WO2016101279A1

WO2016101279A1 - Quick detecting method for synthetic aperture radar image of ship target

Info

Publication number: WO2016101279A1
Application number: PCT/CN2014/095202
Authority: WO
Inventors: 于方杰; 陈戈; 韩勇; 马纯永; 田丰林; 范龙庆; 郭建
Original assignee: 中国海洋大学
Priority date: 2014-12-26
Filing date: 2014-12-26
Publication date: 2016-06-30
Also published as: CN106170819B; CN106170819A

Abstract

A quick detecting method for synthetic aperture radar image of ship target, comprising the following steps: (1) a step of separating land and sea; (2) a step of target screening; (3) setting a background cluster statistical model; (4) GPU processes three types of images respectively according to corresponding constant false alarm rate detection threshold T1 thereto at a GPU platform, and acquires a target area, and the three types of images respectively adapts different processing algorithms to calculate the threshold T1. The method firstly separates the land and sea area, and filters images of the land part, and improves a detection efficiency; secondly, performs a preliminary statistics on shapes, and sets a suitable overall threshold, and preliminary screens the SAR image target, and splits images to several sub-image blocks; finally utilizes CUDA technology to perform a constant false alarm rate detection on the three types of distributed shapes and detects the effective ship target. The present invention can accurately and rapidly accomplish detection of the ship target.

Description

Rapid detection method for ship target of synthetic aperture radar image

Technical field

The invention belongs to the technical field of image processing, and in particular relates to a rapid detection method for a ship target of a synthetic aperture radar image.

Background technique

Ship inspection is a routine task in the world's coastal countries. It has a wide range of applications in civil and military fields. It can carry out water transportation, illegal hunting, smuggling detection and management of specific sea areas and ports, and rescue ships for disasters. The sea area is vast, with an area of more than 3 million square kilometers. It is rich in marine resources. It is of great value and significance to carry out research on ship target detection.

Synthetic Aperture Radar (SAR) is a mature active microwave imaging radar. It has all-weather, all-day, and penetrating capability. It is compared with traditional visible and infrared sensors. The aspect has a unique advantage. With the development of technologies such as embedded technology, integrated circuit technology, and micro-machine manufacturing, SAR has gradually achieved miniaturization and miniaturization. At the same time, due to its low cost, flexibility and access to any designated place, UAVs can complement satellite remote sensing technology and rapidly develop in ocean observation applications. As the resolution of SAR continues to increase, the amount of data information provided by SAR images is also increasing. Quickly and accurately interpreting SAR images and obtaining useful information is an important issue in current SAR target detection. Fast analysis of SAR graphics, traditional serial algorithms have high requirements on system hardware, high-speed CPU, large-capacity memory and hard disk, and the performance improvement of system hardware is still very limited, it is difficult to meet the current slowdown of SAR images. Detect demand.

Utility model content

In order to solve the technical problem that the existing synthetic aperture radar target detection method has high hardware requirements and slow operation speed, the invention provides a rapid detection method for synthetic aperture radar image ship target, which can solve the above problems.

In order to solve the above technical problem, the present invention is implemented by the following technical solutions:

A method for quickly detecting a target image of a synthetic aperture radar, comprising the following steps:

(1) The sea-land separation step, evolving the boundary curve, and separating the sea and land with the boundary curve as the boundary, and obtaining the image of the marine area with effective targets;

(2) Target screening steps, including:

(21), setting the gray threshold T, assigning the index value of the pixel whose gray value is greater than T in the image of the ocean area to the gray value of the pixel, otherwise assigning a value of 0, and establishing an index for all the obtained index values. matrix;

(22) setting a non-zero region in the index matrix as a candidate target region;

(23) dividing the image of the marine area into a plurality of sub-images by using the position of the candidate target area, and each candidate target area corresponds to one sub-image;

(3) Set the background clutter statistical model, including:

(31), respectively calculating the background change index BI of each sub-image;

(32) Setting thresholds TBI1 and TBI2, where TBI1 < TBI2, the sub-images are divided into three categories according to the background change index BI:

If BI≤TBI1, it is a uniform background clutter class;

If TBI1 < BI ≤ TBI2, it is a general uneven background clutter class;

If TBI2<BI, it is a very uneven background clutter class;

(4) Under the GPU platform, the GPU sequentially processes the three types of sub-images according to their corresponding constant false alarm detection thresholds T1 to obtain a target area, and the three types of sub-images respectively use different processing algorithms to calculate a threshold T1. .

Further, in the step (1), the setting method of the boundary curve is:

(11) Initialize the boundary curve C, define the horizontal set function Φ of the region in the boundary curve C, set the narrow band radius, take the point on the boundary curve C as the center, and the narrow band radius as the radius to obtain the narrow band region;

(12) Calculate the minimum value of the energy function of the boundary curve C, and use the Hai's function and the Dirich impact function to obtain the solution of the partial differential equation as:

Where Φ ₀ (x, y) is the level set function of the initial boundary curve C; H(Φ) is the Hai's function, I(x, y) is the image in the narrow-band region, μ, ν, λ ₁ , λ ₂ Representing energy weights separately;

(13) Substituting all the points in the narrowband region into the level set function Φ ₀ (x, y) = 0 of the initialization boundary curve C, and evolving into a new boundary curve, and calculating the level set function of the new boundary curve as Φ ¹ ;

(14), the boundary curve is evolved n times in succession until all points on the image are traversed, and the boundary between land and sea is obtained.

(15), with the boundary between land and sea

Separation of land and sea for the boundary, eliminating land data, and obtaining images of marine areas with effective targets.

Further, in the step (11), the boundary curve is initialized according to the solution short time interval equation |▽T|F=1, where T(x, y, z) is a given point (x, y, z) to the boundary curve. The contraction time, F is the speed parameter. In the initial curve contour, the set speed parameter F is 1, and the point where the distance boundary curve C is equal to or smaller than 1 forms a region to be inspected, and the boundary of the region to be inspected is the boundary curve. C.

Further, in the step (12), the Euler-Lagrangian method is used to solve the minimum value of the energy function of the boundary curve C,

Where L(C) is the length of the closed curve C and S _b (C) is the area of the inner area of the curve C.

Further, in the step (12), the solution of the partial differential equation can be obtained

The iteration formula is:

among them,

The set of functions of the level set function at (x, y),

For forward differential operation.

Further, in the step (31), the calculation method of the background change index BI of the sub-image is:

Where m is the number of pixels included in each sub-image.

Further, in the step (21), the calculation method of the gray threshold T is:

(211), dividing the total gray level of the image of the ocean area into L levels, the total number of pixels of the image of the ocean area is n, and the number of pixels of the k-th gray level is n _k , then the normalization of the k-th gray level The histogram is: p(k)=n _k /n(k=0,1,2...,L-1);

(212), setting the ratio of the candidate target area to

Bring in

Get T.

Further, in the step (4), under the GPU platform, the GPU sequentially processes the three types of sub-images separately:

(41) Initialize the GPU: start CUDA by the CPU, set GPU related parameters, allocate data memory space, and initialize the input sub-image;

(42) Reading the sub-image into the GPU memory: in the CUDA framework, allocating the memory and reading the sub-image from the memory into the GPU memory;

(43) The GPU starts multi-threading and runs the kernel function: the CPU first loads the first type of threshold algorithm into the GPU, and as a multi-threaded kernel function, calculates a threshold, and uses the threshold as T1, which belongs to the first sub-image. A sub-image of a type is used for target detection, and the detection result is returned to the memory and copied to the memory. Secondly, the CPU loads the second type of threshold algorithm into the GPU, calculates the threshold, and uses the threshold as T1 as a multi-threaded kernel function. Target detection is performed on the sub-images belonging to the second category in all the sub-images, and the detection result is returned to the video memory and copied to the memory; again, the CPU loads the third-level threshold algorithm into the GPU as a multi-threaded kernel function, and calculates Threshold value, and using the threshold as T1, performing target detection on sub-images belonging to the second category in all sub-images, Return the test results to the video memory and copy them to memory.

(44) Release GPU resources: When the program is executed, release the GPU memory, recycle the GPU resources, and exit the program.

Further, the first type of sub-images are uniform background clutter, and a threshold is calculated by using a Gaussian distribution statistical model;

The sub-image images of the second type are generally uneven background clutter, and the threshold is calculated by using a Weibull distribution statistical model;

The third type of sub-images are extremely uneven background clutter, and the G ⁰ distribution model is used to calculate the threshold.

Compared with the prior art, the advantages and positive effects of the present invention are: the method for quickly detecting the target image of the synthetic aperture radar of the present invention, firstly separating the land from the ocean region, filtering out the image of the land portion, and improving the detection efficiency; secondly, Perform preliminary statistics on the graph, set appropriate global thresholds, perform preliminary screening on SAR image targets, and segment the image into several sub-image blocks. Finally, use CUDA technology to perform constant false alarm detection on the three distributed graphs to detect effective Ship target.

Other features and advantages of the present invention will become more apparent from the detailed description of the embodiments.

DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the description of the prior art will be briefly described below. Obviously, the drawings in the following description are only It is a certain embodiment of the present invention, and other drawings can be obtained from those skilled in the art without any creative work.

1 is a flow chart showing an embodiment of a method for quickly detecting a synthetic aperture radar image object according to the present invention.

detailed description

The technical solutions in the embodiments of the present invention are clearly and completely described in the following with reference to the accompanying drawings in the embodiments of the present invention. It is obvious that the described embodiments are only a part of the embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts are within the scope of the present invention.

Embodiment 1 This embodiment provides a method for quickly detecting a target image of a synthetic aperture radar, including the following steps:

S1, sea-land separation step, setting boundary curve, and separating sea and land by boundary curve, and obtaining ocean area image with effective data only in marine area;

Generally, land has strong scattering, and it is a bright region in SAR images, which has a great influence on ship target detection. Step S1 separates the land and sea, removes the land area, reduces the influence of the land area on the target detection, and reduces the calculation amount, which is beneficial to improve the speed and accuracy of the target detection.

S2, target screening steps, including:

S21, setting a gray threshold T, assigning an index value of a pixel whose gray value is greater than T in the image of the ocean area to a gray value of the pixel, and otherwise assigning a value of 0, and establishing an index matrix for all the obtained index values;

S22. Set a non-zero region in the index matrix as a candidate target region.

S23. The ocean area image is divided into several sub-images by using the position of the candidate target area, and each candidate target area corresponds to one sub-image;

Since image segmentation is the basis and premise of SAR image interpretation application, step S2 is based on global threshold image segmentation, and the SAR image is segmented into several sub-images to form the basis for the identification of ship targets.

S3. Set a background clutter statistical model, including:

S31, respectively calculating a background change index BI of each sub-image;

S32. Set thresholds TBI1 and TBI2, where TBI1<TBI2, and divide the sub-image into three categories according to the background change index BI:

If BI≤TBI1, it is a uniform clutter class;

If TBI1 < BI ≤ TBI2, it is a general heterogeneous clutter class;

If TBI2 < BI, it is a very uneven clutter class;

The clutter statistical model in the background area is a key factor in determining the performance of the detection algorithm. Due to the relatively variable sea surface conditions, the statistical characteristics of clutter are very complicated. If the statistical model does not describe the clutter characteristics well, it will cause the performance of the constant false alarm detector to deteriorate. The existing constant false alarm target detection algorithm generally adopts global modeling, and uses the same background clutter distribution model for all regions, resulting in the use of the model in the unused area mismatch, making detection The performance is significantly reduced. In order to improve the detection performance, the detection method of the present embodiment fully considers the advantages and disadvantages of each statistical model based on the in-depth analysis of the constant false alarm detection based on different statistical distribution models, and combines the constant false alarm detection algorithm according to the mean value of the SAR sub-image. And variance, the SAR sub-image is divided into three types: uniform background clutter, general uneven background clutter and extremely uneven background clutter. For these three different types, the constant false alarm detection algorithm suitable for this kind of characteristics is adopted to improve the detection accuracy.

S4. Under the GPU platform, the GPU sequentially processes the three types of pixel units according to the threshold T1 to obtain a target area, and the three types of pixel units respectively use different processing algorithms to calculate the threshold T1.

In the graphics processor (GPU) architecture, the Unified Computing Device Architecture (CUDA) technology is adopted, and the algorithm implementation based on three different distributions is optimized according to the characteristics of the GPU, and an efficient constant false alarm target detection algorithm is realized. The CPU implementation greatly shortens the data processing time and can meet the real-time requirements of SAR target detection.

As a preferred embodiment, in the step S1, the setting method of the boundary curve is:

S11, initializing the boundary curve C, defining a horizontal set function Φ of the region in the boundary curve C, setting a narrow band radius, centering on a point on the boundary curve C, and taking a narrow band radius as a radius to obtain a narrow band region;

S12. Calculate the minimum value of the energy function of the boundary curve C, and use the Hai's function and the Dirich impact function to obtain the solution of the partial differential equation as:

S13. Substituting all points in the narrowband region into the level set function Φ ₀ (x, y)=0 of the initialization boundary curve C, and evolving into a new boundary curve, and calculating a level set function of the new boundary curve is Φ ¹ ;

S14. The boundary curve is evolved n times in succession until all points on the image are traversed, and the boundary between the land and the sea is obtained.

S15, the boundary between land and sea

For sea and land separation for the boundary, the land data is eliminated, and the ocean area image with valid data only in the ocean area is obtained.

The detection method of the present embodiment simplifies the initial evolution curve by using the initial boundary parameters under specific conditions on the basis of the advantages of the narrow-band solution in the analysis level set method and the Mumford-Shah model, thereby narrowing the solution and Mumford in the level set method. The -Shah model is effectively combined to quickly obtain the separation effect of the land and sea areas.

Further, in the step S11, the boundary curve is initialized according to the solution short time interval equation |▽T|F=1, where T(x, y, z) is the contraction of the given point (x, y, z) to the boundary curve. Time, F is the speed parameter. Since the characteristics of F and the image are independent of each other, in the initial curve contour, the set speed parameter F is set to 1, and the point where the distance boundary curve C is equal to or smaller than 1 forms a to-be-detected area. The boundary of the area to be inspected is the boundary curve C.

In the step S12, the Euler-Lagrangian method is used to solve the minimum value of the energy function of the boundary curve C,

Further, in the step S12, the solution of the partial differential equation can be obtained.

The iteration formula is:

among them,

The set of functions of the level set function at (x, y),

For forward differential operation.

Further, in the step S31, the calculation method of the background change index BI of the sub-image is:

Where m is the number of pixels included in each sub-image. Since the variance is a measure of the degree of background change, in the SAR image, because of the multiplicative noise, the variance can not accurately represent the degree of background change. Therefore, the detection method of this embodiment introduces each of the background change indices BI. The SAR sub-image has m pixels and calculates BI separately.

In the step S21, the calculation method of the gray threshold T is:

S211. The total gray level of the image of the ocean area is divided into L levels, the total number of pixels of the image of the ocean area is n, and the number of pixels of the kth gray level is n _k , and the normalized square of the kth gray level is obtained. The picture is: p(k)=n _k /n(k=0,1,2...,L-1);

S212, setting the ratio of the candidate target area to

Bring in

Get T.

In the step S4, under the GPU platform, the GPU sequentially processes the three types of pixel units separately:

S41. Initializing the GPU: starting CUDA by the CPU, setting GPU related parameters, allocating data memory space, and initializing the input sub-image;

S42, reading the sub-image into the GPU memory: in the CUDA framework, allocating the memory, and reading the sub-image from the memory into the GPU memory;

S43. The GPU starts multi-threading and runs the kernel function: the CPU first loads the threshold algorithm of the first type into the GPU, and uses the kernel function of the multi-thread to calculate the threshold, and uses the threshold as the T1, and belongs to the first in all the sub-images. The sub-image of the class performs target detection, and returns the detection result to the video memory and copies it to the memory. Secondly, the CPU loads the second type threshold algorithm into the GPU, calculates the threshold, and uses the threshold as T1 as a multi-threaded kernel function. Perform target detection on the sub-images belonging to the second category in all sub-images, return the detection results to the video memory and copy them to the memory; again, the CPU loads the third-level threshold algorithm into the GPU as a multi-threaded kernel function to calculate the threshold. And using the threshold as T1, performing target detection on the sub-images belonging to the second category in all the sub-images, returning the detection result to the video memory and copying it to the memory.

S44. Release the GPU resource: when the program is executed, release the GPU memory, recycle the GPU resource, and exit the program.

In this embodiment, the sub-images of the first type are uniform clutter types, and a threshold is calculated by using a Gaussian distribution statistical model;

The sub-image images of the second type are generally non-uniform clutter, and the threshold is calculated by using a Weibull distribution statistical model;

The sub-images of the third type are extremely heterogeneous clutter types, and the threshold is calculated using a G ⁰ distribution model.

Specifically, for the uniform clutter background SAR image, the constant false alarm detection algorithm based on Gaussian distribution is used to solve the mean μ _{m of the} mixed Gaussian model according to the EM algorithm:

The μ _{m is} taken into the following formula to calculate the detection threshold:

For the general inhomogeneous clutter background SAR image, the constant false alarm detection algorithm based on Weibull distribution is used, and the joint probability density of N independent reference units is assumed to be

Where B is the scale parameter and C is the shape parameter. After f(x) takes the logarithm, the scale parameter and the shape parameter are separately derived to obtain:

Bring B and C into the false alarm probability formula to get the detection threshold:

T _I = B(-lnP _fa ) ^1/C

Clutter for SAR images is uneven, CFAR detection algorithm using G ⁰ distribution, using SKS estimation method, the distribution of the parameter estimates by G ^0, the expression is:

Where n is the equivalent visual number, α is the shape parameter, γ is the scale parameter, and Ψ(·) is the digamma function.

The cumulative amount of the sample logarithm. From the above expression, n, α, γ can be calculated to obtain a probability density expression.

Given the false alarm rate P _fa , can be formulated

Solve the detection threshold T _I . For the G ⁰ distribution, the above integral formula cannot obtain an analytical expression. To this end, the present invention solves by the following method:

(a) order

Initialize the minimum value m=min(I), the maximum value n=max(I), the loop variable n=0, the maximum number of loops N, and the precision ξ;

(b)

If |F(ζ)-(1-P _fa )|≤ξ, execute (d); otherwise, execute (c);

(c) If n < N, perform (d), otherwise, when F (ζ) < 1-P _fa , m = ζ; when F (ζ) > 1-P _fa , n = ζ; then execute ( b);

(d) T _I = ζ, exit the loop.

The above description is not intended to limit the present invention, and the present invention is not limited to the above examples, and variations, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention are also The scope of protection of the present invention.

Claims

A method for rapidly detecting a target image of a synthetic aperture radar, characterized in that it comprises the following steps:

(1) The sea-land separation step, evolving the boundary curve, and separating the sea and land with the boundary curve as the boundary, and obtaining the image of the marine area with effective targets;

(2) Target screening steps, including:

(21) setting a gray threshold T, assigning an index value of a pixel whose gray value is greater than T in the image of the ocean area to a gray value of the pixel, otherwise assigning a value of 0, and establishing all index values obtained An index matrix;

(22) setting a non-zero region in the index matrix as a candidate target region;

(23) dividing the image of the ocean area into a plurality of sub-images by using a position of the candidate target area, each candidate target area corresponding to one sub-image;

(3) Set the background clutter statistical model, including:

(31), respectively calculating the background change index BI of each sub-image;

(32) Setting thresholds TBI1 and TBI2, where TBI1 < TBI2, the sub-images are divided into three categories according to the background change index BI:

If BI≤TBI1, it is a uniform background clutter class;

If TBI1 < BI ≤ TBI2, it is a general uneven background clutter class;

If TBI2<BI, it is a very uneven background clutter class;

(4) Under the GPU platform, the GPU sequentially processes the three types of sub-images according to their corresponding constant false alarm detection thresholds T1 to obtain a target area, and the three types of pixel units respectively use different processing algorithms to calculate a threshold T1. .
The method for rapidly detecting a synthetic aperture radar image object according to claim 1, wherein in the step (1), the setting method of the boundary curve is:

(11) Initialize the boundary curve C, define the horizontal set function Φ of the region in the boundary curve C, set the narrow band radius, take the point on the boundary curve C as the center, and the narrow band radius as the radius to obtain the narrow band region;

(12) Calculate the minimum value of the energy function of the boundary curve C, and use the Hai's function and the Dirich impact function to obtain the solution of the partial differential equation as:

Where Φ 0 (x, y) is the level set function of the initialization boundary curve C; c 1 and c 2 respectively represent the gray average values of the two regions inside and outside the boundary curve, H(z) is the Hai's function, I(x , y) is an image in a narrow band region, and μ, ν, λ 1 , λ 2 respectively represent energy weights;

(13) Substituting all the points in the narrowband region into the level set function Φ 0 (x, y) = 0 of the initialization boundary curve C, and evolving into a new boundary curve, and calculating the level set function of the new boundary curve as Φ 1 ;

(14), the boundary curve is evolved n times in succession until all points on the image are traversed, and the boundary between land and sea is obtained.

(15), with the boundary between land and sea
For sea and land separation for the boundary, the land data is eliminated, and the ocean area image with valid data only in the ocean area is obtained.
The method for rapidly detecting a synthetic aperture radar image object according to claim 2, wherein in the step (11), the boundary curve is initialized according to the solution short time interval equation |▽T|F=1, wherein T(x, y, z) is the contraction time of the given point (x, y, z) to the boundary curve, F is the speed parameter, in the initial curve contour, the set speed parameter F is 1, and the distance boundary curve C is equal to or less than 1. The dots form a region to be inspected, and the boundary of the region to be inspected is the boundary curve C.
The method for rapidly detecting a synthetic aperture radar image object according to claim 3, wherein in the step (12), the Euler-Lagrangian method is used to solve the minimum value of the energy function of the boundary curve C:

Where L(C) is the length of the closed curve C and S b (C) is the area of the inner area of the curve C.
The method for rapidly detecting a synthetic aperture radar image object according to claim 4, wherein in the step (12), the solution of the partial differential equation is obtained
The iteration formula is:

among them,

The set of functions of the level set function at (x, y),
For forward differential operation.
The method for rapidly detecting a synthetic aperture radar image object according to claim 1, wherein in the step (31), the calculation method of the background change index BI of the sub-image is:

Where m is the number of pixels included in each sub-image.
The method for quickly detecting a synthetic aperture radar image object according to any one of claims 1 to 6, wherein in the step (21), the calculation method of the grayscale threshold T is:

(211), dividing the total gray level of the image of the ocean area into L levels, the total number of pixels of the image of the ocean area is n, and the number of pixels of the k-th gray level is n k , then the normalization of the k-th gray level The histogram is: p(k)=n k /n(k=0,1,2...,L-1);

(212), setting the ratio of the candidate target area to
Bring in
Get T.
The method for quickly detecting a synthetic aperture radar image object according to any one of claims 1 to 6, wherein in the step (4), under the GPU platform, the GPU sequentially processes the three types of sub-images separately. The method is:

(41) Initialize the GPU: start CUDA by the CPU, set GPU related parameters, allocate data memory space, and initialize the input sub-image;

(42) Reading the sub-image into the GPU memory: in the CUDA framework, allocating the memory and reading the sub-image from the memory into the GPU memory;

(43) The GPU starts multi-threading and runs the kernel function: the CPU first loads the first type of threshold algorithm into the GPU, and as a multi-threaded kernel function, calculates a threshold, and uses the threshold as T1, which belongs to the first sub-image. A sub-image of a type is used for target detection, and the detection result is returned to the memory and copied to the memory. Secondly, the CPU loads the second type of threshold algorithm into the GPU, calculates the threshold, and uses the threshold as T1 as a multi-threaded kernel function. Target detection is performed on the sub-images belonging to the second category in all the sub-images, and the detection result is returned to the video memory and copied to the memory; again, the CPU loads the third-level threshold algorithm into the GPU as a multi-threaded kernel function, and calculates The threshold value is used as the T1, and the sub-images belonging to the second category in all the sub-images are subjected to target detection, and the detection result is returned to the video memory and copied to the memory.

(44) Release GPU resources: When the program is executed, release the GPU memory, recycle the GPU resources, and exit the program.
The method for rapidly detecting a synthetic aperture radar image object according to claim 8, wherein the first type of sub-images are uniform background clutter, and a Gaussian distribution statistical model is used to calculate a threshold;

The sub-image images of the second type are generally uneven background clutter, and the threshold is calculated by using a Weibull distribution statistical model;

The third type of sub-images are extremely uneven background clutter, and the G 0 distribution model is used to calculate the threshold.