but It is also while increased the levels of precision of extra large land segmentation, it is ensured that be not in excessive because registration is inaccurate after image work difference The excessive problem of caused false-alarm, it is ensured that last change accuracy of detection；

3), the present invention is not classified first to the candidate region after segmentation using grader, but extracts four boundary's information, outer The features such as rectangular information are connect, the prior information with reference to naval vessel on remote sensing images, the false-alarm that setting empirical value carries out primary is picked Remove, so can significantly filter out most false candidate region, reduce the sample number to be sorted of grader treatment below Amount, accelerates system running speed.

Brief description of the drawings

Fig. 1 is flow chart of the method for the present invention；

Fig. 2 is the simulation result figure of regional change of the invention, wherein, (a) is that naval vessel reduces figure, and (b) increases for naval vessel Figure.

Specific embodiment

According to the flow chart of Fig. 1, the present invention mainly takes detailed description below：

S1：For a certain specific harbour that pay close attention to its change it may first have to obtain the history remote sensing images at the harbour, lead to Artificial mark or interception remote sensing images are crossed, the distributed image based on same width remote sensing images of the port area is prepared offline Primitive and mask image；Wherein distributed image primitive is many non-conterminous artificial truncated picture blocks, and must be at least Comprising more than 30% land area, the cartographic feature distributed intelligence comprising the harbour periphery, the figure comprising these distributed intelligences The uniqueness of image registration is can ensure that as primitive；Mask image refer to by people by hand the mode of smearing obtain cover land area Domain, the port area image for exposing water area.

S2：To port area when different phasor carry out the method for registration and comprise the following steps：

S21：For the remote sensing images I of the online 2 width difference phase for obtaining₁And I₂And the distributed image base of step S1 First R, extracts the Feature Descriptor with scale invariability, and be designated as D respectively₁、D₂And D_R；Wherein, feature states sub- D_RIt is multiple The combination of the feature that distributed image primitive R is proposed；

S22：By D₁With D_RMatched, obtained affine matrix H1, by D₂With D_RMatched, obtained affine matrix H2；

S23：Remote sensing images I1 is carried out into resampling according to affine matrix H1, I is obtained₁', by remote sensing images I2 according to affine Matrix H 2 carries out resampling, obtains I₂', by resampling, remote sensing images I1 and remote sensing images I2 is gone to and offline image phase Same visual angle；

S3：Using the mask image of step S1 to the image I after step S2 registrations₁' and I₂' masking operations are carried out, cover land Ground region and phasor I when exposing two width of water area₁" and I₂″；

S4：The I obtained to S3₁" and I₂" water area carry out extra large color homogenization, gray scale is moved in same rank, Concrete operations are the image grey level histogram h1 and h2 for counting two width figure water areas, its gray average is calculated respectively and is compared, Assuming that the gray average of h1 is some higher, then using h2 as benchmark histogram, to I₂" do not operated, but to I₁" sea area Domain carries out histogram criteria, by histogram criteria, can will weaken sea illumination variation, sea color is tended to one Cause, calculate I₂" water area extra large color average, fill I with the extra large color average₁" and I₂" region covered by mask, this master drawing As upper remaining sea and ship, without land area, image I is respectively obtained₁" ' and I₂″′；

S5：Phasor I during two width after sea is homogenized₁" ' and I₂″′；Respective pixel value is poor, obtains phase difference diagram, It is specifically divided into two results --- increased and reduction, the formula for using is as follows：Wherein, P₁And P₂Respectively I₁" ' and I₂" ' in pixel；

S6：Using the image partition method based on maximum entropy, segmentation threshold k is determined, after obtaining binaryzation based on threshold value k Image, thus makees from step S5 region of variation to be split in the two images after difference, obtains the region of variation of two width figures； Wherein, the specific method for obtaining threshold value k is：Image gray levels are divided into C by selected threshold value k₁And C₂Two parts, wherein C₁It is gray scale Level is the pixel of [0,1,2 ..., k], C₂It is that gray level is one group of pixel of [k+1 ..., L-1], is calculated side between maximum kind DifferenceGray value k：

Wherein, P₁K () is set C₁The probability of generation, P₂K () is set C₂The probability of generation, m₁(k) and m₂K () is set C₁And C₂The average gray of pixel, m_GIt is global average；

S7：The region of variation of phasor carries out the void based on priori when the two width port areas for obtaining difference is extracted to S6 Alert to reject, method comprises the following steps：

S71：Bianry image to being obtained by S6 uses UNICOM's zone marker, obtains boundary rectangle information, the length of target With the information such as width；

S72：To bianry image by the way of sliding block scanning, removal points are less than less than given threshold or dispersion and set Determine the UNICOM region of threshold value；

S73：Size according to UNICOM region rejects impossible candidate region, i.e., under current resolution, by length and The judgement that width exceeds possible size is false-alarm；

S74：Impossible candidate region is rejected according to the neighborhood characteristic around target, acquiescence naval vessel surrounding is sea, ash Degree level it is relatively low, its variance with the variance of naval vessel region compared with gap difference it is larger, therefore by variance around candidate region with Candidate region region variance difference is false-alarm less than the regional determination of given threshold；

S75：Using the area information of candidate target itself, including object pixel and boundary rectangle area are than feature, shape Ratio characteristics, energy intensity measurement etc., reject impossible candidate region；

S8：Phasor carries out the false-alarm rejecting based on SVM classifier when two width to S75 port areas are different, and method includes The following steps：

S81：Cut into slices on a collection of naval vessel of interception by hand using history remote sensing images, segmentation behaviour is carried out to the section of each Ship Target Make, extract the features such as rectangular degree, dutycycle, minimum enclosed rectangle area, symmetry, the length-width ratio of section, send into supporting vector It is trained in machine SVM, obtains disaggregated model；

S82：According to remaining candidate region in S7 steps S21 image I₁And I₂Cut on the naval vessel of middle interception fixed size Piece, it is same to each section to extract features described above, classified in feeding SVMs, obtain final Ship Recognition knot Really；

S9：Finally think remaining all Ship Targets, preserve and arrange positional information and the related slices letter of target Breath, section can file and preserve, be browsed with to be checked.As shown in Fig. 2 (a) and (b), the dot in figure is that expression is identified Naval vessel, therefore the method for the present invention can automatically derive the change Ship Target and its position and size at specific harbour etc. letter Breath.

In sum, presently preferred embodiments of the present invention is these are only, is not intended to limit the scope of the present invention. All any modification, equivalent substitution and improvements within the spirit and principles in the present invention, made etc., should be included in of the invention Within protection domain.

Claims (5)

1. a kind of port area Ship Target change detecting method based on remote sensing images, it is characterised in that comprise the following steps：

S1：For pay close attention to harbour, obtain the history remote sensing images at the harbour, prepare offline the port area based on same width The distributed image primitive and mask image of remote sensing images；Wherein distributed image primitive is multiple non-conterminous artificial interceptions Image block, mask image refer to by people by hand the mode of smearing obtain cover land area, expose the harbour of water area Area image；

S2：Phasor carries out registration when different to port area, comprises the following steps：

S21：For the remote sensing images I of the online two width width difference phase for obtaining₁And I₂And the distributed image primitive of step S1 R, extracts the Feature Descriptor with scale invariability, and be designated as D respectively₁、D₂And D_R；

S22：By Feature Descriptor D₁With D_RMatched, obtained affine matrix H1, by D₂With D_RMatched, obtained affine matrix H2；

S23：Remote sensing images I1 is carried out into resampling according to affine matrix H1, I is obtained₁', by remote sensing images I2 according to affine matrix H2 carries out resampling, obtains I₂′；

S3：Using the mask image of step S1 to the image I after step S2 registrations₁' and I₂' masking operations are carried out, when obtaining two width Phasor I₁" and I₂″；

S4：The I obtained to S3₁" and I₂" water area carry out extra large color homogenization, and respectively obtain image I₁" ' and I₂″′；

S5：Phasor I during two width after sea is homogenized₁" ' and I₂" ' respective pixel value is poor, obtains two width phase difference diagrams：

Wherein, P_IncreaseRepresent the pixel in the increase figure in two width difference diagrams, P_ReduceIn representing the reduction figure in two width difference diagrams Pixel, P₁And P₂Respectively I₁" ' and I₂" ' in pixel；

S6：Using the image partition method based on maximum entropy, segmentation threshold k is determined, the figure after binaryzation is obtained based on threshold value k Picture, thus makees from step S5 region of variation to be split in the two images after difference, obtains the region of variation of two width figures；

S7：S6 is extracted the two width port areas that obtain it is different when phasor region of variation carry out the false-alarm based on priori and pick Remove；

S8：Phasor carries out the false-alarm rejecting based on SVM classifier when two width to port area in S7 are different, and method includes following Step：

S81：Cut into slices on a collection of naval vessel of interception by hand using history remote sensing images, cutting operation carried out to the section of each Ship Target, Feature is extracted, is trained in feeding support vector machines, obtain disaggregated model；

S82：According to remaining region of variation in S7 steps S21 image I₁And I₂The naval vessel section of middle interception fixed size, it is right Each section is same to extract feature in S81, is classified in the disaggregated model for sending into support vector machines, obtains final warship Ship recognition result.

2. a kind of port area Ship Target change detecting method based on remote sensing images as claimed in claim 1, its feature It is that in the step S1, mask images include more than 30% land area.

3. a kind of port area Ship Target change detecting method based on remote sensing images as claimed in claim 1, its feature It is that in the S4, the method for extra large color homogenization is：The image grey level histogram h1 and h2 of two width figure water areas are counted, point Its gray average is not calculated and is compared, it is assumed that the gray average of h1 is some higher, then using h2 as benchmark histogram, to I₂" no Operated, to I₁" water area carry out histogram criteria；Calculate I₂" water area extra large color average, with the extra large color Average fills I₁" and I₂" region covered by mask obtains image I₁" ' and I₂″′。

4. a kind of port area Ship Target change detecting method based on remote sensing images as claimed in claim 1, its feature It is that the specific method that threshold value k is obtained in the S6 is：Image gray levels are divided into C by selected threshold value k₁And C₂Two parts, wherein C₁It is that gray level is the pixel of [0,1,2 ..., k], C₂It is that gray level is one group of pixel of [k+1 ..., L-1], is calculated maximum Inter-class varianceGray value k：

${\mathrm{\σ}}_E^2\left(k\right)=P_1\left(k\right){\[m_1\left(k\right)-m_G\]}^2+P_2\left(k\right){\[m_2\left(k\right)-m_G\]}^2$

Wherein, P₁K () is set C₁The probability of generation, P₂K () is set C₂The probability of generation, m₁(k) and m₂K () is set C₁With C₂The average gray of pixel, m_GIt is global average；

$P_1\left(k\right)=\underset{i=0}{\overset{k}{\Σ}}{p}_i;P_2\left(k\right)=\underset{i=k+1}{\overset{255}{\Σ}}{p}_i=1-P_1\left(k\right);m_G=\underset{i=0}{\overset{255}{\Σ}}{\mathrm{ip}}_i;m_1\left(k\right)=\underset{i=0}{\overset{k}{\Σ}}{\mathrm{ip}}_i;m_2\left(k\right)=\underset{i=k+1}{\overset{255}{\Σ}}{\mathrm{ip}}_i.$

5. a kind of port area Ship Target change detecting method based on remote sensing images as claimed in claim 1, its feature It is that the method that the false-alarm based on priori is rejected in the S7 is：

S71：Bianry image to being obtained by S6 uses UNICOM's zone marker；

S72：To the bianry image by the way of sliding block scanning, removal points are less than less than given threshold or dispersion and set Determine the UNICOM region of threshold value；

S73：Size according to UNICOM region rejects impossible candidate region；

S74：Impossible candidate region is rejected according to the neighborhood characteristic around target, i.e.,：By variance and connection around UNICOM region Logical region region variance difference is false-alarm less than the regional determination of given threshold；

S75：Using the area information of candidate target itself, impossible candidate region is rejected.