Ship candidate area detection method and system oriented to large-scene SAR (Synthetic Aperture Radar) image

Abstract

The invention relates to the technical field of radar remote sensing image intelligent processing, in particular to a ship candidate area detection method and system for a large-scene SAR image, and the method comprises the steps: extracting an ocean area of the SAR image according to a submarine topographic map corresponding to the SAR image; the method comprises the following steps: modeling an ocean area of an SAR image based on Gamma distribution, obtaining an initial shape parameter and an initial scale parameter by using a mean value and a variance of pixels of the ocean area of the SAR image, and correcting by using a Le Cam one-step estimation method based on a Fisher information matrix to obtain a target shape parameter and a target scale parameter; based on a preset false alarm rate, calculating a threshold value of each ocean region pixel of the SAR image through inverse Gamma distribution to obtain a threshold value graph; and obtaining a ship candidate area mask image according to the SAR image and the corresponding threshold image. According to the method, the calculation burden can be remarkably reduced under a large-scene SAR image, and the method has higher adaptability and universality.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent processing of radar remote sensing images, and in particular to a method and system for detecting candidate ship areas in large-scene SAR images. Background Art

[0002] With the continuous increase in global trade, fishing activities, and maritime security issues, the demand for accurate and continuous monitoring of the marine environment is becoming increasingly urgent. Satellite remote sensing, especially Synthetic Aperture Radar (SAR) data, has become a core tool for ocean monitoring due to its high stability and wide coverage. SAR images are generated by a radar system actively transmitting microwave signals and receiving the signals reflected from targets on the ground or sea surface. Compared to optical remote sensing, which relies on natural lighting, SAR images are unaffected by clouds, rain, fog, and day-night changes, and have all-weather and all-day imaging capabilities. Its unique imaging mechanism gives it significant advantages such as high resolution, strong penetration, and anti-interference, making it particularly suitable for target detection tasks in complex environments such as the ocean.

[0003] In practical applications, SAR images are widely used for ship candidate detection, target recognition, tracking, and classification. As a fundamental step in the entire processing process, ship candidate detection automatically identifies areas in SAR images where suspected ships may be present, providing prior support for subsequent higher-level tasks such as ship classification, speed estimation, behavior analysis, and trajectory tracking.

[0004] In recent years, with the large-scale deployment and upgrading of high-resolution SAR satellites, the spatial resolution of acquired image data has not only increased significantly, but the volume of data has also exploded. This trend places higher demands on ship candidate detection algorithms: they must not only possess high detection accuracy but also meet the requirements of rapid processing and real-time response to meet the needs of practical application scenarios such as dynamic maritime situation monitoring and early warning command.

[0005] Existing methods for detecting candidate ship regions in SAR images include:

[0006] (1)Constant False Alarm Rate (CFAR) detection based on Gamma distribution fitting: This method models the ocean background echo intensity in SAR images with a Gamma distribution and sets the detection threshold adaptively based on the CFAR theory to identify ship targets with strong scattering characteristics. The Gamma distribution parameters are fitted using maximum likelihood estimation, which involves numerical iterative solution of the log-likelihood function. Although this method has a clear statistical modeling basis and good theoretical explanation, the overall computational overhead is huge when processing high-resolution, wide-range SAR images, and the processing time of a single image can exceed 900 seconds, which cannot meet the real-time and processing efficiency requirements of ship candidate area detection tasks.

[0007] (2)GPU-accelerated cell average CFAR detection: This method estimates the background echo by averaging the pixels in the reference window and sets the detection threshold. With the help of GPU parallel computing, the processing speed can be improved, but the overall efficiency is still insufficient to support real-time processing of large-scale SAR images. In addition, this method is highly dependent on hyperparameters such as sliding window size, guard cell number, and detection threshold, which need to be adjusted manually, resulting in poor adaptability and insufficient generality.

[0008] (3)GPU-accelerated minimum value CFAR detection: This method selects the smaller average value from multiple reference windows around the target when setting the detection threshold, aiming to reduce the risk of false detection caused by strong background echoes on one side. Although GPU acceleration improves efficiency, it still cannot meet the demand for efficient real-time processing in practical applications. At the same time, this method relies on multiple hyperparameter configurations, such as window size and guard cell number, making the parameter tuning process complex and less versatile.

[0009] (4)GPU-accelerated maximum value CFAR detection: This method is similar to the GPU-accelerated minimum value CFAR detection, but selects the larger average value from multiple reference windows around the target as the detection threshold to enhance the suppression ability of high clutter background. Although this method performs better in interference suppression, it still has the problems of limited processing speed and many parameters to adjust.

[0010] In summary, the existing SAR image ship candidate area detection methods have problems such as poor real-time performance and strong parameter dependence when processing large scene images, which cannot meet the real-time image processing requirements in practical applications. SUMMARY

[0011] Therefore, the technical problem to be solved by the present application is to overcome the problems of poor real-time performance and strong parameter dependence in the prior art when processing large scene images, which cannot meet the real-time image processing requirements in practical applications.

[0012] To solve the above technical problems, the application provides a ship candidate area detection method for large-scene SAR images, comprising the following steps:

[0013] An underwater topographic map corresponding to the SAR image is acquired, a sea-land mask map is generated according to the underwater topographic map, and a marine area of the SAR image is extracted;

[0014] The marine area of the SAR image is modeled based on a Gamma distribution, initial shape parameters and initial scale parameters are obtained by using the mean and variance of the marine area pixels of the SAR image, the initial shape parameters and the initial scale parameters are corrected by using a Le Cam one-step estimation method based on a Fisher information matrix, and target shape parameters and target scale parameters are obtained;

[0015] Based on the target shape parameters, the target scale parameters and a preset false alarm rate, a threshold value of each marine area pixel of the SAR image is calculated by using an inverse Gamma distribution, and a threshold value map is obtained;

[0016] The SAR image is compared with the threshold value map corresponding thereto, and a ship candidate area mask map is obtained.

[0017] Preferably, the initial shape parameters and the initial scale parameters are obtained by using the mean and variance of the marine area pixels of the SAR image, the initial shape parameters and the initial scale parameters are corrected by using the Le Cam one-step estimation method based on the Fisher information matrix, and the target shape parameters and the target scale parameters are obtained, comprising the following steps:

[0018] The mean and variance of the marine area pixels of the SAR image are calculated, and the initial shape parameters and the initial scale parameters are calculated according to the mean and the variance;

[0019] The gradient of the shape parameters and the scale parameters in a log-likelihood function is taken as a score function, the score function of the initial shape parameters and the initial scale parameters is constructed, a Fisher information matrix of the initial shape parameters and the initial scale parameters is constructed, and the Le Cam one-step estimation method is used to calculate the target shape parameters and the target scale parameters based on the score function and the Fisher information matrix of the initial shape parameters and the initial scale parameters.

[0020] Preferably, the score function of the initial shape parameters and the initial scale parameters is constructed The formula is as follows:

[0021]

[0022] Wherein, k is the shape parameter, and theta is the scale parameter; k0 is the initial shape parameter, and theta0 is the initial scale parameter; score k represents the score in the k direction; score θscore in the direction of θ; n is the number of ocean region pixels of the SAR image; x i represents the i-th ocean region pixel; ψ(·) is the Digamma function;

[0023] The Fisher information matrix F(k0, θ0) of the initial shape parameter and the initial scale parameter is constructed, and the formula is:

[0024]

[0025] where ψ (2) (·) is the trigamma function;

[0026] Based on the score function and the Fisher information matrix of the initial shape parameter and the initial scale parameter, the target shape parameter and the target scale parameter are calculated by using the Le Cam one-step estimation method, and the formula is:

[0027]

[0028] where k1 is the target shape parameter, θ1 is the target scale parameter, F -1 (k0, θ0) is the inverse matrix of the Fisher information matrix.

[0029] Preferably, based on the GPU tensor calculation framework, the initial shape parameter and the initial scale parameter are obtained by using the mean and variance of the ocean region pixels of the SAR image, and the initial shape parameter and the initial scale parameter are corrected by using the Le Cam one-step estimation method based on the Fisher information matrix, and the target shape parameter and the target scale parameter are calculated, including:

[0030] A plurality of SAR images are input into the GPU as a batch, and a convolution kernel is used as a sliding window to divide each SAR image into a plurality of windows;

[0031] The ocean region pixels of each window are normalized;

[0032] The mean and variance of the normalized ocean region pixels of each window are calculated in parallel to obtain a mean tensor and a variance tensor;

[0033] Based on the mean tensor and the variance tensor, the initial shape parameter and the initial scale parameter of each window are calculated in parallel to obtain an initial shape parameter tensor and an initial scale parameter tensor;

[0034] The gradient of the shape parameter and the scale parameter in the log-likelihood function is used as the score function, and the score of each window in the shape parameter direction and the scale parameter direction is calculated in parallel based on the initial shape parameter tensor and the initial scale parameter tensor to obtain a shape parameter score tensor and a scale parameter score tensor;

[0035] The Fisher information tensor is obtained by parallel computing the Fisher information matrix of each window based on the initial shape parameter tensor and the initial scale parameter tensor;

[0036] The target shape parameter tensor and the target scale parameter tensor are obtained by parallel computing the target shape parameter and the target scale parameter of each window based on the shape parameter score tensor, the scale parameter score tensor and the Fisher information tensor using the Le Cam one-step estimation method.

[0037] Preferably, the mean tensor and the variance tensor are obtained by parallel computing the mean and the variance of each window marine area pixel, and the formula includes:

[0038]

[0039] v[b,i,j]=m2[b,i,j]-m[b,i,j] 2

[0040] Wherein, m[b,i,j] is the value of the position coordinate (i,j) in the mean tensor of the bth SAR image, corresponding to the mean of the i-th row and j-th column window marine area pixel of the bth SAR image; m2[b,i,j] is the value of the position coordinate (i,j) in the square mean tensor of the bth SAR image, corresponding to the square mean of the i-th row and j-th column window marine area pixel of the bth SAR image; v[b,i,j] is the value of the position coordinate (i,j) in the variance tensor of the bth SAR image, corresponding to the variance of the i-th row and j-th column window marine area pixel of the bth SAR image; N h is the size of the convolution kernel in the row direction, N v is the size of the convolution kernel in the column direction; W h [p,q] is the weight element at the position coordinate (p,q) in the convolution kernel W h I[b,si-p,sj-q] is the normalized value of the pixel at the position coordinate (si-p,sj-q) in the bth SAR image in the i-th row and j-th column window; s is the convolution step, and the step is smaller than the size of the convolution kernel; represents rounding.

[0041] Preferably, the initial shape parameter tensor and the initial scale parameter tensor are obtained by parallel computing the initial shape parameter and the initial scale parameter of each window based on the mean tensor and the variance tensor, and the formula includes:

[0042]

[0043] wherein k0[b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the initial shape parameter tensor, corresponding to the initial shape parameter of the i th row and j th column window of the bth SAR image; θ0[b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the initial scale parameter tensor, corresponding to the initial scale parameter of the i th row and j th column window of the bth SAR image; ε is a constant used to prevent the denominator from being zero; m[b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the mean value tensor; v[b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the variance tensor.

[0044] Preferably, based on the initial shape parameter tensor and the initial scale parameter tensor, the score of each window in the shape parameter direction and the scale parameter direction is calculated in parallel to obtain a shape parameter score tensor and a scale parameter score tensor, and the formula includes:

[0045] score k [b, i, j] = d[b, i, j] - n[b, i, j] · ψ(k0[b, i, j]) - n[b, i, j] · log(θ0[b, i, j])

[0046] score θ [b, i, j] = n[b, i, j] · k0[b, i, j] · θ0[b, i, j] - n[b, i, j] · m[b, i, j]

[0047] wherein score k [b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the shape parameter score tensor, corresponding to the score of the i th row and j th column window of the bth SAR image in the shape parameter direction; score θ [b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the scale parameter score tensor, corresponding to the score of the i th row and j th column window of the bth SAR image in the scale parameter direction; ψ(·) is the Digamma function; d[b, i, j] is the sum of the logarithmic values of the ocean region pixels of the i th row and j th column window of the bth SAR image; n[b, i, j] is the number of the ocean region pixels of the i th row and j th column window of the bth SAR image; k0[b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the initial shape parameter tensor; θ0[b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the initial scale parameter tensor; m[b, i, j] is a value at position coordinate (i, j) of the bth SAR image in the mean value tensor;

[0048] The Fisher information matrix of each window is calculated in parallel based on the initial shape parameter tensor and the initial scale parameter tensor to obtain a Fisher information tensor, and the formula includes:

[0049] F 11 [b, i, j] = ψ (2) (k0[b, i, j])·n[b, i, j] + ε′

[0050] F 12 [b, i, j] = F 21 [b, i, j] = -θ0[b, i, j]·n[b, i, j]

[0051] F 22 [b, i, j] = k0[b, i, j]·θ0[b, i, j] 2 ·n[b, i, j] + ε′

[0052] wherein F 11 [b, i, j] is a component of F[b, i, j] in the first row and the first column; F 12 [b, i, j] is a component of F[b, i, j] in the first row and the second column; F 22 [b, i, j] is a component of F[b, i, j] in the second row and the second column; F[b, i, j] is a matrix at a position coordinate (i, j) of the bth SAR image in the Fisher information tensor, corresponding to the Fisher information matrix of the jth window in the ith row of the bth SAR image; ψ(2)(·) is a trigamma function; and ε′ is a normal number.

[0053] Preferably, the target shape parameter and the target scale parameter of each window are calculated in parallel based on the shape parameter score tensor, the scale parameter score tensor and the Fisher information tensor by using a LeCam one-step estimation method to obtain a target shape parameter tensor and a target scale parameter tensor, and the formula includes:

[0054]

[0055] wherein k1[b, i, j] is a value at a position coordinate (i, j) of the bth SAR image in the target shape parameter tensor, corresponding to the target shape parameter of the jth window in the ith row of the bth SAR image; θ1[b, i, j] is a value at a position coordinate (i, j) of the bth SAR image in the target scale parameter tensor, corresponding to the target scale parameter of the jth window in the ith row of the bth SAR image; k0[b, i, j] is a value at a position coordinate (i, j) of the bth SAR image in the initial shape parameter tensor; θ0[b, i, j] is a value at a position coordinate (i, j) of the bth SAR image in the initial scale parameter tensor; score k[b, i, j] is the value at the bth SAR image position coordinate (i, j) in the shape parameter score tensor; score θ [b, i, j] is the value at the bth SAR image position coordinate (i, j) in the scale parameter score tensor; is the component of F-1[b, i, j] in the first row and the first column; is the component of F-1[b, i, j] in the first row and the second column; is the component of F-1[b, i, j] in the second row and the second column; F[b, i, j] is the matrix at the bth SAR image position coordinate (i, j) in the Fisher information tensor, F -1 [b, i, j] is the inverse matrix of F[b, i, j].

[0056] Preferably, based on the target shape parameter tensor, the target scale parameter tensor and the preset false alarm rate, the threshold of each window of the sea area is calculated by the inverse Gamma distribution to obtain a threshold tensor; the threshold of each pixel of the sea area of the SAR image is calculated by adopting the strategy of taking the minimum value in the window overlap area to obtain a threshold map, and the formula is represented as:

[0057]

[0058] wherein, T ext [b, i', j'] is the threshold value at the position coordinate (i', j') of the threshold map corresponding to the bth SAR image; T[b, m', n'] is the value at the position coordinate (m', n') of the bth SAR image in the threshold tensor, corresponding to the threshold value of the m'th row and then'th column window of the bth SAR image; H' and w' are the height and width of the threshold tensor respectively; R m′,n′ is the area of the m'th row and then'th column window in the corresponding SAR image.

[0059] The application also provides a ship candidate area detection system for large scene SAR images, comprising:

[0060] The sea area extraction module is configured to obtain a seabed topographic map corresponding to the SAR image, generate a sea-land mask map according to the seabed topographic map, and extract the sea area of the SAR image.

[0061] The parameter calculation module is configured to model the sea area of the SAR image based on the Gamma distribution, obtain initial shape parameters and initial scale parameters by using the mean and variance of the sea area pixels of the SAR image, and correct the initial shape parameters and the initial scale parameters by using the Le Cam one-step estimation method based on the Fisher information matrix to obtain target shape parameters and target scale parameters.

[0062] a threshold value calculation module configured to calculate a threshold value of each pixel in the marine area of the SAR image based on the target shape parameter, the target scale parameter and a preset false alarm rate through inverse Gamma distribution to obtain a threshold value map;

[0063] a ship candidate area detection module configured to compare the SAR image with the corresponding threshold value map to obtain a ship candidate area mask map.

[0064] The above technical solution of the present application has the following beneficial effects compared with the prior art:

[0065] The ship candidate area detection method for large-scene SAR images provided by the present application first generates a land-sea mask map for the SAR image by using a seabed topographic map, effectively eliminates the interference of land areas in the complex background of the SAR image, and improves the accuracy of modeling and the real-time performance of detection; then models the probability density of the marine area of the SAR image by using Gamma distribution, calculates the shape parameter and the scale parameter required for modeling by combining with the Le Cam one-step estimation method, and only performs one iteration correction, compared with the multi-round iteration optimization process required in the traditional maximum likelihood estimation method, the present application can significantly reduce the calculation burden under large-scene SAR images, while maintaining high parameter estimation accuracy and model fitting effect, to meet the real-time requirement of ship candidate area detection; finally, the threshold value of the marine area pixel is adaptively calculated by using the preset false alarm rate and the inverse Gamma distribution function, the ship candidate area is separated from the sea clutter by using the threshold value, and a ship candidate area mask map is obtained, so that the detection of the ship candidate area is completed. The present application only needs to set one hyperparameter of the false alarm rate, reduces the dependence on artificial parameter adjustment, and has stronger adaptability and universality.

[0066] Further, the present application adopts a parallel processing architecture based on the GPU tensor calculation framework, uniformly represents the sliding window parameter estimation problem as a standard tensor convolution form, efficiently completes the normalization preprocessing, background modeling, parameter estimation and detection threshold generation in a window parallel processing manner in the full image range by mapping the window size to the convolution kernel size, significantly reduces the calculation resource consumption and processing delay while ensuring the estimation accuracy; finally, a fusion strategy of taking the minimum value in the overlapping area of multiple windows is introduced to realize pixel-level threshold expansion, which can map the threshold value tensor to the size of the original SAR image to complete the detection of the ship candidate area, further improves the hit rate of the ship candidate area detection while ensuring the real-time performance of the detection. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to make the content of the present application easier to be clearly understood, the present application will be further described in detail below according to specific embodiments of the present application and in combination with the drawings, in which:

[0068] Figure 1is a flow chart of a ship candidate area detection method for large scene SAR images according to the present application;

[0069] Figure 2 is a structural diagram of a convolution kernel realizing sliding window overlapping tensor calculation;

[0070] Figure 3 is a result example diagram of generating a sea-land mask using a seabed topographic map;

[0071] Figure 4 is a result example diagram of generating a sea-land mask using a global sea-land mask;

[0072] Figure 5 is an example diagram of enlarging a ship-containing region of a SAR image in an XVIEW3 data set;

[0073] Figure 6 is an example diagram of enlarging a generated ship candidate area mask. DETAILED DESCRIPTION

[0074] The present application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it, but the embodiments are not limiting to the present application.

[0075] Referring to Figure 1 The present application provides a ship candidate area detection method for large scene SAR images, specifically comprising:

[0076] S1: Obtain a seabed topographic map corresponding to the SAR image, generate a sea-land mask map according to the seabed topographic map, and extract the marine area of the SAR image.

[0077] The seabed topographic map data can be sourced from a public or authorized marine geographic information platform, including but not limited to GEBCO, ETOPO, and global seabed elevation data published by national surveying and mapping institutions. Such data is stored in a grid format, recording the elevation value of the ground or seabed relative to sea level, wherein a positive value represents a land area, a negative value represents a marine area, and a 0 value corresponds to the position of the coastline. In one specific embodiment of the present application, the seabed topographic map data is sourced from the XVIEW3 data set.

[0078] Specifically, after registering the SAR image with its corresponding seabed topographic map, a sea-land mask map is generated according to the seabed topographic map, including:

[0079] S11: Perform a spatial clipping operation on the seabed topographic map to extract the geographical area consistent with the coverage range of the SAR image; perform a coordinate projection transformation operation on the clipped topographic map to make its projection coordinate system consistent with the SAR image, ensuring alignment consistency in geographical space, to obtain the registered seabed topographic map;

[0080] S12: performing a segmentation operation according to the height threshold of the registered seabed topographic map, determining a pixel region with a height less than 0 as a sea region, and determining a pixel region with a height greater than or equal to 0 as a land region, and generating an initial binary sea-land mask map;

[0081] S13: performing a nearest neighbor interpolation algorithm on the sea-land mask map to ensure spatial consistency of the sea-land mask map and the SAR image at a pixel level, so that the spatial resolution of the sea-land mask map is consistent with that of the SAR image, and obtaining a target sea-land mask map.

[0082] By introducing the sea-land mask map, the land region interference can be effectively shielded in the subsequent sea background statistical modeling process, and the accuracy and robustness of the model parameter estimation are improved. In addition, the sea-land mask map can also be directly applied to the ship candidate region detection stage, and the land region is structurally excluded, so that the non-target coverage rate of the final ship candidate region mask is significantly reduced, and the processing efficiency is improved.

[0083] Preferably, after obtaining the sea-land mask map, the SAR image and its corresponding sea-land mask map are down-sampled, and the formula includes:

[0084] I1 = Resize(I b ; H, W)

[0085] Map f = Resize(Map b ; H, W)

[0086] wherein Resize(·) represents image scaling using area interpolation; I b is the original SAR image; Map b is the sea-land mask map; I1 is the down-sampled SAR image; Map f is the down-sampled floating-point sea-land mask map; and are the height and width of the down-sampled SAR image and floating-point sea-land mask map, and a e (0, 1) is a down-sampling scaling factor; H b and W b are the height and width of the original SAR image and sea-land mask map.

[0087] The floating-point sea-land mask map Map f is subjected to a binaryzation operation, all pixels greater than 0 are set to 1, otherwise set to 0, and the formula is as follows:

[0088]

[0089] wherein Map is the down-sampled binary sea-land mask map, Map(i', j') is the value at the position coordinate (i', j') of the down-sampled binary sea-land mask map, and Mapf (i′, j′) is the value of the floating-point land and sea mask at the position coordinate (i′, j′) after downsampling.

[0090] Through downsampling operations, the image resolution can be effectively reduced, and the data dimension can be significantly reduced while retaining the ship target characteristics as much as possible, thereby reducing the computational complexity and video memory usage of the subsequent constant false alarm rate detection module, further accelerating the detection speed of the entire image, and improving the overall real-time processing capability of the system while maintaining detection accuracy.

[0091] S2: Model the ocean area of ​​the SAR image based on the Gamma distribution, use the mean and variance of the pixels in the ocean area of ​​the SAR image to obtain the initial shape parameters and initial scale parameters, and then use the Le Cam one-step estimation method based on the Fisher information matrix to correct the initial shape parameters and initial scale parameters to obtain the target shape parameters and target scale parameters. Specifically including:

[0092] S201: Model the ocean area of ​​the SAR image based on the Gamma distribution. Its probability density function is as follows:

[0093]

[0094] Where x>0 represents the echo intensity of the pixel in the ocean area of ​​the downsampled SAR image; k>0 is the shape parameter, which controls the degree of skewness of the distribution; θ>0 is the scale parameter, which determines the extension range of the distribution; Γ(k) is the Gamma function, which satisfies is the integration symbol, and t is the integration variable.

[0095] S202: Initial shape parameters and initial scale parameters are obtained using the mean and variance of the pixels in the ocean area of ​​the SAR image.

[0096] Using the moment characteristics of the Gamma distribution, that is, the mathematical expectation (mean) E[x] = kθ and variance Var[x] = kθ of the Gamma distribution 2 , which can be used to establish the relationship between parameters and sample statistics.

[0097] By calculating the mean m and variance v of the pixel echo intensity in the ocean area of ​​the SAR image, and combining the two equations m=kθ,v=kθ in the above moment characteristics 2 , the initial shape parameter can be deduced as: k0 = m 2 / v, the initial scale parameter is: θ0 = v / m.

[0098] S203: The initial shape parameter and the initial scale parameter are corrected by using a Le Cam one-step estimation method based on a Fisher information matrix to obtain a target shape parameter and a target scale parameter.

[0099] Considering that the moment estimation only uses the sample, that is, the low-order moment information of the ocean region pixels, there is a certain deviation. In order to further improve the accuracy of parameter estimation, the Le Cam one-step estimation method is introduced. The method takes the initial value (k0, θ0) as the basis, takes the gradient of the log-likelihood function on the shape parameter and the scale parameter as the score function, and constructs the score function of the initial shape parameter and the initial scale parameter; based on the score function of the initial shape parameter and the initial scale parameter and the Fisher information matrix, the Le Cam one-step estimation method is used to calculate the target shape parameter and the target scale parameter closer to the maximum likelihood estimation.

[0100] The expression of the log-likelihood function is:

[0101]

[0102] Wherein, k is the shape parameter, and θ is the scale parameter; n is the number of ocean region pixels of the SAR image; x i represents the i-th ocean region pixel; Γ(k) is a Gamma function.

[0103] The score function of the initial shape parameter and the initial scale parameter is constructed The formula is:

[0104]

[0105] Wherein, k0 is the initial shape parameter, and θ0 is the initial scale parameter; score k represents the score in the k direction; score θ represents the score in the θ direction; ψ(·) is a Digamma function, and satisfies

[0106] The Fisher information matrix F(k0, θ0) of the initial shape parameter and the initial scale parameter is constructed, and the formula is:

[0107]

[0108] Wherein, ψ (2) (·) is a trigamma function, and satisfies

[0109] Based on the score function of the initial shape parameter and the initial scale parameter and the Fisher information matrix, the Le Cam one-step estimation method is used to calculate the target shape parameter and the target scale parameter, and the formula is:

[0110]

[0111] wherein k1 is a target shape parameter, θ1 is a target scale parameter, F -1 (k0,θ0) is the inverse of the Fisher information matrix.

[0112] The present application can significantly reduce the computational burden under large scene SAR images while maintaining high parameter estimation accuracy and model fitting effect by only one iteration correction compared with the multiple iteration optimization process required in the traditional maximum likelihood estimation method.

[0113] S3: Based on the target shape parameter, the target scale parameter and the preset false alarm rate, the threshold value of each marine area pixel of the SAR image is calculated through the inverse Gamma distribution to obtain a threshold value map.

[0114] The constant false alarm rate detection criterion is based on the statistical hypothesis testing idea, and by constructing a statistical model of the marine area, the detection threshold T is adaptively adjusted under the premise of ensuring the fixed false alarm probability P FA .

[0115] Under the condition of Gamma background modeling, given the shape parameter k, the scale parameter θ and the preset false alarm rate P FA , the marine area pixel threshold value T can be calculated through the inverse cumulative distribution function of the Gamma distribution:

[0116]

[0117] wherein, is the inverse cumulative distribution function of the Gamma distribution.

[0118] S4: The SAR image and its corresponding threshold value map are compared to obtain a ship candidate area mask map.

[0119] The value of the down-sampled SAR image marine area pixel is compared with its corresponding threshold value, if the value of the down-sampled SAR image marine area pixel is greater than its corresponding threshold value, the pixel value is recorded as 1, otherwise as 0, to obtain a ship candidate area mask map.

[0120] Finally, the ship candidate area mask map is up-sampled to restore to the spatial resolution of the original SAR image, and the formula is:

[0121]

[0122] wherein Resize nn (·) represents image scaling using nearest neighbor interpolation; H b and W b are the height and width of the original SAR image respectively. Represents the mask image of the target ship candidate area after restoration to the original spatial resolution.

[0123] Preferably, in order to reduce computing resources and processing delays, this embodiment introduces a GPU tensor calculation framework in S2, uses the mean and variance of the Gamma distribution to obtain initial shape parameters and initial scale parameters, and then uses the Le Cam one-step estimation method based on the Fisher information matrix to correct the initial shape parameters and initial scale parameters, and calculate the target shape parameters and target scale parameters, specifically including:

[0124] S211: Multiple SAR images are input into the GPU as a batch, and the convolution kernel is used as a sliding window to divide each SAR image into multiple windows.

[0125] Multiple downsampled SAR images are input into the GPU as SAR image tensors, which are represented as Where B is the number of SAR images in a batch.

[0126] Convolution kernel element W h Expressed as:

[0127]

[0128] in, Represents a matrix of all 1s; convolution kernel W h The size is N h ×N v ; N h is the size of the convolution kernel in the row direction, N v is the size of the convolution kernel in the column direction.

[0129] Reference Figure 2 As shown, a convolution kernel is used to slide the SAR image to achieve subsequent parallel multi-window computation. The size of the convolution kernel is equivalent to the size of a detector window. In this embodiment, the convolution step size s is smaller than the window size, resulting in a small overlap between adjacent windows. This overlapping strategy effectively mitigates the problem of missed detection of ship targets located exactly at the boundary of adjacent non-overlapping windows, thereby enhancing coverage integrity of the target area.

[0130] S212: Normalize the pixels in the ocean area of ​​each window.

[0131] Considering that the Gamma distribution modeling requires the input to be non-negative continuous values ​​and is greatly affected by the echo amplitude scale, the pixels in the ocean area of ​​each window of the SAR image are normalized to the [0, 1] interval according to the downsampled binary sea and land mask map. The calculation formula is as follows:

[0132]

[0133] where x min and x max are the minimum and maximum value of the ocean region pixels in the current window respectively; x is the value of the current pixel; x' is the normalized value of the current pixel.

[0134] The normalized value range is consistent, which facilitates the numerical stability of the subsequent parameter estimation process. At the same time, this normalization process can be efficiently completed in parallel on the GPU side, without affecting the overall processing efficiency.

[0135] S213: Calculate the mean and variance of the normalized ocean region pixels in each window in parallel, to obtain the mean tensor and variance tensor, including the formula:

[0136]

[0137]

[0138] v[b, i, j] = m2[b, i, j] - m[b, i, j] 2

[0139] where m[b, i, j] is the value of the (i, j) position in the b-th SAR image in the mean tensor, corresponding to the mean of the i-th row and j-th column of the window ocean region pixels in the b-th SAR image; m2[b, i, j] is the value of the (i, j) position in the b-th SAR image in the squared mean tensor, corresponding to the squared mean of the i-th row and j-th column of the window ocean region pixels in the b-th SAR image; v[b, i, j] is the value of the (i, j) position in the b-th SAR image in the variance tensor, corresponding to the variance of the i-th row and j-th column of the window ocean region pixels in the b-th SAR image; N h is the size of the convolution kernel in the row direction, N v is the size of the convolution kernel in the column direction; W h [p, q] is the weight element at the (p, q) position in the convolution kernel W h ; I[b, si-p, sj-q] is the normalized value of the pixel at the (si-p, sj-q) position in the b-th SAR image in the i-th row and j-th column of the window; s is the convolution step size; represents rounding.

[0140] The output size of the mean tensor and the variance tensor is B x H' x W'. H' and W' can be calculated by the following formula:

[0141]

[0142] where N h is the size of the convolution kernel in the row direction, N vis the size of the convolution kernel in the column direction; H and W are the height and width of the input SAR image to the GPU, respectively; s is the convolution step.

[0143] The mean modeling process of the SAR image can be expressed as:

[0144] m = I * W h

[0145] The variance modeling process of the SAR image can be expressed as:

[0146] m2 = I 2 * W h , v = m2 - m 2

[0147] where I is the normalized SAR image tensor; W h is the window convolution kernel; m2 is the squared mean tensor of the SAR image; m is the mean tensor of the SAR image; and v is the variance tensor of the SAR image.

[0148] S214: Based on the mean tensor and the variance tensor, the initial shape parameter and the initial scale parameter of each window are calculated in parallel to obtain an initial shape parameter tensor and an initial scale parameter tensor, and the formula includes:

[0149]

[0150] where k0[b, i, j] is the value at the position coordinate (i, j) of the bth SAR image in the initial shape parameter tensor, corresponding to the initial shape parameter of the i-th row and j-th column window of the bth SAR image; θ0[b, i, j] is the value at the position coordinate (i, j) of the bth SAR image in the initial scale parameter tensor, corresponding to the initial scale parameter of the i-th row and j-th column window of the bth SAR image; ε is a constant used to prevent the denominator from being 0; m[b, i, j] is the value at the position coordinate (i, j) of the bth SAR image in the mean tensor; and v[b, i, j] is the value at the position coordinate (i, j) of the bth SAR image in the variance tensor.

[0151] The output sizes of the initial shape parameter tensor and the initial scale parameter tensor are both B x H' x W'.

[0152] S215: The gradient of the shape parameter and the scale parameter in the log-likelihood function is taken as a score function, and the score of each window in the shape parameter direction and the scale parameter direction is calculated in parallel based on the initial shape parameter tensor and the initial scale parameter tensor to obtain a shape parameter score tensor and a scale parameter score tensor.

[0153] The formula is:

[0154] score k [b, i, j] = d[b, i, j] - n[b, i, j] · ψ(k0[b, i, j]) - n[b, i, j] · log(θ0[b, i, j])

[0155] score θ [b, i, j] = n[b, i, j] · k0[b, i, j] · θ0[b, i, j] - n[b, i, j] · m[b, i, j]

[0156] wherein score k [b, i, j] is the value at the position coordinate (i, j) of the b-th SAR image in the shape parameter score tensor, corresponding to the score of the i-th row and j-th window of the b-th SAR image in the shape parameter direction; score θ [b, i, j] is the value at the position coordinate (i, j) of the b-th SAR image in the scale parameter score tensor, corresponding to the score of the i-th row and j-th window of the b-th SAR image in the scale parameter direction; ψ(·) is the Digamma function; d[b, i, j] is the sum of the logarithmic values of the ocean region pixels of the i-th row and j-th window of the b-th SAR image; n[b, i, j] is the number of the ocean region pixels of the i-th row and j-th window of the b-th SAR image; k0[b, i, j] is the value at the position coordinate (i, j) of the b-th SAR image in the initial shape parameter tensor; θ0[b, i, j] is the value at the position coordinate (i, j) of the b-th SAR image in the initial scale parameter tensor; m[b, i, j] is the value at the position coordinate (i, j) of the b-th SAR image in the mean value tensor.

[0157] The calculation formula of n[b, i, j] is as follows:

[0158]

[0159] wherein Map[b, si-p, sj-q] is the value at the position coordinate (si-p, sj-q) of the b-th down-sampled binary sea-land mask map in the sea-land mask map tensor; the size of the convolution kernel W1 is also N h × N v , and the convolution kernel element is Using the standard two-dimensional convolution operation, it can be expressed as n = Map * W1.

[0160] The calculation formula of d[b, i, j] is as follows:

[0161]

[0162] wherein the convolution kernel Using the standard two-dimensional convolution operation, it can be expressed as d = log(I) * W1.

[0163] S216: calculating the Fisher information matrix of each window in parallel based on the initial shape parameter tensor and the initial scale parameter tensor to obtain a Fisher information tensor, and the formula includes:

[0164] F 11 [b, i, j] = ψ (2) (k0[b, i, j])·n[b, i, j] + ε′

[0165] F 12 [b, i, j] = F 21 [b, i, j] = -θ0[b, i, j]·n[b, i, j]

[0166] F 22 [b, i, j] = k0[b, i, j]·θ0[b, i, j] 2 ·n[b, i, j] + ε′

[0167] Wherein, F 11 [b, i, j] is the component of the first row and the first column of F[b, i, j]; F 12 [b, i, j] is the component of the first row and the second column of F[b, i, j]; F 22 [b, i, j] is the component of the second row and the second column of F[b, i, j]; F[b, i, j] is the matrix at the position coordinate (i, j) of the bth SAR image in the Fisher information tensor, corresponding to the Fisher information matrix of the i-th row and the j-th column window of the bth SAR image; ψ(2)(·) is the trigamma function; ε′ is a small normal number introduced to avoid numerical instability.

[0168] Construct the determinant of the Fisher information tensor above:

[0169] det[b, i, j] = F 11 [b, i, j]·F 22 [b, i, j]-F 12 [b, i, j] 2

[0170] Further, the components of the inverse Fisher information matrix are calculated:

[0171]

[0172] Wherein, det[b, i, j] is the determinant at the position coordinate (i, j) of the bth SAR image in the Fisher information tensor, is the component of the first row and the first column of F-1[b, i, j]; is the component of the first row and the second column of F-1[b, i, j]; F[b, i, j] is the inverse matrix of F[b, i, j]. -1 [b, i, j] is the inverse matrix of F[b, i, j].

[0173] S217: Based on the shape parameter score tensor, the scale parameter score tensor and the Fisher information tensor, the target shape parameter and the target scale parameter of each window are calculated in parallel using the Le Cam one-step estimation method to obtain the target shape parameter tensor and the target scale parameter tensor, and the formula includes:

[0174]

[0175] Wherein, k1[b, i, j] is the value at the SAR image position coordinate (i, j) of the b-th in the target shape parameter tensor, corresponding to the target shape parameter of the i-th row and the j-th column window of the b-th SAR image; θ1[b, i, j] is the value at the SAR image position coordinate (i, j) of the b-th in the target scale parameter tensor, corresponding to the target scale parameter of the i-th row and the j-th column window of the b-th SAR image.

[0176] The output size of the target shape parameter tensor and the target scale parameter tensor is BxH'xW'.

[0177] Compared with the traditional constant false alarm rate detection method which relies on the serial sliding window strategy to estimate the parameters of the image region by region, the embodiment based on the parallel calculation mode of tensor expression unifies the sliding window parameter estimation process into a convolution form, and maps the window size to the tensor convolution kernel size. Under the GPU architecture, the high concurrency operation capability is fully utilized to realize the parallel solution of the window parameter estimation of multiple windows in the whole image range, which significantly reduces the calculation resource consumption and processing delay while ensuring the estimation accuracy.

[0178] Based on the GPU tensor calculation framework, the shape parameter tensor k1[b, i, j] and the scale parameter tensor θ1[b, i, j] are calculated, and then the threshold value of each ocean area pixel of the SAR image is calculated by the inverse Gamma distribution in S3 to obtain the threshold value map, which specifically includes:

[0179] Based on the target shape parameter tensor, the target scale parameter tensor and the preset false alarm rate, the threshold value of each window of the ocean area is calculated by the inverse Gamma distribution to obtain the threshold value tensor:

[0180]

[0181] T[b, i, j] is the value at the SAR image position coordinate (i, j) of the b-th in the threshold value tensor.

[0182] threshold tensor mapping and expanding back to the SAR image tensor Since there is area overlap due to the sliding window, in order to improve the detection accuracy and robustness, the embodiment adopts the strategy of taking the minimum value in the window overlap area to calculate the threshold value of each marine area pixel of the SAR image, and obtains a threshold map, which is expressed by the formula:

[0183]

[0184] wherein, T ext [b, i', j'] is the threshold value at the position coordinate (i', j') of the threshold map corresponding to the bth SAR image; T[b, m', n'] is the value at the position coordinate (m', n') in the threshold tensor, which corresponds to the threshold value of the m'th row andn'th column window of the bth SAR image; H' and W' are the height and width of the threshold tensor, respectively; R m′，n′ is the area of the m'th row andn'th column window in its corresponding SAR image.

[0185] R m′,n′ is calculated by the following formula:

[0186] R m′,n′ = {(i', j') | m's≤i'<m's+N h , n's≤j'<n's+N v}

[0187] wherein, N h is the size of the convolution kernel in the row direction, N v is the size of the convolution kernel in the column direction, and s is the step size of the convolution kernel.

[0188] The SAR image is compared with its corresponding threshold map to obtain a ship candidate area mask map, which is expressed by the formula:

[0189]

[0190] wherein, I[b, i', j'] is the normalized value at the position coordinate (i', j') of the bth SAR image; M[b, i', j'] is the value at the position coordinate (i', j') of the ship candidate area mask map corresponding to the bth SAR image.

[0191] The ship candidate area detection method for large scene SAR image provided by the application first generates a sea-land mask map for the SAR image by using a seabed topographic map, effectively eliminates the land area interference in the complex background of the SAR image, and improves the accuracy of modeling and the real-time performance of detection; then models the probability density of the sea area of the SAR image by using a Gamma distribution, calculates the shape parameter and the scale parameter required for modeling by combining with a Le Cam one-step estimation method, only performs one iteration correction, compared with the multi-round iteration optimization process required in the traditional maximum likelihood estimation method, the application can significantly reduce the calculation burden under the large scene SAR image, while maintaining high parameter estimation accuracy and model fitting effect, to meet the real-time requirement of ship candidate area detection; finally, the threshold of the sea area pixel is adaptively calculated by using a preset false alarm rate and an inverse Gamma distribution function, the ship candidate area is separated from the sea clutter by using the threshold, the ship candidate area mask map is obtained, and the detection of the ship candidate area is completed. The application only needs to set one hyperparameter of the false alarm rate, reduces the dependence on artificial parameter adjustment, and has stronger adaptability and universality.

[0192] Further, the application adopts a parallel processing architecture based on a GPU tensor calculation framework, uniformly represents the sliding window parameter estimation problem as a standard tensor convolution form, efficiently completes the normalization preprocessing, background modeling, parameter estimation and detection threshold generation in a window parallel processing manner in the full image range by mapping the window size to the convolution kernel size, significantly reduces the calculation resource overhead and processing delay while ensuring the estimation accuracy; finally, a fusion strategy of taking the minimum value in the overlapping area of multiple windows is introduced to realize pixel-level threshold expansion, which can map the threshold tensor to the size of the original SAR image, complete the detection of the ship candidate area, and further improve the hit rate of the ship candidate area detection while ensuring the detection real-time performance.

[0193] To verify the effectiveness of the ship candidate area detection method proposed by the application, the SAR image data in the XVIEW3 data set is used for experiment in this embodiment.

[0194] 1. Experimental data

[0195] The XVIEW3 dataset is a large-scale SAR dataset for marine remote sensing target detection, mainly used for ship identification and monitoring of illegal, unreported and unregulated fishing activities. The dataset contains 991 SAR images from the Sentinel-1 satellite, with an average size of 29400x24400 pixels, covering an area of about 4320 square kilometers of sea area. There are 243018 manually verified marine targets in the images, including ships and fixed offshore facilities, providing multi-task annotation information for ship detection and attribute identification, and accompanying auxiliary data such as seabed topography. This embodiment selects 503 SAR images divided by the verification set in the XVIEW3 dataset as input.

[0196] 2. Experimental setup

[0197] This embodiment is based on the PyTorch deep learning framework and is tested in a multi-GPU environment containing 4 NVIDIA RTX3090 graphics cards. To achieve efficient full-image parallel ship candidate region detection on large-sized remote sensing images, a constant false alarm rate detection method based on the Gamma distribution modeling and Le Cam one-step estimation method proposed in the present application is constructed, and a GPU-accelerated sliding window detector is built. The detector uses tensorized computing, fully utilizes the multi-card parallel computing capability, and improves the detection efficiency and throughput. The convolution kernel size (i.e. detector window size) is set to 512x512 pixels, the step size is 384 pixels, the downsampling ratio is set to 0.5, and the false alarm rate is set to 0.2.

[0198] The ship hit rate and the mask coverage rate are selected as evaluation indicators:

[0199] The ship hit rate is used to measure the recognition ability of the detection algorithm for real ship targets, and its calculation formula is:

[0200] r hit =n hit / n total

[0201] where n hit represents the number of ship targets contained in the generated ship candidate region mask; n total represents the total number of all manually annotated real ship targets in the image; the higher the hit rate, the stronger the positioning ability of the detection algorithm for ship targets.

[0202] The mask coverage rate is used to evaluate the sparseness of the detection region generated by the detection algorithm in the whole image, and its calculation formula is:

[0203]

[0204] where n w is the total number of effective detector windows; ai is the pixel area of the i-th detection window determined as a ship candidate region; a total is the total pixel area of the whole SAR image; the lower the mask coverage rate is, the more accurate the target region generated by the detector is, the less the calculation redundancy is, and the faster the speed is in the subsequent detection or classification based on the ship candidate region mask.

[0205] 3. Experimental results

[0206] Table 1 shows the experimental results of the contrast experiment in the sea-land mask extraction process by using the global water-land mask and the seabed topographic map respectively. The global water-land mask refers to the binary raster data used to divide the earth surface area into water bodies (such as oceans, lakes, rivers, etc.) and land (such as continents, islands, etc.) to assist in determining the water-land attribute of each pixel in the image. In this embodiment, the global water-land mask data is selected from the global water-land classification product version 6 released by the National Aeronautics and Space Administration, which is generated based on the multi-temporal vegetation index data collected by the Moderate Resolution Imaging Spectrometer carried by the Earth Observing Satellite Terra.

[0207] Table 1, experimental results of different methods in the sea-land mask extraction task

[0208]

[0209] The experimental results show that the ship target hit rate of the present application using the seabed topographic map as the basis for sea-land mask extraction is about 12.32% higher than that based on the global water-land mask, and the mask coverage rate is reduced by 2.56%, which indicates that this method has higher accuracy and regional focusing ability in target candidate region extraction. The reason is that the seabed topography has strong stability and long-term invariability, is not affected by external factors such as cloud cover, seasonal changes or tidal fluctuations, and does not have the "missing block" or "cloud invalid" area problem common in global water-land masks. In addition, sea-land mask extraction can effectively exclude about 28.15% of the land area, significantly reducing the computational burden in the subsequent constant false alarm rate detection stage.

[0210] Figure 3 is an example of a sea-land mask generated by using the seabed topographic map, Figure 4An example of a result of generating a sea-land mask using a global water-land mask. As can be seen from the two figures, the sea-land mask generated based on the sea floor topographic map is more accurate in coverage, covering only the sea area that actually has the possibility of ship activity; while the mask generated based on the global water-land mask product covers a large number of areas without ship targets including inland water bodies such as rivers and lakes, resulting in a significant increase in redundant areas. This redundant coverage not only causes unnecessary computational burden during subsequent constant false alarm rate detection, prolonging the reasoning time, but also easily introduces false alarms, thereby reducing the overall detection efficiency and system response speed. Therefore, compared with the global water-land mask scheme, the sea-land mask constructed using the sea floor topographic map is more targeted and practical, and performs better in terms of coverage control and computational efficiency.

[0211] Table 2 shows the comparison experimental results of the constant false alarm rate detection method based on Le Cam one-step estimation proposed in the present application with various advanced methods, including constant false alarm rate detection based on Gamma distribution fitting (CPU-Gamma-CFAR), cell average constant false alarm rate detection based on GPU acceleration (GPU-CA-CFAR), minimum constant false alarm rate detection based on GPU acceleration (GPU-SO-CFAR), and the CPU implementation version of the method of the present application (CPU-LeCam-CFAR).

[0212] Table 2, experimental results of different methods in constant false alarm rate task

[0213]

[0214] The experimental results show that the method proposed in the present application achieves a ship hit rate of 96.02%, maintaining similar recognition ability as other high-precision methods. In terms of processing efficiency, the average reasoning time for each large-size SAR image is only 1.15 seconds, which is about 1 / 900 of the time consumed by the traditional CPU-Gamma-CFAR method (939.70 seconds), and is also significantly better than other existing GPU-accelerated methods (4-5 seconds level), showing excellent real-time processing performance. In addition, when the method of the present application is implemented in CPU, it still has a fast processing speed (3.20 seconds / image), further verifying the universality and efficiency advantage of the method. In terms of mask coverage, the method of the present application is only 11.91%, which is significantly lower than the methods of GPU-CA-CFAR (26.09%) and GPU-SO-CFAR (29.77%), indicating that the detection mask generated by the method is more sparse and accurate, effectively reducing the redundant detection area, facilitating the acceleration and resource saving of subsequent processing operations.

[0215] Figure 5 is an example of magnifying the SAR image containing ship area in XVIEW3, Figure 6is an enlarged example diagram of the generated ship candidate region mask. As can be seen from the diagram, the ship candidate region mask generated by the application accurately covers the real ship target in the SAR image in space, and has good target matching capability. At the same time, the mask presents a low sparsity characteristic in the whole large scene image, does not contain a large number of invalid background regions, and effectively avoids redundant detection.

[0216] In summary, the application takes into account the compactness of the mask and the calculation efficiency on the basis of ensuring the accuracy of the candidate region detection, and has a significant comprehensive performance advantage in the ship candidate region detection task of large-size SAR remote sensing images, and has good practical value and engineering promotion potential.

[0217] Based on the above-mentioned ship candidate region detection method for large scene SAR images, the application also provides a ship candidate region detection system for large scene SAR images, comprising:

[0218] An ocean area extraction module is configured to obtain a seabed topographic map corresponding to the SAR image, generate a sea-land mask map according to the seabed topographic map, and extract an ocean area of the SAR image.

[0219] A parameter calculation module is configured to model the ocean area of the SAR image based on a Gamma distribution, obtain an initial shape parameter and an initial scale parameter by using a mean value and a variance of pixels of the ocean area of the SAR image, and correct the initial shape parameter and the initial scale parameter by using a Le Cam one-step estimation method based on a Fisher information matrix to obtain a target shape parameter and a target scale parameter.

[0220] A threshold calculation module is configured to calculate a threshold value of each pixel of the ocean area of the SAR image by using an inverse Gamma distribution based on the target shape parameter, the target scale parameter and a preset false alarm rate to obtain a threshold value map.

[0221] A ship candidate region detection module is configured to compare the SAR image with the threshold value map corresponding thereto to obtain a ship candidate region mask map.

[0222] Those skilled in the art should understand that the embodiments of the application can be provided as a method, a system or a computer program product. Therefore, the application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0223] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0224] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0225] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0226] Obviously, the above-described embodiments are only examples for clarity of description and are not limiting on the embodiments. Based on the above description, one of ordinary skill in the art can further make other different forms of changes or modifications. Here, all the embodiments are not required to be enumerated, and the obvious changes or modifications derived therefrom are still within the protection scope of the present application.

Claims

1. A method for detecting candidate ship regions in large-scale SAR images, characterized in that: include: Obtain the seabed topography map corresponding to the SAR image, generate a sea and land mask map based on the seabed topography map, and extract the ocean area of ​​the SAR image; The ocean area of ​​the SAR image is modeled based on the Gamma distribution. The initial shape parameters and initial scale parameters are obtained using the mean and variance of the pixels in the ocean area of ​​the SAR image. The initial shape parameters and initial scale parameters are then corrected using the Le Cam one-step estimation method based on the Fisher information matrix to obtain the target shape parameters and target scale parameters. Based on the target shape parameters, target scale parameters and preset false alarm rate, the threshold of each ocean area pixel of the SAR image is calculated through the inverse Gamma distribution to obtain the threshold map; The SAR image is compared with its corresponding threshold map to obtain the mask map of the ship candidate area.

2. The method for detecting candidate ship regions in large-scale SAR images according to claim 1, characterized in that: The initial shape parameters and initial scale parameters are obtained using the mean and variance of the pixels in the ocean area of ​​the SAR image. The initial shape parameters and initial scale parameters are then corrected using the Le Cam one-step estimation method based on the Fisher information matrix to obtain the target shape parameters and target scale parameters, including: Calculating the mean and variance of pixels in the ocean area of ​​the SAR image, and calculating initial shape parameters and initial scale parameters based on the mean and variance; The gradient of the log-likelihood function with respect to the shape parameter and the scale parameter is used as the scoring function to construct the score function of the initial shape parameter and the initial scale parameter; the Fisher information matrix of the initial shape parameter and the initial scale parameter is constructed; based on the score function and the Fisher information matrix of the initial shape parameter and the initial scale parameter, the target shape parameter and the target scale parameter are calculated using the Le Cam one-step estimation method.

3. The method for detecting candidate ship regions in large-scale SAR images according to claim 2, characterized in that: Construct a scoring function for the initial shape parameters and initial scale parameters The formula is: Among them, k is the shape parameter, θ is the scale parameter; k0 is the initial shape parameter, θ0 is the initial scale parameter; score k Indicates the score in the k direction; score θ represents the score in the θ direction; n is the number of pixels in the ocean area of ​​the SAR image; x i represents the pixel in the i-th ocean area; ψ(·) is the Digamma function; Construct the Fisher information matrix F(k0,θ0) of the initial shape parameters and initial scale parameters, the formula is: Among them, ψ (2) (·) is the trigamma function; Based on the score function of the initial shape parameter and initial scale parameter and the Fisher information matrix, the target shape parameter and target scale parameter are calculated using the Le Cam one-step estimation method. The formula is: Among them, k1 is the target shape parameter, θ1 is the target scale parameter, F -1 (k0,θ0) is the inverse matrix of the Fisher information matrix.

4. The method for detecting candidate ship regions in large-scale SAR images according to claim 2, wherein: Based on the GPU tensor computing framework, the initial shape parameters and initial scale parameters are obtained using the mean and variance of the pixels in the ocean area of ​​the SAR image. The initial shape parameters and initial scale parameters are then corrected using the Le Cam one-step estimation method based on the Fisher information matrix to calculate the target shape parameters and target scale parameters, including: Multiple SAR images are input into the GPU as a batch, and the convolution kernel is used as a sliding window to divide each SAR image into multiple windows; Normalize the pixels in the ocean area of ​​each window; Parallel calculation of the mean and variance of the normalized ocean area pixels in each window to obtain the mean tensor and variance tensor; Based on the mean tensor and variance tensor, the initial shape parameter and initial scale parameter of each window are calculated in parallel to obtain the initial shape parameter tensor and the initial scale parameter tensor; The gradient of the log-likelihood function with respect to the shape parameter and scale parameter is used as the scoring function. Based on the initial shape parameter tensor and the initial scale parameter tensor, the scores of each window in the shape parameter direction and the scale parameter direction are calculated in parallel to obtain the shape parameter score tensor and the scale parameter score tensor. Based on the initial shape parameter tensor and the initial scale parameter tensor, the Fisher information matrix of each window is calculated in parallel to obtain the Fisher information tensor; Based on the shape parameter score tensor, scale parameter score tensor and Fisher information tensor, the target shape parameter and target scale parameter of each window are calculated in parallel using the LeCam one-step estimation method. Get the target shape parameter tensor and target scale parameter tensor.

5. The method for detecting candidate ship regions in large-scale SAR images according to claim 4, characterized in that: The mean and variance of the pixels in each window ocean area are calculated in parallel to obtain the mean tensor and variance tensor. The formulas include: v[b,i,j]=m2[b,i,j]-m[b,i,j] 2 Where m[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the mean tensor, corresponding to the mean of the pixels in the ocean area of ​​the window in the i-th row and j-th column of the b-th SAR image; m2[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the square mean tensor, corresponding to the mean of the square of the pixels in the ocean area of ​​the window in the i-th row and j-th column of the b-th SAR image; v[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the variance tensor, corresponding to the variance of the pixels in the ocean area of ​​the window in the i-th row and j-th column of the b-th SAR image; N h is the size of the convolution kernel in the row direction, N v is the size of the convolution kernel in the column direction; W h [p,q] is the convolution kernel W h The weight element at the position coordinate (p,q) in the middle; I[b,si-p,sj-q] is the normalized value of the pixel at the coordinate position (si-p,sj-q) of the b-th SAR image in the window of the i-th row and the j-th column; s is the convolution step size, and the step size is smaller than the convolution kernel size; Indicates rounding.

6. The method for detecting candidate ship regions in large-scale SAR images according to claim 5, characterized in that: Based on the mean tensor and variance tensor, the initial shape parameters and initial scale parameters of each window are calculated in parallel to obtain the initial shape parameter tensor and initial scale parameter tensor. The formulas include: Where k0[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the initial shape parameter tensor, corresponding to the initial shape parameter of the window in the i-th row and j-th column of the b-th SAR image; θ0[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the initial scale parameter tensor, corresponding to the initial scale parameter of the window in the i-th row and j-th column of the b-th SAR image; ε is a constant used to prevent the denominator from being zero. m[b,i,j] is the value of the b-th SAR image position coordinate (i,j) in the mean tensor; v[b,i,j] is the value of the b-th SAR image position coordinate (i,j) in the variance tensor.

7. The method for detecting candidate ship regions in large-scale SAR images according to claim 6, wherein: Based on the initial shape parameter tensor and the initial scale parameter tensor, the scores of each window in the shape parameter direction and the scale parameter direction are calculated in parallel to obtain the shape parameter score tensor and the scale parameter score tensor. The formula includes: score k [b,i,j]=d[b,i,j]-n[b,i,j]·ψ(k0[b,i,j])-n[b,i,j]·log(θ0[b,i,j]) score θ [b,i,j]=n[b,i,j]·0[b,i,j]·θ0[b,i,j]-n[b,i,j]·m[b,i,j] Among them, score k [b,i,j] is the value of the b-th SAR image position coordinate (i,j) in the shape parameter score tensor, corresponding to the score of the window in the i-th row and j-th column of the b-th SAR image in the shape parameter direction; score θ [b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the scale parameter score tensor, corresponding to the score of the b-th SAR image window in the i-th row and j-th column in the scale parameter direction; ψ(·) is the Digamma function; d[b,i,j] is the logarithmic sum of the pixels in the ocean area of ​​the i-th row and j-th column window of the b-th SAR image; n[b,i,j] is the number of pixels in the ocean area of ​​the i-th row and j-th column window of the b-th SAR image; k0[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the initial shape parameter tensor; θ0[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the initial scale parameter tensor; m[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the mean tensor; Based on the initial shape parameter tensor and the initial scale parameter tensor, the Fisher information matrix of each window is calculated in parallel to obtain the Fisher information tensor. The formula includes: F 11 [b,i,j]=ψ (2) (k0[b,i,j])·n[b,i,j]+ε' F 12 [b,i,j]=F 21 [b,i,j]=-θ0[b,i,j]·n[b,i,j] F 22 [b,i,j]=k0[b,i,j]·θ0[b,i,j] 2 ·n[b,i,j]+ε' Among them, F 11 [b,i,j] is the component of the first row and first column of F[b,i,j]; F 12 [b,i,j] is the component of the first row and second column of F[b,i,j]; F 22 [b,i,j] is the component of the second row and second column of F[b,i,j]; F[b,i,j] is the matrix at the position coordinate (i,j) of the b-th SAR image in the Fisher information tensor, corresponding to the Fisher information matrix of the window in the i-th row and j-th column of the b-th SAR image; ψ (2) (·) is the trigamma function; ε' is a positive constant.

8. The method for detecting candidate ship regions in large-scale SAR images according to claim 7, characterized in that: Based on the shape parameter score tensor, scale parameter score tensor and Fisher information tensor, the target shape parameter and target scale parameter of each window are calculated in parallel using the Le Cam one-step estimation method to obtain the target shape parameter tensor and target scale parameter tensor. The formulas include: Wherein, k1[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the target shape parameter tensor, corresponding to the target shape parameter of the window in the i-th row and j-th column of the b-th SAR image; θ1[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the target scale parameter tensor, corresponding to the target scale parameter of the window in the i-th row and j-th column of the b-th SAR image; k0[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the initial shape parameter tensor; θ0[b,i,j] is the value at the b-th SAR image position coordinate (i,j) in the initial scale parameter tensor; score k [b,i,j] is the value of the b-th SAR image position coordinate (i,j) in the shape parameter score tensor; score θ [b,i,j] is the value of the b-th SAR image position coordinate (i,j) in the scale parameter score tensor; F -1 [b,i,j] is the component of the first row and first column; F -1 [b,i,j] is the component in the first row and second column; F -1 [b,i,j] is the component of the second row and second column; F[v,i,j] is the matrix at the b-th SAR image position coordinate (i,j) in the Fisher information tensor, F -1 [b,i,j] is the inverse matrix of F[b,i,j].

9. The method for detecting candidate ship regions in large-scale SAR images according to claim 4, characterized in that: Based on the target shape parameter tensor, target scale parameter tensor and preset false alarm rate, the threshold of the ocean area of ​​each window is calculated by inverse Gamma distribution to obtain the threshold tensor. The threshold of each pixel in the ocean area of ​​the SAR image is calculated by taking the minimum value of the window overlapping area to obtain the threshold map, which is expressed as follows: Among them, T ext [b,i',j'] is the threshold value at the position coordinate (i',j') of the threshold map corresponding to the b-th SAR image; T[b,m',n'] is the value at the position coordinate (m',n') of the b-th SAR image in the threshold tensor, corresponding to the threshold value of the window at the m'th row and n'th column of the b-th SAR image; H' and W' are the height and width of the threshold tensor respectively; R m',n' is the area of ​​the window in the m'th row and n'th column in its corresponding SAR image.

10. A ship candidate area detection system for large-scale SAR images, characterized by: include: The ocean area extraction module is used to obtain the seabed topography map corresponding to the SAR image, generate a sea and land mask map based on the seabed topography map, and extract the ocean area of ​​the SAR image; The parameter calculation module is used to model the ocean area of ​​the SAR image based on the Gamma distribution, obtain the initial shape parameters and initial scale parameters using the mean and variance of the pixels in the ocean area of ​​the SAR image, and then use the Le Cam one-step estimation method based on the Fisher information matrix to correct the initial shape parameters and initial scale parameters to obtain the target shape parameters and target scale parameters; The threshold calculation module is used to calculate the threshold of each ocean area pixel in the SAR image through the inverse Gamma distribution based on the target shape parameters, target scale parameters and preset false alarm rate to obtain a threshold map; the ship candidate area detection module is used to compare the SAR image with its corresponding threshold map to obtain a ship candidate area mask map.

Images