CN104036239B - Fast high-resolution SAR (synthetic aperture radar) image ship detection method based on feature fusion and clustering - Google Patents

Fast high-resolution SAR (synthetic aperture radar) image ship detection method based on feature fusion and clustering Download PDF

Info

Publication number
CN104036239B
CN104036239B CN201410234205.6A CN201410234205A CN104036239B CN 104036239 B CN104036239 B CN 104036239B CN 201410234205 A CN201410234205 A CN 201410234205A CN 104036239 B CN104036239 B CN 104036239B
Authority
CN
China
Prior art keywords
target
detection
sar image
region
area
Prior art date
Application number
CN201410234205.6A
Other languages
Chinese (zh)
Other versions
CN104036239A (en
Inventor
杨淑媛
焦李成
王士刚
侯彪
刘芳
刘红英
马晶晶
马文萍
熊涛
刘赵强
Original Assignee
西安电子科技大学
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 西安电子科技大学 filed Critical 西安电子科技大学
Priority to CN201410234205.6A priority Critical patent/CN104036239B/en
Publication of CN104036239A publication Critical patent/CN104036239A/en
Application granted granted Critical
Publication of CN104036239B publication Critical patent/CN104036239B/en

Links

Abstract

The invention discloses a fast high-resolution SAR (synthetic aperture radar) image ship detection method based on feature fusion and clustering. The fast high-resolution SAR image ship detection method comprises the following steps: on the basis of the back scattering characteristics of each ground object and the prior information of a ship target in an SAR image, positioning a target potential position index map by an Otsu algorithm and range constraint; on the index map, pre-screening to obtain a detection binary segmentation map by a CFAR (constant false alarm rate) algorithm based on a local contrast; carrying out morphological processing to a detection result, and extracting a potential target slice from the SAR image and a detected binary segmentation map according to a processing result; and carrying out K-means clustering to the extracted slice by a designed identification feature to obtain a final identification result. According to the fast high-resolution SAR image ship detection method based on feature fusion and clustering, the data volume of a detection stage is effectively reduced by pre-processing, and point-to-point detection is not needed/the time of point-to-point detection is saved. Meanwhile, a target identification problem under the condition of insufficient training samples at present can be solved by the designed characteristic and a non-supervision clustering method, the target can be effectively positioned, and the size of the target can be estimated.

Description

The quick Ship Detection of High Resolution SAR image that feature based merges and clusters

Technical field

The invention belongs to technical field of image processing, is related to a kind of feature based fusion and the quick warship of SAR image for clustering Ship detection method, for understanding and the interpretation of diameter radar image.

Background technology

Synthetic aperture radar (Synthetic Aperture Radar, SAR) is a kind of master being imaged with microwave Dynamic formula sensor, its imaging are not limited by conditions such as weather, illumination, can carry out round-the-clock, whole day to interesting target The observation of time.Multiband, complete polarization, the appearance of the SAR system of different working modes cause comprehensive earth observation to be possibly realized. At present, the interpretation capability of SAR image cannot meet the demand that the great amount of images collected is processed, and this directly results in SAR image The development of Interpretation Technology.It is the important means of realizing marine surveillance, fishery management and control that Ship Target Detection is carried out using SAR image, And China territorial waters is wide, marine resources enrich, carry out the detection of SAR image naval vessel significant.

The eighties in 20th century, Lincoln laboratory is proposed to be known based on the synthetic aperture radar automatic target of layering attention mechanism (SAR ATR) tertiary treatment flow process, does not become the common recognition of scientific circles.The model adopts layered shaping mode, first to view picture It is clearly not mesh target area to remove that SAR image carries out preview or detection, obtains potential target region;Then to potential target Region is differentiated, removes natural clutter false-alarm therein to obtain target region of interest;Finally area-of-interest is carried out More complicated feature extraction and classification, to realize the purpose of target recognition.With the in-depth for processing, data volume to be dealt with Can be fewer and feweri, and computation complexity can be increasing, such that it is able to realize efficient understanding and interpretation.SARATR is for target The real-time and accuracy rate of identification suffers from higher requirement, and needs are developed algorithm quickly and efficiently to meet actual need Ask.

Shown in SAR image using Ship Target and surrounding sea more than the algorithm of current SAR image naval vessel detection Feature difference, by setting an adaptive threshold being detected.Classical constant false alarm rate (CFAR) detector is basis The difference of Ship Target and surrounding sea gray feature, according to the false alarm rate for setting come adaptive after being modeled with normal distribution to background Detection threshold value should be found, theoretic constant false alarm rate can be reached.CFAR detections are a kind of simple and effective detection methods, according to Following formula carries out the presence or absence for judging impact point.

Wherein, μROIIt is the average in object support region, μCIt is the average of clutter supporting zone, σCIt is clutter supporting zone Standard deviation, λ are relevant with a default false alarm rate constants.If above formula meets at the certain point, the pixel to be measured is judged to It is set to target pixel points, is otherwise clutter pixel.This method detection speed by the way of pointwise detection is restricted, together When there are problems that the poor fitting of model causes accuracy of detection to be affected, with positioned at the extremely limited observation picture in sliding window region There is statistical deviation come the parameter for estimating normal distribution in plain value, this frequently can lead to model mismatch and makes the property of detector Can be severely impacted.There is bulk redundancy information in SAR image, the partitioning algorithm and region operation based on Otsu can be Detection-phase reduces data volume, improves interpretation speed and efficiency.Above pretreatment is combined with two-parameter CFAR and is expected to The overall performance of detector is lifted on the premise of keeping constant false alarm rate.

The design of descriminator is two classification problems.Current method is based on having a supervision message, need substantial amounts of Typical target and false-alarm are cut into slices come the parameter for obtaining grader, such as support vector machine (SVM) and QD.And train in practical situations both The acquisition of data takes time and effort very much, and training data is in different sensors, imaging circumstances and target type in itself Under obtain, it is difficult to characterize data to be tested.Texture, region and the contrast that will be extracted in artwork section and binary map section The Feature Fusion of degree is got up and can provide more strong distinguishing ability.Based on target and the unsupervised K- of clutter feature priori Means clustering methods, according to target and clutter aggregation features of the section in feature space, with the mode of cluster to target and miscellaneous Ripple section makes a distinction, and can break away from the dependence to training sample, realizes efficient discriminating.

The content of the invention

For above-mentioned the deficiencies in the prior art, the present invention proposes a kind of feature based fusion and the SAR image of cluster is fast Speed ship object detection method, to reduce data volume, reduce computing cost, while using the unsupervised clustering of feature priori To break away from the restriction problem of training sample, finally realize efficient target detection with positioning.

Realize that technical thought of the invention is:The difference and target of different atural object scattering propertiess are combined in pretreatment stage Can not possibly there is mesh target area in the prior information of size, removal;Detection-phase utilizes target and surrounding clutter scattering propertiess Difference, realizes that with reference to local contrast feature prescreening obtains potential target region;Jing after Morphological scale-space, from artwork and inspection Area-of-interest section is proposed in binary map after survey;Design effective feature and according to feature priori, with the mode reality of cluster The discriminating of existing Ship Target.

Feature based fusion of the present invention and the quick Ship Detection of High Resolution SAR image for clustering, including it is as follows Step:

(1) pretreatment based on Terrain Scattering characteristic and priori

The Otsu optimum segmentation threshold value for making inter-class variance maximum is found according to the normalization histogram of image 1a), will be original SAR image is divided into bright area and dark areas two parts, obtains binary segmentation figure.

1b) to step 1a) area is much larger than during the binary segmentation figure that obtains removes binary segmentation figure after carrying out holes filling The connected region of Ship Target area, obtains the index map in potential target region.

(2) CFAR based on local contrast is detected

2a) the size according to Ship Target, chooses the object support area size needed for CFAR detections and clutter Support The size in domain.

Mirror reflection extension is carried out around border to original SAR image and index map 2b), the false alarm rate of detection is preset To adaptively determine the detection threshold value at each point according to background information in detection process.

2c) on the potential target position indicated by index map, the average energy and hollow cunning in object support region are calculated The standard deviation of the average energy and its energy of clutter supporting zone in dynamic window, it is equal according to the pixel of pixel supporting zone to be measured Whether the magnitude relationship of the detection threshold value at value and the point is judging the point as target pixel points.

After 2d) hollow sliding window moves the result for obtaining detecting in original SAR image, to the detection knot for obtaining Fruit carries out corroding expansive working to remove isolated test point and with holes filling supplementing the impact point being blanked.

(3) region of interesting extraction and its feature construction and fusion

3a) all connected regions are extracted in the binary segmentation figure from after detection and remove connected region of the area much smaller than target Domain, according to the barycenter and the size of default section of remaining area, extracts from the binary segmentation figure after original SAR image and detection Go out the section of area-of-interest.

3b) logarithm standard deviation is calculated as this field strength undulatory property of tolerance in each original SAR image section extracted Feature;The number that eight connectivity region is found in the binary segmentation figure section after each detection extracted is most strong as description region The feature of scattering point space divergence;Meanwhile, the target area position calculation correspondence according to indicated by the binary segmentation figure after detection Original SAR image in energy meansigma methodss as target area average energy tolerance.

Logarithm standard deviation, eight connectivity number of regions and target area average energy are normalized into shape after fusion respectively 3c) Into the characteristic vector with higher resolving ability, as the comprehensive description of each area-of-interest;Target area and clutter region Feature aggregation properties are presented on feature space, and for each feature of target area and clutter region is all have priori Guidance.

(4) target of feature based priori and K-means clusters differentiates

4a) setting cluster classification number, maximum iteration time and the initial cluster center determined by feature priori, according to Euclidean distance is measuring the similarity of each sample to be tested and cluster centre, and sorts out each sample with this, until K-means Cluster reaches convergence.

4b) according to the sample class for finally giving, the area-of-interest position gathered corresponding to the sample of target is found, And the region labeling for being finally defined as Ship Target in original SAR image with the bounding box of suitable size is out.

The minimum of target is extracted from the binary segmentation figure after the detection being demarcated as corresponding to the region of Ship Target 4c) Boundary rectangle, so as to draw the length and width of the Ship Target.

Further, step 1a) in, it is assumed that original SAR image is I, and its size is m × n, with making inter-class variance most Big optimum segmentation threshold value ToptDivide the image into into area pellucida domain and dark areas two parts, optimum segmentation threshold value ToptObtain according to following formula :

Wherein, mGIt is the average gray value of image, m (k) is pixel average of the gray value less than k, P1K () is that gray value is little Ratio shared by pixel in k.

Further, step 1b) in, holes filling is carried out to the binary segmentation figure after Otsu segmentations and is connected with completion Region, after counting the size of all connected regions, region of the area much larger than 3 times of Ship Target size is removed.

Further, step 3a) in, the method that all connected regions are extracted in the binary segmentation figure from after detection It is:Morphological operation is carried out to the binary segmentation figure after detection, the morphological operation includes being connected come completion with holes filling Region, removes isolated point with corrosion and expansion and connects abutment points, so as to obtain all connected regions.

Further, step 2c) in, the criterion of target pixel points is:If pixel supporting zone to be measured Pixel average is judged as target pixel points more than the detection threshold value at the point, the then point;If pixel supporting zone to be measured Pixel average is judged as clutter pixel less than or equal to the detection threshold value at the point, the then point.

Compared with prior art, it is an advantage of the invention that:

1st, prior information of the present invention according to the difference and naval vessel size of each atural object Electromagnetic Scattering Characteristics, optimum by Otsu The constraint of Threshold segmentation and connected region area, efficiently reduces the hunting zone of target, improves detection speed, overcome Pointwise detects brought time and computing cost.

2nd, separately designed in the section that the present invention is extracted in original SAR image and binary segmentation image and reflected for target Other new feature, and feature is combined to form the feature group with the ability of sign, have more than classical diagnostic characteristicses Good differentiation performance.Simultaneously under the guidance of feature prior information, by target and clutter in the aggregation of feature space, use Unsupervised clustering method realizes automatic target discriminating, effectively overcomes the problem of lack of training samples, has obtained more Effect is interpreted preferably.

The present invention is described in further details below with reference to drawings and Examples.

Description of the drawings

Fig. 1 is the flow chart of the present invention.

Fig. 2 is that the CFAR that the present invention is adopted detects hollow sliding window schematic diagram.

Fig. 3 (a) is original SAR image.

Fig. 3 (b) is the binary segmentation figure after Otsu segmentations, wherein the optimal threshold for obtaining is 60.

Fig. 3 (c) is the index map Jing after holes filling and region are removed.

Fig. 3 (d) is two-parameter CFAR testing result figures, wherein Pfa=0.03.

Fig. 4 is the ROC curve of CFAR detection of the present invention under not preprocessed and preprocessed.

Fig. 5 is the area-of-interest that the present invention is extracted from original SAR image and binary segmentation image.

Fig. 6 (a) is the logarithm standard deviation feature for extracting section.

Fig. 6 (b) is the eight connectivity number of regions feature for extracting section.

Fig. 6 (c) is the target area average energy feature for extracting section.

Fig. 7 (a) is original SAR image.

Fig. 7 (b) is the potential target administrative division map for differentiating prelocalization.

Fig. 7 (c) is result of three features of present invention employing after K-means discriminatings.

Fig. 7 (d) is result of three feature of Lincoln laboratorys after K-means discriminatings.

Fig. 8 is the target size estimated result that the present invention is obtained.

Specific embodiment

Embodiment 1:

There is the acquisition of data volume in prior art mentioned in the background and computing cost is too big, and receive training sample Restriction the problems such as, make up these technological deficiencies to reach, realize efficient target detection with positioning purpose, the present invention provide A kind of feature based fusion and the SAR image quick Ship Target Detection method for clustering.

Realize that technical thought of the invention is:The difference and target of different atural object scattering propertiess are combined in pretreatment stage Can not possibly there is mesh target area in the prior information of size, removal;Detection-phase utilizes target and surrounding clutter scattering propertiess Difference, realizes that with reference to local contrast feature prescreening obtains potential target region;Jing after Morphological scale-space, from artwork and inspection Area-of-interest section is proposed in binary map after survey;Design effective feature and according to feature priori, with the mode reality of cluster The discriminating of existing Ship Target.

The feature based fusion and the quick Ship Target Detection method of SAR image for clustering, as shown in figure 1, including such as Lower step:

Step 1:Pretreatment based on Terrain Scattering characteristic and priori

1.1) assume that original SAR image is I, its size is m × n, using optimum segmentation threshold value T that inter-class variance is maximumopt Original SAR image is divided into into area pellucida domain and dark areas two parts, optimum segmentation threshold value ToptObtain according to following formula:

Wherein, mGIt is the average gray value of image, m (k) is pixel average of the gray value less than k, P1K () is that gray value is little Ratio shared by pixel in k.

1.2) holes filling is carried out with completion connected region to the binary segmentation figure after Otsu segmentations, counts all connected regions After the size in domain, region of the area much larger than 3 times of Ship Target size is removed, the index map in potential target region is obtained Iindex

Step 2:CFAR based on local contrast is detected

2.1) according to the distance between the size and naval vessel on naval vessel priori presetting the hollow sliding window needed for CFAR detections The each parameter of mouth, as shown in Figure 2.The present invention by object support region be sized to 3 × 3 with to isolate strong clutter point press down System, the radius of protection zone are set to 15 to avoid other targets from being leaked in hollow sliding window, and clutter supports the number of pixel 120 are set to, to represent the surrounding clutter information of pixel to be measured.

2.2) due to the presence of hollow sliding window, need to original SAR image I and index map IindexCarry out mirror reflection It is extended to (m+30) × (n+30) sizes.In CFAR algorithms, false alarm rate PfaSelection should as far as possible in true verification and measurement ratio and Compromise is reached between false alarm rate, if PfaChoosing it is excessive, substantial amounts of clutter point, and P can be included in testing resultfaThat what is selected is too small Impact point can be lost again causes missing inspection.False alarm rate P in the present inventionfaElect 0.03 as.At every bit, if calculating clutter support The mean μ in regionCAnd standard deviation sigmaCAfterwards, corresponding detection threshold value T can be tried to achieve by following formula:

2.3) calculate the pixel mean μ of pixel supporting zone to be measuredROIAnd be compared with detection threshold value T for obtaining, if μROIThen the point is judged as target pixel points to > T, is otherwise clutter pixel.According to index map in the original SAR image of view picture Instruction move the detection that hollow sliding window completes to all potential target points successively, so as to obtain the two-value point of CFAR detections Cut figure Idet

2.4) performance in order to verify pretreatment to the effect of detection-phase and to DP-CFAR CFAR detectors is entered Row analysis, by default false alarm rate P in CFARfaIt is respectively set as from 0 starting with 0.01 as the centrifugal pump for being spaced 0.25, every The testing result of CFAR is obtained under individual false alarm rate, and the Groundtruth figures of hand labeled are matched in advance, calculated true Verification and measurement ratio PdWith real false alarm rate Pf, false alarm rate and verification and measurement ratio are drawn out CFAR inspections respectively as abscissa and vertical coordinate Survey experimenter's operation curve (ROC) of device.It is identical with the above for process that is not preprocessed and directly carrying out CFAR detections, Its corresponding ROC curve can be obtained.

Step 3:Region of interesting extraction and its feature construction and fusion

3.1) for the binary segmentation figure I after detectiondetCarry out morphological operation, including with holes filling come completion connect Region, removes isolated point with corrosion and expansion and connects abutment points, so as to obtain more regular region.

3.2) barycenter and area information of all connected regions, if the area of certain connected region is too small, explanation are extracted This region is the nontarget area comprising clutter, can be excluded as the probability of target area.According to this experiment SAR image The prior information of spatial resolution and naval vessel size, for region of the area less than 50 (minimum target half) is rejected.With Centered on remaining connected region barycenter, 64 × 64 are extracted respectively in the binary segmentation image from after original SAR image and detection big Little section, so as to obtain the section at potential target place.The potential target number that hypothesis is extracted is n, then scheme from original SAR The section extracted as in is expressed as C1,C2…,Cn, the section extracted in the binary segmentation figure from after detection is respectively B1, B2…,Bn, the size of each section is 64 × 64.

3.3) from C1,C2…,CnThe middle logarithm standard deviation feature for calculating the section of each potential target is characterizing in this region The fluctuation information of reflex strength.Computing formula is as follows:

Wherein, N be section in pixel sum and

3.4) from B1,B2…,BnThe middle eight connectivity areal for calculating each potential target section is strong in this region to characterize The distribution situation of the scattering space of points.Under normal circumstances, strong scattering point in target area is more assembled for clutter, so eight The number of connected region is fewer than clutter.

3.5) combine B1,B2…,BnWith C1,C2…,CnThe target area average energy for calculating each potential target section is used To characterize the average scattering intensity of target in this region.Presence due to detecting false-alarm, has the clutter region compared with strong scattering characteristic Can present with structure as target class and to target formed disturb.This feature can be according to the point institute for being detected as object pixel The scattering strength average of respective pixel is distinguishing target and clutter.Its computing formula is as follows:

Wherein, niIt is the number of pixels that impact point is judged in i-th section;More than it is various in, Ci(r is a) i-th In potential target section at r row a row pixel scattering strength value.

3.6) features above is combined into into matrix form, represents three features of section per a line respectively, have n rows, knot It is really as follows:

Assume that the significance level of each feature is identical, as each characteristic dimension is different, need to return every string of F One changes operation to constitute the feature group of resolving ability.In the present invention, we are by each row normalization in F between 0 and 1.

Step 4:The target of feature based priori and K-means clusters differentiates

4.1) because the result of cluster only has target and clutter, the cluster number in K-means is set to into 2, according to Initial cluster center is set to [1,0,1] and [0,1,0] by the feature priori of target and clutter, and maximum iteration time is set to 100.Often The secondary Euclidean distance for asking for each sample and cluster centre, by sample be classified as therewith closer to a class, while with being classified as each class Sample average updating such center, till cluster centre no longer changes or reaches maximum iteration time.

4.2) aggregation being had from target and clutter feature space, in theory target be finally classified as first Class and clutter is classified as Equations of The Second Kind.To differentiate redness is used in original SAR image for mesh target area with reference to final cluster result Square frame marks out to determine its position.Extract the minimum external world rectangle frame for surrounding target simultaneously according to the binary map of detection, according to This can be used to the size for estimating target.

Embodiment 2:

The effect of the present invention is further illustrated by following emulation experiment.

(1) experiment simulation condition:

The data adopted by this experiment are TerraSAR high-resolution SAR images.The data are by TerraSAR Synthetic aperture of the resolution in the one regional coastline of the Straits of Gibraltar that satellite is obtained under HH polarization modes, X-band for 1m Radar image, comprising types of ground objects such as ocean, mountain range, building, river, harbour and naval vessels, wherein naval vessel mesh to be detected Indicate 21 and target type is different.The scene size covered by experimental data is 2987m × 4134m, is each pixel 8bits Gray-scale maps.This experiment CPU be Intel (R) Core (TM) i5-3470, dominant frequency be 3.2GHz, the interior WINDOWS7 for saving as 4G MATLAB 2012a in operating system using 32 are emulated.

(2) target detection and discriminating Performance evaluation criterion:

(2a) verification and measurement ratio and false alarm rate

Assume that SAR image size is M × N and the total number of object pixel is Ntarget, then clutter pixel in SAR image Sum is Nclutter=M × N-Ntarget.If the number of the object pixel for detecting is Ndt, the false-alarm number of generation is Ndc, then it is real Border false-alarm probability and actually detected probability are:

In advance refer to figure with artificial demarcation, by the testing result obtained under different false alarm rates with compare with reference to figure with calculating Two values above.The result for obtaining is depicted as into ROC curve, curve is more big with the area of transverse axis institute enclosing region, shows Detection performance is better.

(2b) total error number, fail to report number, total accuracy and target accuracy

The discriminating stage with total error number, fail to report number, total accuracy and target accuracy evaluating diagnostic characteristicses and discriminating The quality of device, its definition difference are as follows:

Total error number is judged to the number of clutter false-alarm for target and clutter false-alarm be judged to the number of target and, that is, have

ne=ntc+nct

Wherein, neIt is total error number, ntcThe number of clutter false-alarm is judged to for target, number, n is also referred to as failed to reportctIt is empty for clutter Police is judged to the number of target.

If the number of slices extracted by the discriminating stage is n, real goal number of slices therein is m, then total accuracy and target Accuracy is respectively defined as:

Generally, SAR ATR require that detection and discriminating stage should be before guarantee verification and measurement ratio and target accuracy be 1 Put, false alarm rate is reduced as far as possible by adjusting parameter preset, so as to realize complete, the efficient positioning to target.

(2c) interpret the time

The interpretation time refers to the total time-consuming from input SAR image to output testing result.As SAR ATR have to real-time Very high requirement, is generally required and is completed the positioning to target in a short period of time, so run time should be as far as possible It is short reaching application request.

(3) experiment content

Experiment one

Entered using the feature based fusion designed by the present invention and the quick Ship Detection of High Resolution SAR image for clustering Row target detection, experimental result are as shown in Figure 3.Wherein:

Fig. 3 (a) is original SAR image;

Fig. 3 (b) is the binary segmentation figure after Otsu segmentations, wherein the optimal threshold for obtaining is 60;

Fig. 3 (c) is the index map Jing after holes filling and region are removed;

Fig. 3 (d) is two-parameter CFAR testing result figures, wherein Pfa=0.03;

From Fig. 3 (a), Fig. 3 (b), Fig. 3 (c), Fig. 3 (d), the present invention is significantly reduced at data by pretreatment Reason amount so that the false alarm rate of detection is greatly lowered, so as to accelerate the speed of detection-phase.Fig. 4 delineates not preprocessed With in the case of two kinds of pretreatment, experimenter's operation curve of resulting two-parameter CFAR detections.It can be seen that Pretreatment causes experimenter's operation curve to left, increased the area surrounded by curve, so as to improve detection-phase Performance.Table 1 lists the calculating used time in the case of two kinds.

1 detection-phase of table calculates the used time

As it can be seen from table 1 pretreatment and the combination of two-parameter CFAR cause the data volume for detecting greatly to reduce, so as to drop The low time complexity of algorithm.

Experiment two

Testing result to obtaining is carried out after the Morphological scale-spaces such as burn into expansion and holes filling, from original SAR image and Extract the slice map comprising potential target that size is for 64 × 64, the part knot of experiment in binary segmentation image after detection simultaneously Fruit is as shown in Figure 5.There is very strong similarity between the section extracted as seen from Figure 5, it is desirable to have the method for effect carries out area to which Point.

Logarithm standard deviation corresponding to each section, eight connectivity areal and target area average energy three are calculated respectively The value of individual feature, will be combined into characteristic vector and forms description to cutting into slices after feature normalization, experimental result is as shown in Figure 6.Its In:

Fig. 6 (a) is the logarithm standard deviation feature for extracting section;

Fig. 6 (b) is the eight connectivity number of regions feature for extracting section;

Fig. 6 (c) is the target area average energy feature for extracting section;

Based on each section similarity characteristically, need to combine multiple features and could be formed to target and clutter The fine differentiation of false-alarm.

According to target slice and false-alarm section prior information characteristically, cluster to realize for mesh with K-means Target discriminating, as a result as shown in Figure 7.Wherein:

Fig. 7 (a) is original SAR image;

Fig. 7 (b) is the potential target administrative division map for differentiating prelocalization;

Fig. 7 (c) is result of three features of present invention employing after K-means discriminatings;

Fig. 7 (d) is result of three feature of Lincoln laboratorys after K-means discriminatings.

From Fig. 7 (a), Fig. 7 (b), Fig. 7 (c), Fig. 7 (d), the present invention can overcome SAR image to instruct in actual applications Practice the rare problem of sample, Ship Target Detection can be effectively completed by Feature Fusion is combined with unsupervised clustering Task.Designed feature has compared with classical three feature of Lincoln laboratorys preferably distinguishes performance, has obtained more Good identification result.

High Resolution SAR image provides more abundant scene information, and result that can be after discriminating is estimating target Size, so as to being that marine surveillance and fishery management and control provide beneficial help.The minimum enclosed rectangle of gained partial target Figure is as shown in figure 8, the resolution with reference to SAR image is estimated that the size of each target.

Table 2 lists comparing result of the present invention with Lincoln laboratory method on performance is differentiated, wherein Lincoln experimental reality Three features be respectively logarithm standard deviation, fractal dimension and arrangement energy ratio.

Total error number of the 2 liang of stack features of table under K-means, fail to report number, total accuracy and target accuracy

As seen from Table 2, traditional three feature of Lincoln laboratory of aspect ratio designed by the present invention is with more representational, while The discrimination method clustered based on K-means achieves more satisfied result, and every identification beacon is close to ideal value.

To sum up, the present invention can be greatly lowered the data complexity of SARATR, improve the detection performance of detector, while Overcome the problem of lack of training samples, more efficiently target differentiated with clutter false-alarm, reduce the time needed for interpretation, Understand in SAR image and have certain application prospect with interpretation field.

Embodiments of the present invention are described above in conjunction with accompanying drawing, but the present invention are not limited to above-mentioned embodiment, In the ken that one skilled in the relevant art possesses, can be making on the premise of without departing from present inventive concept Various change.

Claims (5)

1. the quick Ship Detection of High Resolution SAR image that feature based merges and clusters, it is characterised in that including following step Suddenly:
(1) pretreatment based on Terrain Scattering characteristic and priori
The Otsu optimum segmentation threshold values for making inter-class variance maximum 1a) are found according to the normalization histogram of image, original SAR is schemed As being divided into bright area and dark areas two parts, binary segmentation figure is obtained;
1b) to step 1a) area is much larger than naval vessel during the binary segmentation figure that obtains removes binary segmentation figure after carrying out holes filling The connected region of target area, obtains the index map in potential target region;
(2) CFAR based on local contrast is detected
2a) the size according to Ship Target, the object support area size and clutter supporting zone needed for selection CFAR detections Size;
Mirror reflection is carried out around border to original SAR image and index map 2b) and expands to the size being adapted to needed for detection, in advance The false alarm rate of detection is set the detection threshold value at each point is adaptively determined according to background information in detection process;
2c) on the potential target position indicated by index map, the average energy and hollow sliding window in object support region are calculated The standard deviation of the average energy and its energy of clutter supporting zone in mouthful, according to the pixel average of pixel supporting zone to be measured with Whether the magnitude relationship of the detection threshold value at the point is judging the point as target pixel points;
After 2d) hollow sliding window moves the result for obtaining detecting in original SAR image, the testing result to obtaining is entered Row corrosion expansive working is removing isolated test point and with holes filling supplementing the impact point being blanked;
(3) region of interesting extraction and its feature construction and fusion
3a) all connected regions are extracted in the binary segmentation figure from after detection and remove connected region of the area much smaller than target, According to the barycenter and the size of default section of remaining area, extract from the binary segmentation figure after original SAR image and detection The section of area-of-interest;
3b) logarithm standard deviation is calculated as the spy for measuring this field strength undulatory property in each original SAR image section extracted Levy;The number in eight connectivity region is found in the binary segmentation figure section after each detection extracted as description region most strong scattering The feature of space of points divergence;Meanwhile, the corresponding original of target area position calculation according to indicated by the binary segmentation figure after detection Tolerance of the meansigma methodss of energy as target area average energy in beginning SAR image;
3c) logarithm standard deviation, eight connectivity number of regions and target area average energy are normalized after fusion respectively and form tool There is the characteristic vector of higher resolving ability, as the comprehensive description of each area-of-interest;Target area and the spy in clutter region Levy and aggregation properties are presented on feature space, and for each feature of target area and clutter region is all the finger that has priori Lead;
(4) target of feature based priori and K-means clusters differentiates
4a) setting cluster classification number, maximum iteration time and the initial cluster center determined by feature priori, according to Euclidean Distance is measuring the similarity of each sample to be tested and cluster centre, and sorts out each sample with this, until K-means is clustered Reach convergence;
4b) according to the sample class for finally giving, the area-of-interest position gathered corresponding to the sample of target is found, is used in combination The bounding box of suitable size is finally defined as the region labeling of Ship Target out in original SAR image;
4c) the minimum external of target is extracted from the binary segmentation figure after the detection being demarcated as corresponding to the region of Ship Target Rectangle, so as to estimate the approximate length and width of the Ship Target.
2. the quick Ship Detection of High Resolution SAR image that feature based according to claim 1 merges and clusters, its It is characterised by, step 1a) in, it is assumed that original SAR image is I, and its size is m × n, with the optimum for making inter-class variance maximum Segmentation threshold ToptDivide the image into into area pellucida domain and dark areas two parts, optimum segmentation threshold value ToptObtain according to following formula:
T o p t = m a x k [ m G P 1 ( k ) - m ( k ) ] 2 P 1 ( k ) [ 1 - P 1 ( k ) ]
Wherein, mGIt is the average gray value of image, m (k) is pixel average of the gray value less than k, P1K () is that gray value is less than k Ratio shared by pixel.
3. the quick Ship Detection of High Resolution SAR image that feature based according to claim 1 merges and clusters, its It is characterised by:Step 1b) in, holes filling is carried out with completion connected region to the binary segmentation figure after Otsu segmentations, is united After counting the size of all connected regions, region of the area much larger than 3 times of Ship Target size is removed.
4. the quick Ship Detection of High Resolution SAR image that feature based according to claim 1 merges and clusters, its It is characterised by, step 3a) in, the method that all connected regions are extracted in the binary segmentation figure from after detection is:To detection Binary segmentation figure afterwards carries out morphological operation, and the morphological operation is included with holes filling come completion connected region, with corruption Erosion and expansion remove isolated point and connect abutment points, so as to obtain all connected regions.
5. the quick Ship Detection of High Resolution SAR image that feature based according to claim 1 merges and clusters, its It is characterised by, step 2c) in, the criterion of target pixel points is:If the pixel average of pixel supporting zone to be measured More than the detection threshold value at the point, then the point is judged as target pixel points;If the pixel average of pixel supporting zone to be measured Less than or equal to the detection threshold value at the point, then the point is judged as clutter pixel.
CN201410234205.6A 2014-05-29 2014-05-29 Fast high-resolution SAR (synthetic aperture radar) image ship detection method based on feature fusion and clustering CN104036239B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201410234205.6A CN104036239B (en) 2014-05-29 2014-05-29 Fast high-resolution SAR (synthetic aperture radar) image ship detection method based on feature fusion and clustering

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201410234205.6A CN104036239B (en) 2014-05-29 2014-05-29 Fast high-resolution SAR (synthetic aperture radar) image ship detection method based on feature fusion and clustering

Publications (2)

Publication Number Publication Date
CN104036239A CN104036239A (en) 2014-09-10
CN104036239B true CN104036239B (en) 2017-05-10

Family

ID=51467005

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201410234205.6A CN104036239B (en) 2014-05-29 2014-05-29 Fast high-resolution SAR (synthetic aperture radar) image ship detection method based on feature fusion and clustering

Country Status (1)

Country Link
CN (1) CN104036239B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104408482B (en) * 2014-12-08 2019-02-12 电子科技大学 A kind of High Resolution SAR Images object detection method
CN106170819B (en) * 2014-12-26 2019-04-05 中国海洋大学 A kind of diameter radar image Ship Target rapid detection method
CN105205450B (en) * 2015-08-24 2018-08-07 辽宁工程技术大学 A kind of SAR image target extraction method based on irregular mark point process
CN105223561B (en) * 2015-10-23 2017-10-24 西安电子科技大学 Radar ground target discriminator design method based on spatial distribution
CN107977667B (en) * 2016-10-21 2019-02-19 西安电子科技大学 SAR target discrimination method based on semi-supervised coorinated training
CN106803078A (en) * 2017-01-18 2017-06-06 哈尔滨工业大学 A kind of SAR image Ship Target dividing method
CN110334706A (en) * 2017-06-30 2019-10-15 清华大学深圳研究生院 A kind of images steganalysis method and device

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102208017A (en) * 2011-05-30 2011-10-05 西安电子科技大学 Ship detection method based on high-resolution synthetic aperture radar images
US8422738B1 (en) * 2008-08-25 2013-04-16 The United States Of America As Represented By The Secretary Of The Navy Adaptive automated synthetic aperture radar vessel detection method with false alarm mitigation
CN103400156A (en) * 2013-07-04 2013-11-20 西安电子科技大学 CFAR (Constant False Alarm Rate) and sparse representation-based high-resolution SAR (Synthetic Aperture Radar) image ship detection method

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102622607B (en) * 2012-02-24 2013-09-25 河海大学 Remote sensing image classification method based on multi-feature fusion

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8422738B1 (en) * 2008-08-25 2013-04-16 The United States Of America As Represented By The Secretary Of The Navy Adaptive automated synthetic aperture radar vessel detection method with false alarm mitigation
CN102208017A (en) * 2011-05-30 2011-10-05 西安电子科技大学 Ship detection method based on high-resolution synthetic aperture radar images
CN103400156A (en) * 2013-07-04 2013-11-20 西安电子科技大学 CFAR (Constant False Alarm Rate) and sparse representation-based high-resolution SAR (Synthetic Aperture Radar) image ship detection method

Also Published As

Publication number Publication date
CN104036239A (en) 2014-09-10

Similar Documents

Publication Publication Date Title
Huang et al. A multidirectional and multiscale morphological index for automatic building extraction from multispectral GeoEye-1 imagery
Prieto et al. Using self-organising maps in the detection and recognition of road signs
US8611604B2 (en) Object detection device
Guo et al. Classification of airborne laser scanning data using JointBoost
CN102622607B (en) Remote sensing image classification method based on multi-feature fusion
Yu et al. Learning hierarchical features for automated extraction of road markings from 3-D mobile LiDAR point clouds
Cord et al. Automatic road defect detection by textural pattern recognition based on AdaBoost
Xu et al. Multiple-entity based classification of airborne laser scanning data in urban areas
Corcoran et al. Segmentation performance evaluation for object-based remotely sensed image analysis
Chen et al. Vehicle detection in high-resolution aerial images via sparse representation and superpixels
CN103902976B (en) A kind of pedestrian detection method based on infrared image
CN102214298B (en) Method for detecting and identifying airport target by using remote sensing image based on selective visual attention mechanism
CN103400147B (en) Target fish recognition method and system based on image procossing
CN104166841A (en) Rapid detection identification method for specified pedestrian or vehicle in video monitoring network
CN102044069B (en) Method for segmenting white blood cell image
WO2017190574A1 (en) Fast pedestrian detection method based on aggregation channel features
CN103048329B (en) A kind of road surface crack detection method based on active contour model
Hoogs et al. A common set of perceptual observables for grouping, figure-ground discrimination, and texture classification
CN103442209B (en) Video monitoring method of electric transmission line
CN102646200B (en) Image classifying method and system for self-adaption weight fusion of multiple classifiers
Derpanis et al. Classification of traffic video based on a spatiotemporal orientation analysis
CN102750703B (en) Remote sensing image airport automatic detecting method based on linear cutting
CN103034863A (en) Remote-sensing image road acquisition method combined with kernel Fisher and multi-scale extraction
Kim et al. 3D classification of power-line scene from airborne laser scanning data using random forests
Suruliandi et al. Local texture patterns-a univariate texture model for classification of images

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
TR01 Transfer of patent right
TR01 Transfer of patent right

Effective date of registration: 20180702

Address after: 710077 Shaanxi provincial science and technology resources co ordination center, room 10, 10 zhang85 Road, Xi'an high tech Zone, A405, China

Patentee after: Xi'an Wien think Union Data Technology Co., Ltd.

Address before: No. 2 Taibai Road, Xi'an, Shaanxi Province, Shaanxi

Patentee before: Xidian University