A SAR image target detection method based on sparse enhancement and Bayesian saliency

Abstract

The application provides a SAR image target detection method based on SAR image sparse enhancement and Bayesian saliency detection; comprising: step one: inputting a matched filter SAR image, obtaining initial values of regularization parameters of a SAR sparse enhancement method by calculating gray statistical characteristics of the image; step two: solving a convolution sparse feature enhancement algorithm model based on L1 norm by using an alternating direction multiplier method, and obtaining a SAR sparse enhancement image; step three: calculating a Bayesian saliency map of the SAR sparse enhancement image, binarizing the saliency map by using a threshold detection method, and realizing final target detection. The method can reduce SAR speckle noise by using a sparse image enhancement method, reduce background gray values, highlight target regions by using a SAR image saliency detection method, further reduce background pixel values, and finally realize efficient detection of ground and sea surface targets of interest.

Description

TECHNICAL FIELD

[0001] The present application relates to the SAR image enhancement and target detection technology, specifically to a SAR image target detection method based on sparse enhancement and Bayesian saliency. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) is an active microwave sensor, which can realize long-time, high-resolution observation of the target on the ground and sea surface in all-weather and all-day, and has a wide range of applications in military, civil and other fields. With the gradual maturity of SAR system, the resolution of the image obtained is getting higher and higher, and the imaging quality is getting better and better. The interpretation and identification of the target of interest in the SAR image have become an important content of SAR application. In the interpretation of SAR image, how to automatically detect high-value targets in a large scene SAR image is the frontiers and research difficulties of SAR application, which has attracted widespread attention from scholars at home and abroad. At present, the detection and identification technology of ship targets on the sea surface without land background and low sea conditions, and high-brightness vehicle targets with relatively pure ground background has gradually matured, and a large number of automatic target detection methods have appeared, mainly including Constant False Alarm Rate (CFAR) target detection method, target detection method based on visual attention model, target detection method based on complex image, and neural network target detection method based on data-driven (see Du Lan, Wang Zhaocheng, Wang Yan, et al. Research Progress of Single-channel SAR Target Detection and Identification in Complex Scene. Journal of Radars, 2020, Vol. 9, No. 1). The above methods have their own advantages and disadvantages. The SAR target detection based on CFAR mainly models the clutter, and rarely considers the characteristics of the target. The SAR image detection based on complex data mainly uses the target electromagnetic scattering information contained in the complex number, but the complex data SAR image is difficult to obtain. The neural network-based method needs to design a very complex target detection network in advance, and the timeliness is poor. The SAR image saliency detection method based on visual attention mechanism is a very important target detection method. Saliency detection only focuses on the more salient and prominent objects in the scene, and ignores other objects in the scene. The visual signals selected by this mechanism are sent to the high-level cortex of the brain, which can greatly speed up the brain's target detection and identification speed, so the target detection method based on saliency is an efficient and accurate method, which has been widely used in optical image detection. However, the traditional matched filter SAR image has a lot of speckle noise due to its special coherent imaging mechanism. In addition, when the background is complex, the matched filter SAR image is prone to have a large gray scale of background pixels, which leads to confusion between foreground and background, and the saliency extraction accuracy is greatly reduced. For the problem of SAR image saliency detection, scholars have proposed some improved methods.Wang et al. proposed a pattern recurrence saliency detector in 2016, which uses the intensity contrast of image patches to highlight the target while suppressing the coherent speckle noise (see Wang, Saliency detector for SAR images based on pattern recurrence, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 9, no. 7, 2016). Ni et al. proposed a SAR image target saliency detection method based on background context information in 2018, which adaptively selects background image patches based on the statistical mean of SAR image patches, and then uses the non-similarity construction of these background image patches and the current image patches to calculate the saliency map (see Ni, Background context-aware-based SAR image saliency detection, IEEE Geoscience and Remote Sensing Letters, vol. 15, no. 9, 2018). The above methods are all based on the traditional optical image saliency detection method and increase the SAR image-based processing unit, which increases the computational complexity and makes the method unable to be applied to large scene SAR images. At present, there is still a lack of detection means for suppressing non-uniform terrain / sea clutter, natural clutter and other target false alarm interference in large scene and complex background SAR images, and for real-time and accurate capture of high-value target detection.

[0003] Sparse SAR imaging method is a new type of SAR imaging method, which can obtain SAR images with lower sidelobes, smaller background noise and higher resolution compared to traditional matched filter SAR imaging method. The above characteristics are very advantageous for subsequent SAR image interpretation, especially SAR image target detection. To obtain sparse SAR images, a feasible method is to directly perform sparse processing on the basis of traditional matched filter SAR images, also known as SAR image sparse enhancement. This method takes the traditional matched filter SAR image as input, and through sparse regularization processing, it can obtain sparse SAR images with lower sidelobes, smaller background noise and higher resolution, without the need to change the SAR system hardware, and can be directly applied to existing SAR systems, with broad application prospects. At present, there are still few target detection methods based on sparse SAR images, and the fusion of sparse SAR image enhancement and saliency detection method is beneficial to greatly improve the target detection performance of existing saliency detection methods. SUMMARY

[0004] In view of the problems in the prior art, the present application provides a SAR image target detection method based on sparse enhancement and Bayesian saliency, which is realized through the following steps:

[0005] Step 01) input a matched filter SAR image, obtain the initial value of the regularization parameter of the SAR sparse enhancement method by calculating the gray statistical characteristics of the image; the specific process is as follows:

[0006] Step 01) input a matched filter SAR image Wherein, Y MF represents the SAR image obtained by matched filtering, which is referred to as "matched filter SAR image" for short, represents that the image size is MxN pixels; the image gray value is normalized to the range of [0, 255], that is, 8bit quantization; and then the gray histogram of the image is calculated;

[0007] Step 02) obtain the sparse rate of the image by using the gray histogram, and obtain the initial value of the regularization parameter by using the sparse rate;

[0008] The sparse rate of the matched filter SAR image is defined as sr=N hi / N MF , wherein N MF represents the number of all pixel points in the matched filter SAR image, N hi represents the number of pixel points with a gray value higher than 220; the greater the value of sr, the more the number of foreground pixel points in the image, and a smaller initial value of the sparse regularization parameter needs to be set; the smaller the value of sr, the more the number of background pixel points in the image, and a larger initial value of the sparse regularization parameter needs to be set to reduce the background noise intensity;

[0009] The initial value and variation of the sparse regularization parameter λ are defined as follows:

[0010]

[0011] Wherein, ini represents the initial value, k represents the iteration number of the SAR sparse enhancement method, and var(Y MF ) represents the variance of the input image;

[0012] Step 02) obtain the sparse rate of the image by using the gray histogram, and obtain the initial value of the regularization parameter by using the sparse rate;

[0013] Step 03) ADMM operator based on matrix multiplication, design L1 norm based convolution sparse feature enhancement algorithm model;

[0014] Firstly, a SAR image sparse enhancement model is constructed, and the model is based on the following expression:

[0015] Y MF = X + N (2)

[0016] wherein is a sparse enhancement result, with a size of MxN pixels; denotes the difference between the sparse enhancement result and the original image; X is solved by an L1 norm sparse regularization method; specifically, the estimated value of X is solved by the following sparse model

[0017]

[0018] wherein, denotes the estimated value of X, denotes the F norm of a matrix, || ||1 denotes the L1 norm, and λ is a regularization parameter; denotes the convolution operation of the matrix X and the matrix C, and the specific expression of C is:

[0019]

[0020] wherein c1∈[0.5, 1] is a constant between 0.5 and 1; the convolution operation is used to eliminate the scattered points that have been set to zero in the target; the sliding step is set to 1 in the convolution process, and the edge points are removed after the convolution, so that the size of the image after the convolution is consistent with that of the input image; ε represents a small constant, and ⊙ represents the Hadamard product of a matrix;

[0021] Step 04) calculating the specific expression of three-step iteration in the ADMM algorithm to obtain the SAR sparse enhancement result in one iteration process;

[0022] using the ADMM algorithm of matrix multiplication, using the iteration steps, and calculating the sparse enhancement image step by step auxiliary matrix Lagrange multiplier matrix Specifically, set X, Z, β, P c the initial value is wherein I is an identity matrix, and P c denotes a connected domain matrix, denotes the connected domain matrix in the kth iteration, which controls the termination condition of the algorithm; the iteration optimization expressions of the three matrices X, Z, and β are as follows:

[0023]

[0024]

[0025] β (k+1) = β (k) + ρ(X (k+1)-Z (k+1) ) (7)

[0026] where k denotes the iteration number, the upper index (k+1) denotes the corresponding value in the (k+1)th iteration; in formula (5), L(·) denotes the objective function, and p denotes a penalty factor for controlling the reconstruction accuracy and convergence speed of the algorithm; in formula (6), S(·;·) denotes an iterative soft threshold operation, and T denotes a threshold matrix, and the expression is as follows:

[0027]

[0028] where sign(·) denotes a sigmoid nonlinear function, |·| denotes taking an absolute value, and max(x, 0) denotes taking the maximum of x or 0; the constant e is a small constant;

[0029] Step three: calculating the Bayesian saliency map of the SAR sparse enhancement image, using a threshold detection method to binarize the saliency map, and realizing the final target detection;

[0030] Step 05) calculating a superpixel block using the SAR sparse enhancement result, and calculating a Bayesian prior saliency map based on the superpixel block;

[0031] The SAR sparse enhancement image is used as input to calculate the superpixel block, and a single linear iterative clustering (SLIC) algorithm is directly used for superpixel segmentation. The gray difference d i and the distance difference d s between the pixel point i and the superpixel seed point j are used to measure the similarity D ij between the pixel points, and then the SLIC algorithm is used for segmentation, and the specific measurement expression is as follows:

[0032]

[0033] where l i denotes the gray value at the pixel point i, (x i , y i ) denotes the position of the pixel point i; the segmented superpixel s i is obtained through the SLIC algorithm, i=1, 2,..., N s ; and N s is the number of superpixel blocks;

[0034] After s i is obtained, the Bayesian prior saliency map P(sal|s i ) of the superpixel block is calculated as follows:

[0035]

[0036] where ωpos (s i , s j ) represents the distance weight, d int (s i , s j ) represents the Euclidean distance value of all pixel points in the superpixel block s j from the center pixel point of s i , and the value is normalized to [0, 1]; the distance parameter σ pos controls the contribution degree of the distance between superpixels; d shape (s i ) represents the number of all pixel points in s i ;

[0037] After calculating P(sal|s i ), the minimum-maximum scale method is used to obtain the normalized prior saliency map The formula is:

[0038]

[0039] Wherein, P min (sal|s i ) represents the minimum saliency map value, and P max (sal|s i ) represents the maximum saliency map value;

[0040] After obtaining the normalized prior saliency map of all superpixel blocks by formula (11), all are combined into the prior saliency map of the whole image

[0041] Step 06) judge the change of the connected domain with the largest number of pixel points in the top 10 of the normalized prior saliency map When the pixel points and the corresponding pixel value change range of the 10 connected domains are less than the set threshold, the final Bayesian saliency map is calculated by using the normalized prior saliency map ;

[0042] First, the normalized prior saliency map is used to calculate whether the superpixel of the image belongs to the foreground fg or the background bg, and the mean value of the pixel value of the prior saliency map is used as the threshold for segmentation. If the pixel value of the normalized prior saliency map of the superpixel is greater than the mean value, it is considered that the superpixel belongs to the foreground region fg, otherwise it belongs to the background, and the specific formula is as follows:

[0043]

[0044] Wherein s idenotes the i-th superpixel block, and mean(·) denotes the mean value of the gray scale of the prior salient pixels in the superpixel block;

[0045] Further, the top 10 largest connected domains in the foreground fg are calculated, and the pixel values of the remaining smaller connected domains are set to 0, and the processed foreground region is counted as is an M*N matrix;

[0046] Determination: if or the iteration number k>N s ,

[0047] Then, k=k+1, and the calculation formula (5) is returned;

[0048] If the feature enhancement result X is output, and the normalized prior saliency map of the whole image is

[0049] The final Bayesian prior saliency map is calculated, and for a pixel point z in the whole SAR image, the final saliency map value of the point is calculated by using the following formula:

[0050]

[0051] where P(z|sal) and P(z|bg) respectively represent the posterior probabilities of the salient region and the background region, N fg (z) represents the number of pixel points in the background region bg obtained by using formula (12) and having the same gray scale value as the z point, N bg (z) represents the number of pixel points in the foreground region fg and having the same gray scale value as the z point; N fg and N bg respectively represent the total number of pixel points in the foreground and the background in the SAR image after the processing by formula (12).

[0052] Step 07) The Bayesian saliency Figure Two value is detected by using a threshold detection method, and the final target detection result is obtained.

[0053] Specifically, a global threshold Tz based on P(sal|z) is set, when the Bayesian saliency map value P(sal|z) of the pixel point z in the image is greater than the threshold, the pixel value of the z point is set to 1, that is, the pixel position of 1 is the target region, otherwise, it is set to 0, and the pixel position of 0 is the background region; a final binary image, that is, a target detection result, is obtained.

[0054] In one specific embodiment of the present application, in Step 04), p=0.99, and e -6 .

[0055] In another specific embodiment of the present application, in Step 05),

[0056] The present application aims to combine the sparse SAR image enhancement and SAR image saliency detection method, and proposes a SAR image target detection method based on sparse enhancement and Bayesian saliency. The sparse image enhancement method is used to reduce the SAR coherent speckle noise and reduce the background gray value. The SAR image saliency detection method is used to highlight the target area and further reduce the background pixel value, and finally realize the efficient detection of the ground sea surface target of interest.

[0057] The present application has the beneficial effects that: a convolution L1 norm SAR sparse enhancement method is proposed, which uses convolution operation to eliminate the hollow points and non-continuous edges of the target area, and lays a good foundation for subsequent calculation of image superpixel blocks and Bayesian prior saliency map. A Bayesian saliency map calculation method based on SAR sparse enhancement image is proposed, which can effectively improve the performance of the existing saliency detection method. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The algorithm flowchart of the method is shown in Figure 1.

[0059] Figure 2 The original matched filter SAR image and its corresponding gray histogram are shown in Figure 2, wherein the left column is three matched filter SAR images, and the right column is the histogram corresponding to the left image.

[0060] Figure 3 The processing results of the proposed method and the comparative method on a sea surface scene SAR image are shown in Figure 3; wherein (a)-(d): matched filter SAR image, corresponding superpixel segmentation result, Bayesian prior saliency map, Bayesian saliency map; (e)-(h): image enhancement result based on the traditional L1 sparse enhancement method, λ=50, corresponding superpixel segmentation result, Bayesian prior saliency map, Bayesian saliency map; (i)-(l): image enhancement result based on the traditional L1 sparse enhancement method, λ=130, corresponding superpixel segmentation result, Bayesian prior saliency map, Bayesian saliency map; (m)-(p): image enhancement result based on the L1 ADMM sparse enhancement method of the present application, ini=5, corresponding superpixel segmentation result, Bayesian prior saliency map, Bayesian saliency map; (q)-(u): change of the top 10 connected domains.

[0061] Figure 4Figures (a) to (h) show a sea scene SAR image target detection case, where (a) input matched filter SAR image; (b) image enhancement result based on the L1 ADMM sparse enhancement method of the present application, ini = 10; (c) corresponding superpixel segmentation result, (d) Bayesian prior saliency map after 5 iterations; (e) Bayesian saliency map, (f) binary map based on threshold segmentation; (g) target detection ground truth; (h) final target detection result.

[0062] Figure 5 Figures (a) to (d) show a ground scene SAR image target detection case, where (a) input matched filter SAR image; (b) image enhancement result based on the L1 ADMM sparse enhancement method of the present application, ini = 20; (c) corresponding superpixel segmentation result, (d) Bayesian prior saliency map after 7 iterations.

[0063] Figure 6 Figures (a) to (c) show ROC curves of ship target detection in (a) and (b). Figure 4 Figures (a) to (c) show ROC curves of ship target detection in (a) and (b). Figure 6 Figures (a) to (c) show ROC curves of ship target detection in (a) and (b). Figure 5 Figures (a) to (c) show ROC curves of ship target detection in (a) and (b). DETAILED DESCRIPTION

[0064] The present application will be further described below in conjunction with the accompanying drawings and examples of the present application.

[0065] As shown in Figure 1 , a SAR image target detection method based on sparse enhancement and Bayesian saliency of the present application is realized by the following steps:

[0066] Step 1: input a matched filter SAR image, and obtain the initial value of the regularization parameter of the SAR sparse enhancement method by calculating the gray statistical characteristics of the image. Specifically as follows:

[0067] Step 01) input a matched filter SAR image (as shown in Figure 2 (a)), where Y MF represents a SAR image obtained by matched filtering (referred to as "matched filter SAR image"), represents an image with a size of M x N pixels. The image gray value is normalized to the range of [0, 255], i.e. 8bit quantization. Then the gray histogram of the image (as shown in Figure 2 (b)) is calculated;

[0068] Step 02) obtain the sparsity of the image by using the gray histogram, and calculate the initial value of the regularization parameter by using the sparsity;

[0069] The sparsity of the matched filter SAR image is defined as sr=N hi / N MF , wherein N MF represents the number of all pixel points in the matched filter SAR image, N hi represents the number of pixel points with a gray value higher than 220. The greater the value of sr, the more pixel points of the foreground in the image, and a smaller initial value of the sparse regularization parameter needs to be set. The smaller the value of sr, the more pixel points of the background in the image, and a larger initial value of the sparse regularization parameter needs to be set to reduce the background noise intensity.

[0070] The initial value and variation of the sparse regularization parameter λ are defined as follows:

[0071]

[0072] wherein ini represents the initial value, k represents the iteration number of the SAR sparse enhancement method, and var(Y MF ) represents the variance of the input image.

[0073] Step 02: using the Alternating Direction Method of Multipliers (ADMM) (see Yang, A fast alternating direction method for tvl1-l2 signal reconstruction from partial fourier data, IEEE Journal of Selected Topics in Signal Process, 2010, vol. 4, no. 2) to solve the L1 norm-based convolution sparse feature enhancement algorithm model and obtain the SAR sparse enhancement image.

[0074] Step 03) ADMM operator based on matrix multiplication, design L1 norm-based convolution sparse feature enhancement algorithm model;

[0075] First, construct the SAR image sparse enhancement model, which is based on the following expression:

[0076] Y MF = X + N (2)

[0077] wherein is the sparse enhancement result, with a size of M×N pixels. The difference between the sparse enhancement result and the original image, including system noise, sidelobes to be removed, background clutter, etc. X can be solved by L1 norm sparse regularization method. In the present application, the estimated value of X is solved by the following sparse model

[0078]

[0079] Wherein, The estimated value of X is represented by X, The F norm of the matrix is represented by || || F, the L1 norm is represented by || || 1, and λ is a regularization parameter; The convolution operation of matrix X and matrix C is represented by X * C, and the specific expression of C is as follows:

[0080]

[0081] Wherein c1 is a constant between 0.5 and 1, and c1 is a constant between 0.5 and 1. The convolution operation can eliminate the scattering points that have been zeroed in the target. The sliding step in the convolution process is set to 1, and the edge points are removed after convolution, so that the size of the image after convolution is consistent with the input image. ε represents a small constant, and ⊙ represents the Hadamard product of the matrix. The present application solves formula (3) by using the ADMM algorithm based on matrix multiplication.

[0082] Step 04) Calculate the specific expression of three-step iteration in the ADMM algorithm to obtain the SAR sparse enhancement result in one iteration process;

[0083] The present application uses the ADMM algorithm based on matrix multiplication, uses the iteration step, and calculates the sparse enhancement image step by step Auxiliary matrix Lagrange multiplier matrix Specifically, set X, Z, β, P c The initial value is Wherein I is an identity matrix, and P c The connected domain matrix is represented by P, The connected domain matrix in the kth iteration is represented by P k, which controls the termination condition of the algorithm. The iteration optimization expression of the three matrices X, Z and β is as follows:

[0084]

[0085]

[0086] β (k+1) =β (k) +ρ(X (k+1) -Z (k+1) ) (7)

[0087] where k denotes the iteration number, and the upper index (k+1) denotes the corresponding value in the (k+1)th iteration. In formula (5), L(·) denotes the objective function, and p denotes a penalty factor for controlling the reconstruction accuracy and convergence speed of the algorithm, which is set to p=0.99 in the present application. In formula (6), S(·;·) denotes an iterative soft threshold operation, and T denotes a threshold matrix, which is expressed as follows:

[0088]

[0089] where sign(·) denotes a sigmoid nonlinear function, |·| denotes taking an absolute value, and max(x,0) denotes taking the maximum value of x or 0. The constant e is a small constant, and e=1.e -6 .

[0090] Step 03: Calculate the Bayesian saliency map of the SAR sparse enhancement image, and binarize the saliency map by using a threshold detection method to achieve the final target detection.

[0091] Step 05) Calculate the superpixel block by using the SAR sparse enhancement result, and calculate the Bayesian prior saliency map based on the superpixel block.

[0092] The present application uses the SAR sparse enhancement image as input to calculate the superpixel block. The coherent speckle noise of the SAR sparse enhancement image has been suppressed, and the intensity of the background region has also been reduced. Therefore, it is not necessary to use a more complex calculation step to calculate the superpixel block, and the traditional simple linear iterative clustering (SLIC) algorithm can be directly used for superpixel segmentation (see Achanta, SLIC Superpixels Compared to State-of-the-Art Superpixel Methods, IEEE Trans. Pattern Anal. Mach. Intell, 2012, vol. 34, no. 11). The gray difference d i and the distance difference d s between the pixel point i and the superpixel seed point j are used to measure the similarity D ij between the pixel points, and then the SLIC algorithm is used for segmentation. The specific measurement expression is as follows:

[0093]

[0094] where l i denotes the gray value at the pixel point i, (x i , y iThe superpixel s represents the position of pixel i. The superpixel s after segmentation can be obtained using the SLIC algorithm. i i = 1, 2, ... N s N s This represents the number of superpixel blocks.

[0095] Get s i Then, the Bayesian prior saliency map P(sal|s) of the superpixel block is calculated. i )as follows:

[0096]

[0097] Where ω pos (s i s j ) represents the distance weight, d int (s i s j ) represents superpixel block s j All pixels in the middle and s i The Euclidean distance value of the center pixel (this value is obtained by calculating the Euclidean distance between two pixel blocks), is normalized to the range [0, 1]. Distance parameter σ pos The contribution of the distance between superpixels is controlled in this invention. d shape (s i ) represents s i The number of all pixels in the array.

[0098] The calculation yields P(sal|s) i Then, the normalized prior saliency map is obtained using the min-max scaling method. The formula is:

[0099]

[0100] Among them, P min (sal|s i P represents the minimum saliency value. max (sal|s i ) represents the largest saliency value.

[0101] After obtaining the normalized prior saliency map of all superpixel blocks using formula (11), all of them can be... Combined into the prior saliency map of the whole image

[0102] Step 06) Determine the normalized prior saliency map The changes in the connected components with the most pixels are analyzed. When the range of changes in the pixels and corresponding pixel values ​​of these 10 connected components is less than a set threshold, a normalized prior saliency map is used. The final Bayesian saliency map is calculated.

[0103] First, normalized prior saliency maps are used. To determine whether a superpixel belongs to the foreground (fg) or background (bg) of an image, segmentation is performed using the mean pixel value of the prior saliency map as a threshold. If the normalized prior saliency map of the superpixel... If a pixel value is greater than the mean, the superpixel is considered to belong to the foreground region fg; otherwise, it belongs to the background. The specific formula is as follows:

[0104]

[0105] Where s i Let represent the i-th superpixel block, and mean(·) represent the mean gray value of the prior salient pixels in the superpixel block.

[0106] Further calculate the top 10 largest connected components in the foreground region fg, set the pixel values ​​of the remaining smaller connected components to 0, and the processed foreground region is counted as... It is an M×N matrix.

[0107] Judgment: If Or the number of iterations k > N s ,

[0108] Then: Let k = k + 1, and continue to return to the calculation formula (5).

[0109] if The output is the feature enhancement result X, and the normalized prior saliency map of the entire image.

[0110] The final Bayesian prior saliency map is calculated using the following formula for a given pixel z in the entire SAR image:

[0111]

[0112] Where P(z|sal) and P(z|bg) represent the posterior probabilities of the salient region and the background region, respectively, and N fg (z) represents the number of pixels in the background region bg obtained using formula (12) that have the same gray value as point z, N. bg (z) represents the number of pixels in the foreground region fg that have the same gray value as point z. N fg and N bg These represent the total number of pixels in the foreground and background of the SAR image after processing by formula (12), respectively.

[0113] Step 07) Use threshold detection method to make Bayesian saliency Figure Two Value-based analysis yields the final target detection result.

[0114] Specifically, a global threshold Tz based on P(sal|z) is set. When the Bayesian saliency map value P(sal|z) at pixel z in the image is greater than this threshold, the pixel value at z is set to 1, meaning the location of the pixel with a value of 1 is the target region; otherwise, it is set to 0, and the location of the pixel with a value of 0 is the background region. The final binary image, i.e., the target detection result, is obtained.

[0115] Example: An experimental method for SAR image target detection based on sparse enhancement and Bayesian saliency.

[0116] Experiments: To verify the effectiveness of the proposed method, multiple SAR images of land and sea scenes with complex backgrounds were selected for experiments. In the sea scene images, the target was a ship, and the background included land and ports. In the land scene SAR images, the target was a vehicle, and the background included buildings, roads, grass, and trees. The Constant False Alarm Rate (CFAR) detection method was used, and a saliency detection method based on matched-filter SAR images was compared with a traditional L1 norm-based sparse enhancement method. All experiments were conducted in MATLAB R2016 using an Intel Core i5-5257 CPU with a clock speed of 2.7GHz and 8GB of memory.

[0117] The experimental parameters are set as follows:

[0118] c1 = 1, ρ = 0.99, N s =100, ε=1·e -6 ,

[0119] Experimental results are as follows Figures 3-6 As shown. Figure 3 This display shows the saliency detection results for a sea scene image using different methods. The input MF SAR image contains two ship targets and a large area of ​​land background. The intensity of the land background is similar to, or even higher than, the intensity of the ship targets, which poses a significant challenge to saliency detection. From... Figure 3 As can be seen from (a)-(d), using MF SAR images as input data does not yield good salience detection results. Due to the complexity and high intensity of the background, the superpixel method cannot accurately extract the target contour, nor can it highlight the target region from the Bayesian prior image. Figure 3 (e)-(l) use sparse feature-enhanced images obtained through traditional L1 sparse enhancement methods as input data. Two different regularization parameters λ are chosen to test their performance.Figure 3 As can be seen from (e)-(h), although the traditional L1-based IST method significantly suppresses sea clutter, many high-intensity background superpixels still remain. Therefore, the target region in the Bayesian prior map is not prominent enough, and this is also true for the target region in the final Bayesian prior map. Figure 3 As can be seen from (i)-(l), if a large λ is chosen, the feature enhancement result will become more sparse. Because traditional L1 sparse enhancement methods produce holes and discontinuities in the target region under large λ conditions, the superpixel method cannot obtain good segmentation results, especially for the ground background region (where most superpixels are square). This leads to poor performance when calculating the Bayesian prior map. Figure 3 (m)-(p) show the results of the method proposed in this invention. The initial value of λ was calculated as ini = 5, and convergence was achieved after 5 iterations. It can be seen that the superpixel method can obtain good target contours while also segmenting the background region. A complete target region with continuous contours can be obtained from the final Bayesian prior saliency map. Figure 3 (q)-(u) shows the changes in the top 10 connected domains in Pc. The proposed method obtains the final saliency map in only 5 iterations, which is much more time-efficient compared to the previously mentioned method that requires nearly 100 iterations.

[0120] Figure 4 This shows the target detection results for another SAR image of a sea surface scene using the proposed method. The initial value of λ was calculated using the intensity histogram as ini = 10, and convergence was required after 5 iterations. From... Figure 4 As shown in Figure (b), despite the algorithm choosing a relatively large λ, there are very few holes on the ship target. In Figure 4(d), the foreground region is bright with a smooth and complete outline. In the Bayesian saliency map, there is a large intensity difference between the foreground and background, while the pixel values ​​of the foreground are relatively average. Figure 4 (e) Display the saliency plot, Figure 4 (f) Displays the binary image obtained by the global thresholding method. Figure 4 (h) shows the final detection results, where the minimum bounding rectangle of the real target is marked with a green box. It can be seen that the proposed method can obtain accurate outlines of all ship targets with fewer false alarms.

[0121] Figure 5This displays target detection results for a mini SAR ground scene SAR image with a resolution of 0.1 m × 0.1 m and 1638 × 2510 pixels. The image has a complex foreground and background, containing vehicles and other terrain features such as buildings, roads, grass, and trees. Some terrain features have intensities similar to, or even higher than, those of vehicles. The initial value of λ was calculated as ini = 20, indicating that the intensity difference between the background and foreground is small. Figure 5 In (b), it can be seen that low-intensity roads and grass are suppressed, while vehicle outlines are preserved. The proposed method requires 7 iterations to obtain the final prior map. From Figure 5 As can be seen in (d)-(e), the foreground includes vehicle targets, buildings, and trees, with the target outlines being smooth and complete. It is difficult to extract only vehicle targets from this SAR image without generating clutter false alarms. Figure 5 (f) shows the binary image after adding size information to remove clutter much larger than the vehicle. Although false alarms are still present in the binary image, all target regions are well preserved. This is very helpful for subsequent target feature extraction, especially for extracting the target's length, width, and other geometric features. It can be seen that the proposed method can preserve the target's outline to the greatest extent.

[0122] Figure 6 Display ROC curves for different methods. For example... Figure 6 As shown, the ROC curve obtained by the method of the present invention is superior to that of the comparative method, especially in... Figure 6 In (b), the background of the SAR image is very complex, and the intensity difference between the background and the foreground is very small.

Claims

1. A SAR image target detection method based on sparse enhancement and Bayesian saliency, characterized in that, This method is implemented through the following steps: Step 1: Input a matched-filtered SAR image. By calculating the image's gray-level statistical characteristics, obtain the initial values ​​of the regularization parameters for the SAR sparse enhancement method; details are as follows: Step 01) Input matched-filtered SAR image Among them, Y MF This refers to the SAR image obtained through matched filtering, and is simply called a "matched-filtered SAR image". The image size is M×N pixels; the grayscale values ​​of the image are normalized to the range of [0, 255], i.e., 8-bit quantization; then the grayscale histogram of the image is calculated. Step 02) Obtain the sparsity of the image using the grayscale histogram, and use the sparsity to calculate the initial value of the regularization parameter; Define the sparsity of a matched-filtered SAR image as sr = N hi / N MF , where N MF N represents the total number of pixels in a matched-filtered SAR image. hi This indicates the number of pixels with a grayscale value higher than 220. The larger the value of sr, the more pixels there are in the foreground of the image, and the more necessary it is to set a smaller initial value for the sparsity regularization parameter. On the other hand, the smaller the value of sr, the more pixels there are in the background of the image, and the more necessary it is to set a larger initial value for the sparsity regularization parameter to reduce the intensity of background noise. The initial value and variation of the sparse regularization parameter λ are defined as follows: λ (k) =(ini+k)·var(Y MF ) Where ini represents the initial value, k represents the number of iterations for calculating the SAR sparse enhancement method, and var(Y) MF ) represents the variance of the input image; Step 2: Solve the L1 norm-based convolutional sparse feature enhancement algorithm model using the alternating direction multiplier method to obtain the SAR sparse enhancement image; Step 03) Based on the ADMM operator of matrix multiplication, design a convolution sparse feature enhancement algorithm model based on L1 norm; First, a sparse enhancement model for SAR images is constructed, which is based on the following expression: AND MF =X+N (2) in For sparse enhancement results, the size is M×N pixels; This indicates the difference between the sparse enhancement result and the original image; X is solved using the L1 norm sparse regularization method; specifically, the estimated value of X is solved using the following sparse model. in, This represents the estimated value of X. Let ||||1 denote the F norm of the matrix, ||||1 denotes the L1 norm, and λ is the regularization parameter; This represents the convolution operation between matrix X and matrix C, where C is specifically expressed as: Where c1∈[0.5,1] is a constant between 0.5 and 1; the scattering points in the target that have been set to zero are eliminated by the convolution operation; the sliding step size is set to 1 during the convolution process, and the edge points are removed after the convolution so that the size of the convolved image is consistent with the input image; ε represents a small constant, and ⊙ represents the Hadamard product of the matrices; Step 04) Calculate the specific expressions for the three iterations in the ADMM algorithm to obtain the SAR sparse enhancement results in one iteration process; The ADMM algorithm using matrix multiplication is used to compute sparse enhancement images step by step through iterative steps. Auxiliary matrix Lagrange multiplier matrix Specifically, let X, Z, β, P be defined. c The initial value is Where I is the identity matrix, P c Represents a connected component matrix. Let X represent the connected component matrix in the k-th iteration, which controls the termination condition of the algorithm; the iterative optimization expressions for the three matrices X, Z, and β are as follows: b (k+1) =b (k) +ρ(X (k+1) -Z (k+1) ) (7) Where k represents the number of iterations, and the superscript... (k+1) Let L(·) represent the corresponding value in the (k+1)th iteration; in formula (5), L(·) represents the objective function, and ρ represents the penalty factor, which is used to control the reconstruction accuracy and convergence speed of the algorithm; in formula (6), S(·;·) represents the iterative soft thresholding operation, and T represents the threshold matrix, as shown in the following expression: Where sign(·) represents the sigmoid nonlinear function, |·| represents taking the absolute value, max(x, 0) represents taking the maximum value between x and 0; the constant ε is a small constant; Step 3: Calculate the Bayesian saliency map of the SAR sparse enhancement image, and use the threshold detection method to binarize the saliency map to achieve the final target detection; Step 05) Calculate superpixel blocks using SAR sparse enhancement results, and calculate Bayesian prior saliency maps based on superpixel blocks; Superpixel blocks are computed using SAR sparse enhanced images as input. Superpixel segmentation is performed directly using the SLIC (Single Linear Iterative Clustering) algorithm, utilizing the gray-level difference d between pixel i and superpixel seed j. i and distance difference d s To measure the similarity between pixels D ij Then, the SLIC algorithm is used for segmentation, with the specific metric expression as follows: Among them l i Represents the grayscale value at pixel i, (x i y i The superpixel s is obtained by the SLIC algorithm, where i represents the position of pixel i. i i = 1, 2, ... N s N s The number of superpixel blocks; Get s i Then, the Bayesian prior saliency map P(sal|s) of the superpixel block is calculated. i )as follows: Where ω pos (s i s j ) represents the distance weight, d int (s i s j ) represents superpixel block s j All pixels in the middle and s i The Euclidean distance value of the center pixel is normalized to the range [0, 1]; the distance parameter σ pos It controls the contribution of the distance between superpixels; d shape (s i ) represents s i The number of all pixels in the image; The calculation yields P(sal|s) i Then, the normalized prior saliency map is obtained using the minimum-maximum scaling method. The formula is: Among them, P min (sal|s i P represents the minimum saliency value. max (sal|s i ) represents the largest saliency value; After obtaining the normalized prior saliency maps of all superpixel blocks using formula (11), all of them... Combined into the prior saliency map of the whole image Step 06) Determine the normalized prior saliency map The changes in the connected components with the most pixels are analyzed. When the range of changes in the pixels and corresponding pixel values ​​of these 10 connected components is less than a set threshold, a normalized prior saliency map is used. The final Bayesian saliency map is calculated. First, normalized prior saliency maps are used. To determine whether a superpixel belongs to the foreground (fg) or background (bg) of an image, segmentation is performed using the mean pixel value of the prior saliency map as a threshold. If the normalized prior saliency map of the superpixel... If a pixel value is greater than the mean, the superpixel is considered to belong to the foreground region fg; otherwise, it belongs to the background. The specific formula is as follows: Where s i Let represent the i-th superpixel block, and mean(·) represent the mean gray value of the prior salient pixels in the superpixel block; Further calculate the top 10 largest connected components in the foreground region fg, set the pixel values ​​of the remaining smaller connected components to 0, and the processed foreground region is counted as... It is an M×N matrix; Judgment: If Or the number of iterations k > N s , Then: Let k = k + 1, and continue to return to the calculation formula (5); if The output is the feature enhancement result X, and the normalized prior saliency map of the entire image. The final Bayesian prior saliency map is calculated using the following formula for a given pixel z in the entire SAR image: Where P(z|sal) and P(z|bg) represent the posterior probabilities of the salient region and the background region, respectively, and N fg (z) represents the number of pixels in the background region bg obtained using formula (12) that have the same gray value as point z, N bg (z) represents the number of pixels in the foreground region fg that have the same gray value as point z; N fg and N bg These represent the total number of pixels in the foreground and background of the SAR image after processing by formula (12); Step 07) Use the threshold detection method to binarize the Bayesian saliency map to obtain the final target detection result; Specifically, a global threshold Tz based on P(sal|z) is set. When the Bayesian saliency map value P(sal|z) at pixel z in the image is greater than the threshold, the pixel value at z is set to 1, that is, the pixel position with a value of 1 is the target region. Otherwise, it is set to 0, and the pixel position with a value of 0 is the background region. The final binary image is obtained, that is, the target detection result.

2. The SAR image target detection method based on sparse enhancement and Bayesian saliency as described in claim 1, characterized in that, In Step 04), ρ=0.99, ε=1·e -6 。 3. The SAR image target detection method based on sparse enhancement and Bayesian saliency as described in claim 1, characterized in that, In Step 05),

Images