CN112766032A - SAR image saliency map generation method based on multi-scale and super-pixel segmentation - Google Patents

Abstract

The invention provides a method for detecting a significant target of an SAR (synthetic aperture radar) image based on multi-scale superpixels, belonging to the field of signal processing, in particular to the field of synthetic aperture radar image feature extraction. The method comprises the frequency characteristic of a super-pixel level, the local contrast of the super-pixel level and the global contrast of the super-pixel level. And linearly integrating the same features under different scales to obtain the salient feature maps of the features, so that the superpixel features obtained under multiple scales can participate in the fusion of the feature maps. And finally multiplying the characteristic saliency maps to obtain a final saliency target detection result map. Finally, simulation experiments are carried out on SAR images in different scenes, and compared with various classical saliency detection algorithms and a two-parameter CFAR detection method, the method proves that the algorithm has a remarkable target detection effect superior to that of a comparison algorithm under most conditions, and the inhibition effect on background clutter is higher than that of the comparison algorithm; under comprehensive evaluation, the SAR image significant target detection method has a good SAR image significant target detection effect.

Description

SAR image saliency map generation method based on multi-scale and super-pixel segmentation

Technical Field

The invention belongs to the field of signal processing, in particular to the field of synthetic aperture radar image feature extraction.

Background

Synthetic Aperture Radar (SAR) is a microwave active high-resolution imaging Radar, and can overcome the influence of extreme weather such as thunderstorm, snowstorm and the like, and realize all-day and all-weather active operation tasks. By utilizing the powerful penetration capability of electromagnetic waves, the SAR can detect and detect targets hidden under vegetation and ruins, obtain large-area ground surface high-resolution images in a short time and provide help which cannot be achieved by optical images and infrared. Therefore, the SAR plays an indispensable role in civilian aspects such as disaster detection, agricultural investigation, environmental monitoring, building surveying and mapping and military fields such as military target identification, battlefield armed investigation and ground strike due to unique advantages

High resolution and wide imaging area are main advantages of the SAR image, but with the development of the SAR imaging technology, High latitude (High-dimension), High resolution (High-resolution) and Multi-mode (Multi-mode) become main development trends of the SAR image. The SAR image has been developed from the first dozens of pixels to tens of thousands of counting units, the texture information, the shape information and the edge information are more and more abundant, and meanwhile, the wide imaging range causes the image to contain more and more target types and the scene to be more and more complex. It is obvious that these large and complex data cannot be processed efficiently by means of conventional manual methods.

Many other studies have been to represent images by changing the resolution of the image, i.e. by down-sampling, to form a gaussian pyramid structure. Every time the original image is subjected to the down-sampling operation, image information is lost 3/4, and the gaussian pyramid structure including the original image generally has nine layers, so that a large amount of target information is lost when the down-sampling operation is completed.

Disclosure of Invention

The invention aims to solve the technical problem that the prior art can not quickly and accurately locate and find a special area in a scene for a large number of images.

In order to extract the salient object more accurately and effectively, original image information can be reserved while multi-scale local information is obtained, and the information integrity of the multi-scale image is guaranteed. Firstly, segmenting an image in different scales by utilizing an SLTC algorithm; the SLTC algorithm is a simple linear iterative clustering algorithm for short, and is a super-pixel segmentation method utilizing the relation between color similarity and spatial distance. The method can generate compact and uniform superpixel blocks, has higher comprehensive evaluation in the aspects of operation speed, edge adhesion, segmentation shape and the like, and can obtain satisfactory effect; and then, on the basis of studying the ITTI algorithm, the biological mechanism of non-uniform multi-scale sampling of the human eyes on the visual image is learned, and the defect that the shape characteristics of the target contour are lost in the ITTI algorithm is improved by utilizing the good adhesion of the superpixel segmentation methods with different scales and sizes to the edge. Meanwhile, based on the inherent characteristics of the SAR image, the SAR image is subjected to characteristic extraction suitable for the SAR image by using the superpixel as a basic unit without using simple characteristics, wherein the characteristic extraction comprises superpixel-level frequency characteristics, superpixel-level local contrast and superpixel-level global contrast. The method comprises the following steps of performing linear integration on the same characteristic under different scales to obtain each characteristic saliency map, and ensuring that superpixel characteristics obtained under multiple scales can participate in fusion of the characteristic maps, so that the technical scheme of the invention is a SAR image saliency target detection method based on multi-scale superpixels, and the method comprises the following steps:

step 1: segmenting the multi-scale superpixel image;

presetting M scales, namely a scale 1-a scale M, by using an SLIC algorithm (linear iterative clustering algorithm), wherein the M is more than or equal to 1; respectively obtaining super pixel blocks under different scales;

SLIC segmentation is carried out on the input image under the scale 1 to obtain N₁A plurality of superpixel blocks;

SLIC segmentation is carried out on the input image under the scale 2 to obtain N₂A plurality of superpixel blocks;

······

SLIC segmentation is carried out on the input image under the scale M to obtain N_MA plurality of superpixel blocks;

in the above way, the superpixel segmentation results under different scales are obtained;

step 2: extracting features based on the superpixel blocks;

step 2.1: extracting frequency characteristics of the superpixel blocks;

under the scale 1, each super pixel block in the image after the SLIC algorithm processing is represented as S_{1_i}，i＝1,2,...,N₁(ii) a With equation (1), the frequency of the superpixel block per block is obtained:

wherein, I_fre(m, n) denotes the frequency of the superpixel block, λ denotes the empirical parameter, fre_(m,n)Representing the frequency of the gray value of a pixel point with the coordinate of (m, n) in the SAR image appearing in the image, n_(m,n)Representing the number of the points corresponding to the gray value appearing in the image, and N representing the total number of the pixel points of the input SAR image; and calculating a frequency characteristic value by using the following formulas (2) and (3):

num(1_i)＝Number(S_{1_i})/4，i＝1,2，...,N₁ (2)

wherein num (1_ i) represents taking S_{1_i}1/4 pixels with the largest intra-pixel frequency eigenvalue, Number (S)_{1_i}) Representing the ith super pixel S_{1_i}The frequency characteristic values in the sequence are sorted from large to small and the pixel number is calculatedAmount f_frep(S_{1_i}) Represents a pair S_{1_i}The maximum 1/4 pixel frequency characteristic value in the average is taken as S_{1_i}The super-pixel frequency characteristic value;

to N sequentially₁Extracting frequency characteristics of the super pixel blocks to obtain N₁Dimension feature vector f_{frep_1}(S_{1_1}), f_{frep_1}(S_{1_2}),......,f_{frep_1}(S_{1_N1})](ii) a For the ith super-pixel block, the frequency characteristic of each pixel inside the super-pixel block is f_frep(S_{1_i}) Represents; obtaining a frequency characteristic diagram Freq _ map (1) under the scale 1;

similar to the above steps, obtaining a frequency characteristic map of scale 2, Freq _ map (2);

................................

obtaining a frequency characteristic diagram of a scale M, Freq _ map (M);

step 2.2: extracting local contrast characteristics of the superpixel blocks;

calculating a superpixel block S using the following equation (4)_{1_i}Local contrast frequency f_local(S_{1_i})：

Wherein S is_{1_j}Representation and superpixel block S_{1_i}All super pixels in the neighborhood; for enhancing algorithm robustness, super-pixel block S_{1_i}The gray values of the inner pixels are sorted from large to small, the largest 1/8 pixels are selected and averaged, and the peak characteristic f of each super pixel is obtained_peak(S_{1_i}) The mean value of the gray values of the pixels in each super pixel is f_mean(S_{1_i}) Finally, the local contrast frequency f is obtained according to the formula (4)_local(S_{1_i})

Based on the segmentation result of the scale 1, S is sequentially paired_{1_i}Performing local contrast feature extraction to obtain N₁Dimensional local contrast feature vector [ f ]_{local_1}(S_{1_1}),f_{local_1}(S_{1_2}),......,f_{local_1}(S_{1_N1})](ii) a For the ith super-pixel block, the local contrast characteristic frequency of each pixel inside the super-pixel block is f_local(S_{1_i}) Represents; after normalization processing is carried out, a Local contrast characteristic map (1) under the scale 1 is obtained;

similarly, with the steps, Local _ map (2) based on scale 2 is obtained;

........................

obtaining a Local _ map (M) based on the scale M;

step 2.3, extracting the global contrast characteristic of the superpixel block;

superpixel block S_{1_i}The average value of the gray levels of all the pixels in the super pixel is used as the initial value f of the super pixel_i(ii) a The area S is calculated by the following method_iGlobal contrast characteristic value f_glob(S_{1_i})：

Wherein, G (S)_{1_i}) Representing the global contrast, f, of the i-th block superpixel at scale 1_jRepresents the mean value of the gray levels, num (S) of all the pixels in the jth super-pixel_{1_j}) Representing a super pixel area S_{1_j}The threshold value T is based on G (S)_{1_i}) An empirical value constant is selected for the distribution;

in sequence to S_{1_i}Performing feature extraction of global contrast to obtain N₁Dimensional global contrast feature vector f_{global_1}(S_{1_1}), f_{global_1}(S_{1_2}),......,f_{global_1}(S_{1_N1})](ii) a For the ith super-pixel block, the global contrast characteristic frequency of each pixel inside the super-pixel block is f_global(S_{1_i}) Represents; obtaining a global contrast characteristic map global _ map (1) under the scale 1;

similarly, with the steps, obtaining a global feature map Local _ map (2) based on the scale 2;

........................

obtaining global feature map (M) based on the scale M;

and step 3: multi-scale fusion;

the frequency saliency map SM is obtained by linearly superposing Freq _ map (1) to Freq _ map (M)_freq；

Performing linear superposition on Local _ map (1) -Local _ map (M) to obtain a Local feature saliency map SM_local；

Performing linear superposition on Global _ map (1) -Global _ map (M) to obtain a Global feature saliency map SM_global；

And 4, step 4: multiplying the three graphs obtained in the step 3 by using the following formula (8):

SM＝N(SM_freq·SM_local·SM_global) (8)

obtaining a final total saliency map SM; where N (-) represents the normalization operation.

The invention has the beneficial effects that:

the obvious target detection result obtained by the algorithm more accurately depicts the shape information of the target, and meanwhile, the suppression effect on the background clutter is obviously superior to that of other comparison algorithms, and the algorithm is greatly helpful for reducing the false alarm rate.

Drawings

Fig. 1 is a basic flow chart of the algorithm.

Fig. 2 shows an example of SLIC algorithm with different numbers, where (a) is a result graph with 100 superpixels, and (b) is a result graph with 200 superpixels.

Fig. 3 shows two examples of superpixel-based frequency profiles, where (a) is the original and (b) is the superpixel frequency profile.

Fig. 4 is a local contrast characteristic diagram based on super-pixels, in which (a) is an original image and (b) is a local contrast characteristic diagram of super-pixels.

Fig. 5 is a superpixel-based global contrast feature map, where (a) is the original image and (b) is the superpixel-based global contrast feature map.

FIG. 6 is a global contrast profile, where the top row is a tank SAR imaging picture from the M-Star database with a resolution of 128 × 128, and the bottom row is a port ship SAR picture from the SSDD database with a resolution of 366 × 345; wherein, (a) is the original image, and the upper and lower lines of the images (b) to (f) are respectively the saliency maps obtained by adopting different algorithms for the two images; the graph (b) is the standard significant target result of the artificial labeling, which is called a standard significant result Graph (GT), the graph (c) is the result of the significant target detection method using the invention, the graph (d) is the detection result of the RC algorithm, the graph (e) is the detection result of the ITTI algorithm, and the graph (f) is the detection result of the two-parameter CFAR algorithm.

Detailed Description

The flow of the algorithm is shown in fig. 1, and the specific implementation steps are as follows:

step 1: based on SLIC algorithm, multi-scale super-pixel segmentation is carried out on input image

M scales, namely the scale 1-the scale M, are preset, and M is more than or equal to 1.

SLIC segmentation is carried out on an input image under the scale 1 to obtain N₁A plurality of superpixel blocks;

performing SLIC segmentation on the input image under the scale 2 to obtain N₂A plurality of superpixel blocks;

······

SLIC segmentation is carried out on the input image under the scale M to obtain N_MA plurality of superpixel blocks;

in the above, the super-pixel segmentation results at M scales are obtained.

FIG. 2 is an example of superpixel segmentation at different scales; for the corresponding input image of the graph (a), take N₁The segmentation result is shown in fig. (a) as 100; for the corresponding input image of the image (b), take N₁The result of the segmentation is shown in fig. (b) at 200.

Step 2: feature extraction based on superpixel blocks;

the step is divided into three parallel submodules;

step 2.1: extracting the frequency characteristics of the superpixel segmentation;

sequentially comparing the superpixel blocks S obtained in the step 1 based on the segmentation result of the scale 1_{1_i}，i＝1,2,...,N₁Frequency feature extraction is carried out by using formulas (1), (2) and (3) to obtain an N1-dimensional frequency feature vector f_{frep_1}(S_{1_1}),f_{frep_1}(S_{1_2}),......, f_{frep_1}(S_{1_N1})]. For the ith (1. ltoreq. i. ltoreq.N)₁) A super-pixel block, the frequency characteristic of each pixel inside the super-pixel block is f_{frep_1}(S_{1_1}) And (4) showing. After normalization, the frequency feature map Frep _ map (1) at scale 1 is obtained.

Similarly, with the above steps, a frequency feature map Frep _ map (2) based on scale 2 is obtained.

........................

A frequency feature map Frep _ map (M) based on the scale M is obtained.

Step 2.2: local contrast feature extraction based on superpixel segmentation

Sequentially comparing the superpixel blocks S obtained in the step 1 based on the segmentation result of the scale 1_{1_i}，i＝1,2,...,N₁Using formulas (3) and (4) to extract the local contrast characteristics to obtain an N₁Dimensional local contrast feature vector [ f ]_{local_1}(S_{1_1}), f_{local_1}(S_{1_2}),......,f_{local_1}(S_{1_N1})]. For the ith (1. ltoreq. i. ltoreq.N)₁) A super-pixel block, wherein the local contrast characteristic of each pixel inside the super-pixel block is represented by f_{local_1}(S_{1_i}) And (4) showing. After normalization, Local contrast characteristic map Local _ map (1) at scale 1 is obtained.

Similarly, with the above steps, Local _ map (2) based on scale 2 is obtained.

........................

Local _ map (M) is obtained based on the scale M.

Step 2.3: superpixel-based global contrast feature extraction

Sequentially comparing the superpixel blocks S obtained in the step 1 based on the segmentation result of the scale 1_{1_i}And carrying out feature extraction of global contrast by using formulas (5), (6) and (7) to obtain N₁Dimensional global contrast feature vector f_{global_1}(S_{1_1}),f_{global_1}(S_{1_2}),......, f_{global_1}(S_{1_N1})]. For the ith (1. ltoreq. i. ltoreq.N)₁) A super-pixel block, wherein the global contrast characteristic of each pixel inside the super-pixel block is f_{global_1}(S_{1_i}) And (4) showing. Thus, a global contrast feature map global _ map (1) at scale 1 is obtained.

Similarly, with the above steps, a global feature map global _ map (2) based on scale 2 is obtained.

........................

Obtaining a global feature map (M) based on the scale M.

And step 3: extracting the feature graphs of the frequency feature, the local contrast feature and the global contrast feature respectively through the step 2, and then linearly overlapping the feature graphs of the same feature among different scales to avoid target omission caused by that the superpixel scale on a certain scale is not matched with the target and the target superpixel is not included, so that a frequency feature significant graph F based on superpixel segmentation is obtained_frepLocal contrast saliency map F_localGlobal contrast saliency map F_global

And 4, step 4: fusing the three feature maps by using the formula (8) to generate a final saliency map

The results shown in fig. 6 are obtained by performing simulation experiments under various algorithms on two sets of SAR images. As seen from fig. 6(c), the frequency of the highlighted pixel points is low, a complete target area is obtained, and the shape information of the target is accurately depicted, so that compared with other comparison algorithms, the algorithm provided by the invention has a better filtering effect on background noise, a lower false alarm result, and better display on the contour information of the target.

Claims (1)

1. A SAR image salient target detection method based on multi-scale superpixels comprises the following steps:

step 1: segmenting the multi-scale superpixel image;

presetting M scales, namely a scale 1-a scale M, by using a linear iterative clustering algorithm, wherein the M is more than or equal to 1; respectively obtaining super pixel blocks under different scales;

SLIC segmentation is carried out on the input image under the scale 1 to obtain N₁A plurality of superpixel blocks;

SLIC segmentation is carried out on the input image under the scale 2 to obtain N₂A plurality of superpixel blocks;

······

SLIC segmentation is carried out on the input image under the scale M to obtain N_MA plurality of superpixel blocks;

in the above way, the superpixel segmentation results under different scales are obtained;

step 2: extracting features based on the superpixel blocks;

step 2.1: extracting frequency characteristics of the superpixel blocks;

under the scale 1, each super pixel block in the image after the SLIC algorithm processing is represented as S_1-i，i＝1,2,...,N₁(ii) a With equation (1), the frequency of the superpixel block per block is obtained:

wherein, I_fre(m, n) denotes the frequency of the superpixel block, λ denotes the empirical parameter, fre_(m,n)Representing the frequency of the gray value of a pixel point with the coordinate of (m, n) in the SAR image appearing in the image, n_(m,n)Representing the number of the points corresponding to the gray value appearing in the image, and N representing the total number of the pixel points of the input SAR image; and calculating a frequency characteristic value by using the following formulas (2) and (3):

num(1_i)＝Number(S_{1_i})/4，i＝1,2，...,N₁ (2)

wherein num (1_ i) represents taking S_{1_i}1/4 pixels with the largest intra-pixel frequency eigenvalue, Number (S)_{1_i}) Representing the ith super pixel S_1-iThe frequency eigenvalues in (f) are sorted from large to small and the number of pixels is calculated_frep(S_{1_i}) Represents a pair S_{1_i}The maximum 1/4 pixel frequency characteristic value in the average is taken as S_{1_i}The super-pixel frequency characteristic value;

to N sequentially₁Extracting frequency characteristics of the super pixel blocks to obtain N₁Dimension feature vector f_{frep_1}(S_{1_1}),f_{frep_1}(S_{1_2}),……,f_{frep_1}(S_{1_N1})](ii) a For the ith super-pixel block, the frequency characteristic of each pixel inside the super-pixel block is f_frep(S_{1_i}) Represents; obtaining a frequency characteristic diagram Freq _ map (1) under the scale 1;

similar to the above steps, obtaining a frequency characteristic map of scale 2, Freq _ map (2);

…………………………

obtaining a frequency characteristic diagram of a scale M, Freq _ map (M);

step 2.2: extracting local contrast characteristics of the superpixel blocks;

calculating a superpixel block S using the following equation (4)_{1_i}Local contrast frequency f_local(S_{1_i})：

Wherein S is_1-jRepresentation and superpixel block S_1-iAll super pixels in the neighborhood; for enhancing algorithm robustness, super-pixel block S_1-iThe gray values of the inner pixels are sorted from large to small, and the largest gray value is selected1/8 number of pixels and averaging to obtain the peak feature f of each super-pixel_peak(S_{1_i}) The mean value of the gray values of the pixels in each super pixel is f_mean(S_{1_i}) Finally, the local contrast frequency f is obtained according to the formula (4)_local(S_{1_i})

Based on the segmentation result of the scale 1, S is sequentially paired_{1_i}Performing local contrast feature extraction to obtain N₁Dimensional local contrast feature vector [ f ]_{local_1}(S_{1_1}),f_{local_1}(S_{1_2}),……,f_{local_1}(S_{1_N1})](ii) a For the ith super-pixel block, the local contrast characteristic frequency of each pixel inside the super-pixel block is f_local(S_{1_i}) Represents; after normalization processing is carried out, a Local contrast characteristic map (1) under the scale 1 is obtained;

similarly, with the steps, Local _ map (2) based on scale 2 is obtained;

……………………

obtaining a Local _ map (M) based on the scale M;

step 2.3, extracting the global contrast characteristic of the superpixel block;

superpixel block S_{1_i}The average value of the gray levels of all the pixels in the super pixel is used as the initial value f of the super pixel_i(ii) a The area S is calculated by the following method_iGlobal contrast characteristic value f_glob(S_{1_i})：

Wherein, G (S)_1-i) Representing the global contrast, f, of the i-th block superpixel at scale 1_jRepresents the mean value of the gray levels, num (S) of all the pixels in the jth super-pixel_{1_j}) Representing a super pixel area S_{1_j}The threshold value T is based on G (S)_{1_i}) An empirical value constant is selected for the distribution;

in sequence to S_{1_i}Performing feature extraction of global contrast to obtain N₁Dimensional global contrast feature vector f_{global_1}(S_{1_1}),f_{global_1}(S_{1_2}),……,f_{global_1}(S_{1_N1})](ii) a For the ith super-pixel block, the global contrast characteristic frequency of each pixel inside the super-pixel block is f_global(S_1-i) Represents; obtaining a global contrast characteristic map global _ map (1) under the scale 1;

similarly, with the steps, obtaining a global feature map Local _ map (2) based on the scale 2;

……………………

obtaining global feature map (M) based on the scale M;

and step 3: multi-scale fusion;

the frequency saliency map SM is obtained by linearly superposing Freq _ map (1) to Freq _ map (M)_freq；

Performing linear superposition on Local _ map (1) -Local _ map (M) to obtain a Local feature saliency map SM_local；

Performing linear superposition on Global _ map (1) -Global _ map (M) to obtain a Global feature saliency map SM_global；

And 4, step 4: multiplying the three graphs obtained in the step 3 by using the following formula (8):

SM＝N(SM_freq·SM_local·SM_global) (8)

obtaining a final total saliency map SM; where N (-) represents the normalization operation.

Images