A wide swath SAR image low false alarm target detection method combined with shape confidence

Abstract

The application discloses a wide-width SAR image low false alarm target detection method combined with shape confidence, and comprises the following steps: acquiring a SAR image; sequentially performing constant false alarm detection and morphological processing on the SAR image to obtain a detection image; inputting the detection image into a trained discrimination network to obtain a first discrimination image; performing segmentation processing of the SAR image into sea and land, and performing discrimination on the first discrimination image again based on the segmentation result of the sea and land to remove the land part and obtain a second discrimination image; and finally discriminating the second discrimination image based on shape confidence to complete target detection of the SAR image. The application can effectively distinguish between a ship target and a false target, realize low false alarm detection of the ship target, and improve the discrimination accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar image processing, and particularly relates to a wide-width SAR image low false alarm target detection method combined with shape confidence. BACKGROUND

[0002] As an active microwave sensor, synthetic aperture radar (SAR) can realize real-time remote detection of stationary sensitive targets (such as tanks, armored vehicles, ships, etc.) at all times and in all weather, and has important application value in military and civilian fields. SAR automatic target recognition (ATR) technology is an important branch of SAR image interpretation, aiming to effectively obtain and identify targets of interest from complex ground scenes. SAR automatic target recognition can quickly and accurately complete the specified SAR image target recognition task through a computer, and therefore has attracted widespread attention.

[0003] The existing SAR image automatic target recognition method usually adopts a three-stage processing flow, that is, firstly, a region of interest (ROI) is extracted from the entire SAR image; then, the region is preprocessed to eliminate non-target regions; finally, feature extraction is performed on the target region, and the extracted features are used for target classification and recognition. The target detection and identification stages are used to obtain the ROI, and the quality of the ROI obtained by the two stages directly affects the task complexity of the recognition stage.

[0004] At present, a lot of researches have been made on SAR image target detection, and in the constant false alarm rate (CFAR) detection method commonly used in the current SAR ship rapid detection, low false alarm has always been the research focus. Xiong Wei et al. proposed a high-resolution SAR image ship detection method based on a support vector machine (SVM) according to the characteristics of high-resolution marine background and targets. After the image is divided and screened using the support vector machine (SVM), the screened region is finely detected using the maximum entropy threshold (KSW). This method has good detection effect on TerraSAR-X data, but when there is a large area of land, the false alarm rate will increase, and the detection effect is greatly affected by the land environment factor.

[0005] In addition, for ship detection in coastal images, GAO Bing et al. proposed a two-step hierarchical detection method. After target rough detection based on K-lognormal mixed distribution constant false alarm detection, false targets are removed according to the motion characteristics and focusing effect of the ship itself. This method has high detection efficiency in complex sea surface conditions, but for SAR images with complex and discontinuous coastlines and unclear contours, it cannot effectively remove the land edge false alarm. Jiao et al. proposed an improved SAR image ship detection algorithm based on Faster R-CNN, which solves the detection problem in multiple scenes and multiple scales, but has the problems of difficulty in obtaining and labeling SAR images, large network model computation, and is not conducive to real-time detection.

[0006] Therefore, how to provide a target detection method for SAR images with large areas of land and complex near-port scenes has become a problem to be solved. SUMMARY

[0007] In order to solve the above problems existing in the prior art, the present application provides a wide SAR image low false alarm target detection method combined with shape confidence. The technical problem to be solved by the present application is solved by the following technical scheme:

[0008] A wide SAR image low false alarm target detection method combined with shape confidence, the target detection method comprising:

[0009] Step 1, obtaining a SAR image;

[0010] Step 2, performing constant false alarm detection and morphological processing on the SAR image in turn to obtain a detection image;

[0011] Step 3, inputting the detection image into a trained discrimination network to obtain a first discrimination image, the discrimination network being used for discriminating false alarms due to clutter;

[0012] Step 4, performing sea and land segmentation processing on the SAR image, and based on the sea and land segmentation result, performing discrimination again on the first discrimination image to remove the land part to obtain a second discrimination image;

[0013] Step 5, performing final discrimination on the second discrimination image based on shape confidence to complete target detection of the SAR image.

[0014] In an embodiment of the present application, the step 2 comprises:

[0015] Step 2.1, performing constant false alarm detection on the SAR image by using a constant false alarm detection method to obtain a binary image;

[0016] Step 2.2, morphological processing is performed on the binary image to remove connected domains that do not meet preset conditions, to obtain a detection image.

[0017] In an embodiment of the present application, the discrimination network comprises 5 layers of convolution layers, 3 layers of maximum pooling layers, and 1 layer of Dropout layer, wherein the first layer of convolution layer conv1, the first layer of maximum pooling layer MaxPool1, the second layer of convolution layer conv2, the second layer of maximum pooling layer MaxPool2, the third layer of convolution layer conv3, the third layer of maximum pooling layer MaxPool3, the fourth layer of convolution layer conv4, the Dropout layer, and the fifth layer of convolution layer are sequentially connected in series.

[0018] In an embodiment of the present application, the step 4 comprises:

[0019] Step 4.1, judging whether the SAR image contains a need for sea and land segmentation processing by using latitude and longitude, if needed, performing median filtering processing on the SAR image to obtain a filtered SAR image, and performing step 4.2 on the filtered SAR image;

[0020] Step 4.2, comparing the pixel value of each pixel point of the gray histogram of the filtered SAR image with two randomly initialized numerical values, and clustering the pixel value of each pixel point in the gray histogram with the numerical value closer to it into a class to obtain a first clustering result and a second clustering result;

[0021] Step 4.3, calculating the pixel mean of the first clustering result and the pixel mean of the second clustering result, comparing the pixel value of each pixel point of the gray histogram of the filtered SAR image with the pixel mean of the first clustering result and the pixel mean of the second clustering result, and clustering the pixel value of each pixel point in the gray histogram with the pixel mean closer to it into a class to obtain a third clustering result and a fourth clustering result;

[0022] Step 4.4, comparing the pixel mean of the third clustering result and the pixel mean of the fourth clustering result with the pixel mean of the two clustering results of the last iteration, if the pixel means are the same, stopping iteration to obtain a sea-land segmentation binary image, if the pixel means are not the same, continuing to compare the pixel value of each pixel point of the gray histogram of the filtered SAR image with the pixel mean of the third clustering result and the pixel mean of the fourth clustering result until the pixel means of the two iterations are the same, to obtain a sea-land segmentation binary image;

[0023] Step 4.5, mapping the first discrimination image into the sea-land segmentation binary image to obtain a mapping image;

[0024] Step 4.6, respectively calculate the proportion of sea and the proportion of land occupied by the detection frame in the corresponding region of the mapping image;

[0025] Step 4.7, judge the relationship between the proportion of sea and the proportion of land and the first threshold value, if the proportion of land is greater than the first threshold value, it is a false alarm of the land part, then remove the region corresponding to the detection frame from the mapping image, if the proportion of sea is greater than the first threshold value, retain the region of the mapping image corresponding to the detection frame, to obtain the second discrimination image.

[0026] In an embodiment of the present application, the step 5 comprises:

[0027] Step 5.1, map the second discrimination image to the SAR image, extract the region of the SAR image mapped by the second discrimination image to obtain a target suspected slice;

[0028] Step 5.2, perform logarithmic transformation on the target suspected slice to obtain a transformed image;

[0029] Step 5.3, perform target region segmentation on the transformed image to obtain a segmentation image;

[0030] Step 5.4, perform morphological processing on the segmentation image to obtain a discrimination slice image;

[0031] Step 5.5, extract the maximum distance feature, target diameter feature and count feature from the discrimination slice image;

[0032] Step 5.6, determine the shape confidence of the discrimination slice image based on the discrimination fusion feature value obtained by the maximum distance feature, the target diameter feature and the count feature, compare the shape confidence with a third threshold value, if the shape confidence is less than the third threshold value, it is a target, if the shape confidence is greater than the third threshold value, it is a clutter, and the target detection of the SAR image is completed.

[0033] In an embodiment of the present application, the step 5.3 comprises:

[0034] Judge the relationship between each pixel value of the transformed image and a comparison condition, if the pixel value of the transformed image is less than the comparison condition, the pixel value is a clutter, and the clutter part is removed, if the pixel value of the transformed image is greater than or equal to the comparison condition, the pixel value is a target, after comparing all pixel values of the transformed image with the comparison condition, the segmentation image is obtained, and the comparison condition is μ+pσ, wherein μ is the pixel mean value of the transformed image, p is a constant, and σ is the standard deviation of the transformed image.

[0035] In an embodiment of the present application, the maximum distance feature is the maximum distance from all target pixels in the identification slice image to the centroid of the target region.

[0036] The target diameter feature is the diagonal length of the minimum circumscribed rectangle of the target region of the identification slice image.

[0037] The count feature is the proportion of the number of pixels with intensity greater than a second threshold in the target region of the identification slice image to the total number of pixels in the target region.

[0038] In an embodiment of the present application, the calculation formula of the identification fusion feature value is:

[0039]

[0040] Wherein, f is the identification fusion feature value, distance is the maximum distance feature, diameter is the target diameter feature, and count is the count feature.

[0041] In an embodiment of the present application, the calculation formula of the shape confidence is:

[0042]

[0043] Wherein, P conf is the shape confidence.

[0044] The present application has the following beneficial effects:

[0045] 1. The present application is aimed at the complex land and near-port scene in wide SAR images, and the land part may have inconsistent light and shade. After constant false alarm detection, a large number of false targets may appear in these places. At the same time, the port at the edge of the land may also produce false targets similar to the target due to the complex near-port scene. The use of sea-land segmentation results combined with a shape confidence based on identification features can effectively distinguish between ship targets and false targets, and realize low false alarm detection of ship targets.

[0046] 2. The present application takes advantage of the identification network in the classification problem, and converts the identification of the target detection frame and the clutter detection frame into the classification of the slice for processing. The slice data set is made to train the network, and the cost-sensitive learning method is used to solve the imbalance problem of the training sample. Compared with the traditional method, the accuracy of identification is improved. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a flowchart of a wide SAR image low false alarm target detection method combined with shape confidence provided by an embodiment of the present application;

[0048] Figure 2 is a flowchart of another shape confidence combined wide swath SAR image low false alarm target detection method provided by an embodiment of the present application;

[0049] Figure 3 is a framework diagram of a discrimination network provided by an embodiment of the present application;

[0050] Figure 4 is a distribution diagram of a discrimination fusion feature value provided by an embodiment of the present application;

[0051] Figure 5 is a function plot of shape confidence change with feature value provided by an embodiment of the present application. DETAILED DESCRIPTION

[0052] The present application will be further described in detail below with specific embodiments, but the embodiments of the present application are not limited thereto.

[0053] Embodiment One

[0054] Please refer to Figure 1 and Figure 2 , Figure 1 is a flowchart of a shape confidence combined wide swath SAR image low false alarm target detection method provided by an embodiment of the present application, Figure 2 is a flowchart of another shape confidence combined wide swath SAR image low false alarm target detection method provided by an embodiment of the present application. The present application proposes a shape confidence combined wide swath SAR image low false alarm target detection method, which is a traditional target detection method for wide swath SAR images with large area land and complex near-port scenes, and the method can reduce the false alarm rate. The shape confidence combined wide swath SAR image low false alarm target detection method proposed by the present application includes steps 1-5, wherein:

[0055] Step 1, obtaining a SAR image, which is a wide swath SAR image.

[0056] Step 2, sequentially performing constant false alarm rate (CFAR) detection and morphological processing on the SAR image to obtain a detection image.

[0057] Specifically, first, the SAR image is subjected to constant false alarm rate detection to preliminarily detect the targets and false alarms of the SAR image, and then the morphological processing is performed on the image after the constant false alarm rate detection to determine the preliminary detection result.

[0058] In this embodiment, step 2 can specifically include:

[0059] Step 2.1: Perform constant false alarm rate (CFAR) detection on the SAR image using the constant false alarm rate (CFAR) detection method to obtain a binary image.

[0060] Specifically, for pixel-by-pixel two-dimensional CFAR detection of wide-span SAR images, this embodiment uses an adaptive window-based CFAR algorithm. For a pixel to be detected in the input SAR image, a hollow sliding window is used to obtain the background clutter pixels around the pixel to be detected. A warning window also needs to be set around the pixel to prevent the target pixel from leaking into the background window. For example, if the maximum size of the ship is estimated to be L, the size of the warning window is set to 2.

[0061] Furthermore, for the pixel to be detected x i Based on the selected clutter statistical distribution probability model c(x) and false alarm rate P... fa By sliding the hollow window, the detection threshold T corresponding to the target to be detected is calculated, and then (the pixel to be detected x) is... i The pixel value (mean of background window pixels) / standard deviation of the background window is compared with the detection threshold T. If it is greater than the detection threshold T, then the pixel to be detected is... i For the target pixel, if it is smaller than the detection threshold T, then the pixel to be detected is x. i These pixels represent clutter background pixels, thus ultimately achieving constant false alarm rate (CFAR) detection for the entire SAR image.

[0062] Step 2.2: Perform morphological processing on the binary image to remove connected components that do not meet the preset conditions, and obtain the detection image.

[0063] Specifically, morphological processing is performed on the binary image after CFAR detection, specifically using image opening and closing operations. By setting a reasonable size of the structuring element for the morphological operation, suspected targets with small connected components are removed, i.e., noise isolated points are removed. These noise isolated points are connected components that do not meet the preset conditions, thus obtaining the preliminary detection result, i.e., the detection image.

[0064] Step 3: Input the detected image into the trained discrimination network to obtain the first discrimination image. The discrimination network is used to identify false alarms caused by clutter.

[0065] Specifically, the trained discrimination network is used to discriminate the detection image obtained in step 2, eliminating a large number of false alarms caused by clutter, and obtaining the first discrimination image.

[0066] In this embodiment, please refer to Figure 3The discrimination network consists of 5 convolutional layers, 3 max pooling layers, and 1 dropout layer. The first convolutional layer (conv1), the first max pooling layer (MaxPool1), the second convolutional layer (conv2), the second max pooling layer (MaxPool2), the third convolutional layer (conv3), the third max pooling layer (MaxPool3), the fourth convolutional layer (conv4), the dropout layer, and the fifth convolutional layer are connected in series.

[0067] The first convolutional layer, conv1, takes a 128×128 image as input. The parameter settings for all convolutional layers are as follows: Figure 3 As shown, after the convolutional operations of the first four layers, a non-linear mapping is performed using the ReLU activation function. Finally, the fifth convolutional layer, conv5, outputs a two-dimensional vector to determine whether it is a target or clutter. The mathematical expression of the ReLU activation function is:

[0068] f(x) = ax(0,x).

[0069] Where f(x) is the ReLU activation function.

[0070] In this embodiment, a method for training an identification network is also provided, the training method comprising:

[0071] S1. Create a slice dataset from the publicly available SSDD dataset and divide the slices into ship slices and clutter slices for use in the training process of the discrimination network.

[0072] Because the slice sizes are inconsistent, in order to ensure normal training results, all slices are preprocessed. The size of the slices is modified without changing the aspect ratio of the target. Specifically, interpolation can be used to fill the empty parts with the smallest pixel value in the slice, so that the slice size in the training dataset is uniform.

[0073] S2. The discrimination network constructed above is trained using the above training dataset. Since there is a large difference in the number of training samples, the training loss ratio is modified using a cost-sensitive learning method when training the discrimination network to solve the problem of imbalanced training samples and obtain the optimal network weights.

[0074] Step 4: Perform ocean and land segmentation on the SAR image, and based on the results of ocean and land segmentation, perform further identification on the first identification image to remove the land portion, thereby obtaining the second identification image.

[0075] In one specific embodiment, step 4 may specifically include:

[0076] Step 4.1: Use latitude and longitude to determine whether the SAR image contains elements that need to be segmented into ocean and land. If so, perform median filtering on the SAR image to obtain the filtered SAR image. Then, execute step 4.2 on the filtered SAR image. If not, proceed directly to step 5.

[0077] Specifically, it can be determined whether the SAR image in this embodiment contains land by using a map library file that includes global land and sea latitude and longitude information. If it does, the SAR image needs to be processed by median filtering. If it does not, step 5 is executed directly.

[0078] Step 4.2: Compare the pixel values ​​of the grayscale histogram of the filtered SAR image with two randomly initialized values, and cluster the pixel values ​​of each pixel in the grayscale histogram with the values ​​that are closer to it into one class to obtain the first clustering result and the second clustering result.

[0079] Step 4.3: Calculate the pixel mean of the first clustering result and the pixel mean of the second clustering result. Compare the pixel values ​​of the gray-level histogram of the filtered SAR image with the pixel mean of the first clustering result and the pixel mean of the second clustering result. Cluster the pixel value of each pixel in the gray-level histogram with the pixel mean of the nearest pixel to obtain the third clustering result and the fourth clustering result.

[0080] Step 4.4: Compare the pixel mean values ​​of the third and fourth clustering results with the pixel mean values ​​of the two clustering results from the previous iteration. If the pixel mean values ​​are the same, stop the iteration and obtain the land-sea segmentation binary image. If the pixel mean values ​​are different, continue to compare the pixel values ​​of the grayscale histogram of the filtered SAR image with the pixel mean values ​​of the third and fourth clustering results until the pixel mean values ​​of the two iterations are the same, and obtain the land-sea segmentation binary image.

[0081] Step 4.5: Map the first identification image onto the land-sea segmentation binary image to obtain the mapped image.

[0082] Step 4.6: Calculate the proportion of the ocean and the proportion of the land in the corresponding regions of the detection box on the mapped image.

[0083] Specifically, since there are many connected components after CFAR detection, each connected component on the mapped image is enclosed in a detection box, and then the proportion of ocean and land occupied in each detection box is calculated.

[0084] Step 4.7: Determine the relationship between the proportions of ocean and land occupation and the first threshold. If the proportion of land occupation is greater than the first threshold, it is a false alarm for the land portion, and the region corresponding to the detection box containing the land portion is removed from the mapped image. If the proportion of ocean occupation is greater than or equal to the first threshold, the region of the mapped image corresponding to the detection box is retained, and the retained region is used as the second discrimination image. The first threshold can be set according to the size of the ship target.

[0085] Step 5: Perform final identification on the second discrimination image based on shape confidence to complete target detection of the SAR image.

[0086] In one specific embodiment, step 5 may specifically include:

[0087] Step 5.1: Map the second identification image onto the SAR image, extract the region of the SAR image onto which the second identification image is mapped, and obtain the suspected target slice.

[0088] Step 5.2: Perform a logarithmic transformation on the suspected target slice to obtain the transformed image.

[0089] Specifically, a logarithmic transformation is performed on the suspected target slice, which converts multiplicative noise in the SAR image into additive noise, expands low grayscale values, compresses high grayscale values, and adjusts the pixel dynamic range to obtain the transformed image G.

[0090] Step 5.3: Perform target region segmentation on the transformed image to obtain a segmented image.

[0091] Specifically, the relationship between each pixel value of the transformed image and the comparison condition is determined. If the pixel value of the transformed image is less than the comparison condition, the pixel value is considered clutter and is removed. If the pixel value of the transformed image is greater than or equal to the comparison condition, the pixel value is considered the target. All pixel values ​​of the transformed image are compared with the comparison condition to obtain the segmented image. The comparison condition is μ + pσ, where μ is the mean pixel value of the transformed image, p is a constant set empirically, and σ is the standard deviation of the transformed image.

[0092] In other words, the transformed image is segmented into its target region:

[0093]

[0094] Here, 1 represents the segmented binary image. When the value is greater than the comparison condition, pixel (i,j) is judged as the target and G1(i,j) is 1; otherwise, pixel (i,j) is judged as clutter and G1(i,j) is 0.

[0095] Step 5.4: Perform morphological processing on the segmented image to remove connected components that do not meet the preset conditions, and obtain the identification slice image.

[0096] Specifically, the segmented image obtained in the above steps is subjected to morphological processing, specifically using image opening and closing operations to remove isolated noise points with high intensity and fill in the parts of the target region with weak scattering. Since the size of ships in the ocean is inconsistent, the size of the connected components is set according to empirical values ​​to traverse all connected components. Parts smaller than the threshold are judged as clutter, and finally, the preprocessed discriminative slice image is obtained.

[0097] Step 5.5: Extract the maximum distance feature, target diameter feature, and count feature from the identification slice image.

[0098] Specifically, the maximum distance feature is the maximum distance from all target pixels in the discriminant slice image to the centroid of the target region.

[0099] The target region is the area in the identification slice image where the pixel value is not zero, and the centroid of the target region is the mean of the pixel coordinates of the non-zero pixel values ​​in the identification slice image.

[0100] The target diameter feature is the length of the diagonal of the smallest bounding rectangle of the target region in the slice image.

[0101] The counting feature is the proportion of pixels with a pixel intensity greater than a second threshold in the target region of the slice image to the total number of pixels in the entire target region.

[0102] Optionally, the second threshold is 98% of the maximum intensity value in the clutter.

[0103] Step 5.6: Based on the discrimination fusion feature values ​​obtained from the maximum distance feature, target diameter feature, and count feature, determine the shape confidence of the discrimination slice image, compare the shape confidence with the third threshold. If the shape confidence is less than the third threshold, it is a target; if the shape confidence is greater than the third threshold, it is clutter. Target detection of the SAR image is thus completed.

[0104] Since ships have certain morphological characteristics, this invention uses them as the basis for screening shape confidence. The method for calculating the fusion identification features designed in this invention calculates the shape confidence by analyzing and statistically analyzing the numerical distribution range of different features.

[0105] In this embodiment, the formula for calculating the identification fusion feature value is:

[0106]

[0107] Where f is the discrimination fusion feature value calculated based on the suspected target slice, distance is the maximum distance feature, diameter is the target diameter feature, and count is the count feature. The distribution of the discrimination fusion feature values ​​is as follows: Figure 4 As shown in the image.

[0108] Based on the distribution of the fused discriminative feature values, the shape confidence of the slice is calculated using the following formula:

[0109]

[0110] Among them, P conf To plot the shape confidence as a function of eigenvalues, see the following graph: Figure 5 As shown.

[0111] After statistically analyzing the distribution of feature values ​​for slice identification, it was found that the maximum distance feature value had the highest confidence level in the range of 70-50, the target diameter feature value had the highest confidence level in the range of 120-170, and the count feature value had the highest confidence level in the range of 0.05-0.45.

[0112] This invention performs a final identification of the remaining suspected target detection boxes by appropriately selecting the discrimination confidence level, and filters out false targets generated in complex ports at the edge of land based on the prior morphological features of the ship, thus completing the ship target detection of wide-span SAR images.

[0113] This invention addresses the issue of wide-swath SAR images containing complex land and near-port scenes. The land portion often exhibits areas of inconsistent brightness, making it prone to generating numerous false targets after constant false alarm detection (CFAR). Additionally, ports located at the land's edge also generate false targets that resemble the actual targets due to the complex near-port scenes. By utilizing land-sea segmentation results and combining them with a shape confidence score based on discriminative features, this invention can effectively distinguish between ship targets and false targets, achieving low CFAR detection for ship targets.

[0114] This invention leverages the advantages of discriminative networks in classification problems, transforming the discrimination between target detection boxes and clutter detection boxes into slice classification. A slice dataset is created to train the network, and a cost-sensitive learning approach is used to address the imbalanced training sample problem, thereby improving the discrimination accuracy compared to traditional methods.

[0115] The effectiveness of this invention can be further illustrated by the following experimental data:

[0116] I. Experimental Conditions

[0117] 1) Experimental data

[0118] The experimental data used to train the discrimination network in the experiment came from the SSDD public dataset. Slices of targets and clutter were manually obtained from the dataset, and labels were classified. Finally, the dataset used to train the discrimination network was made, which contained 3,000 clutter slices and 12,000 target slices. 90% of the experimental dataset was randomly selected as training sample data and 10% as test sample data.

[0119] The wide-swath SAR images used for the detection experiment contain large areas of land and complex near-port scenes, for example... Figure 1 Examples Figure 2 The size and number of ship targets included are given in Table 1. False alarm type 1 represents false targets that are not ships, such as thin bright lines or waves and islands. False alarm type 2 represents false targets caused by errors.

[0120] Table 1. Size and number of targets included in the SAR images used in the experiment.

[0121]

[0122] 2) Evaluation Criteria

[0123] In the ship detection results, TP (TruePositives) indicates that the target was correctly identified; TN (TrueNegatives) indicates that the background in the image was not identified as the target, i.e., no false alarm occurred; FP (FalsePositives) indicates that the background in the image was identified as the target, i.e., a false detection occurred; and FN (FalseNegatives) indicates that the target was incorrectly identified as the background, i.e., a missed detection occurred.

[0124] Precision: Accuracy, representing the proportion of correct predictions out of all predicted targets;

[0125] Recall: Recall rate, which represents the proportion of targets that are correctly predicted out of the total number of targets;

[0126] FA: False Alarm, the number of detection boxes that are incorrectly identified as targets;

[0127] 3) Experiment Content

[0128] For the SAR image data described above, after CFAR detection, the following four identification methods were used to conduct experiments and the detection results are shown in Table 2.

[0129] Experiment 1: Using Support Vector Machines for identification based on slice features;

[0130] Experiment 2: Authentication using only the authentication network;

[0131] Experiment 3: Using a discrimination network combined with land-sea segmentation for discrimination;

[0132] Experiment 4: After identification by the discrimination network, joint identification is performed based on land-sea segmentation and shape confidence.

[0133] Table 2 Comparison of Experimental Results

[0134]

[0135] In Table 2, the method proposed in this invention eliminates false alarms by using a three-step joint identification method after detection using traditional methods. The results show that the number of false alarms in the final detection result of this invention is much less than that of other detection methods. It ensures a high accuracy rate within the allowable decrease range of the detection rate and achieves a good low false alarm detection effect in ship detection of wide-span SAR images.

[0136] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0137] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0138] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for low false alarm target detection in wide-swath SAR images combining shape confidence, characterized in that, The target detection method includes: Step 1: Acquire SAR images; Step 2: Perform constant false alarm rate (CFAR) detection and morphological processing on the SAR image sequentially to obtain the detection image; Step 3: Input the detected image into the trained discrimination network to obtain the first discrimination image. The discrimination network is used to identify false alarms caused by clutter. Step 4: Perform ocean and land segmentation processing on the SAR image, and perform further identification on the first identification image based on the ocean and land segmentation results to remove the land portion and obtain the second identification image; Step 5: Perform final identification on the second discrimination image based on shape confidence to complete target detection of the SAR image; Step 4 includes: Step 4.1: Use latitude and longitude to determine whether the SAR image contains elements that need to be segmented into ocean and land. If so, perform median filtering on the SAR image to obtain a filtered SAR image. Then, execute step 4.2 on the filtered SAR image. Step 4.2: Compare the pixel values ​​of the grayscale histogram of the filtered SAR image with two randomly initialized values, and cluster the pixel values ​​of each pixel in the grayscale histogram with the values ​​that are closer to it into one class to obtain the first clustering result and the second clustering result. Step 4.3: Calculate the pixel mean of the first clustering result and the pixel mean of the second clustering result. Compare the pixel values ​​of the grayscale histogram of the filtered SAR image with the pixel mean of the first clustering result and the pixel mean of the second clustering result. Cluster the pixel value of each pixel in the grayscale histogram with the pixel mean of the nearest pixel to obtain the third clustering result and the fourth clustering result. Step 4.4: Compare the pixel mean of the third clustering result and the pixel mean of the fourth clustering result with the pixel mean of the two clustering results of the previous iteration. If the pixel mean is the same, stop the iteration and obtain the land-sea segmentation binary image. If the pixel mean is different, continue to compare the pixel value of the grayscale histogram of the filtered SAR image with the pixel mean of the third clustering result and the pixel mean of the fourth clustering result until the pixel mean of the two iterations is the same and obtain the land-sea segmentation binary image. Step 4.5: Map the first identification image onto the land-sea segmentation binary image to obtain the mapped image; Step 4.6: Calculate the proportion of ocean and land occupied by the detection box in the corresponding area of ​​the mapped image, respectively; Step 4.7: Determine the relationship between the proportion of the ocean and the proportion of the land and the first threshold. If the proportion of the land is greater than the first threshold, it is a false alarm for the land portion. Then, the area corresponding to the detection box where the land portion is located is removed from the mapping image. If the proportion of the ocean is greater than the first threshold, the area of ​​the mapping image corresponding to the detection box is retained to obtain the second identification image.

2. The method for detecting low false alarm targets in wide-swath SAR images according to claim 1, characterized in that, Step 2 includes: Step 2.1: Perform constant false alarm rate (CFAR) detection on the SAR image using the constant false alarm rate (CFAR) detection method to obtain a binary image; Step 2.2: Perform morphological processing on the binary image to remove connected components that do not meet the preset conditions, and obtain the detection image.

3. The method for detecting low false alarm targets in wide-swath SAR images according to claim 1, characterized in that, The discrimination network comprises 5 convolutional layers, 3 max pooling layers, and 1 dropout layer. The first convolutional layer (conv1), the first max pooling layer (MaxPool1), the second convolutional layer (conv2), the second max pooling layer (MaxPool2), the third convolutional layer (conv3), the third max pooling layer (MaxPool3), the fourth convolutional layer (conv4), the dropout layer, and the fifth convolutional layer are connected in series.

4. The method for detecting low false alarm targets in wide-swath SAR images according to claim 1, characterized in that, Step 5 includes: Step 5.1: Map the second identification image onto the SAR image, extract the region of the SAR image onto which the second identification image is mapped, and obtain a suspected target slice; Step 5.2: Perform a logarithmic transformation on the suspected target slice to obtain the transformed image; Step 5.3: Perform target region segmentation on the transformed image to obtain a segmented image; Step 5.4: Perform morphological processing on the segmented image to obtain the identification slice image; Step 5.5: Extract the maximum distance feature, target diameter feature, and count feature from the identified slice image; Step 5.6: Based on the discrimination fusion feature value obtained from the maximum distance feature, the target diameter feature, and the counting feature, determine the shape confidence of the discrimination slice image, compare the shape confidence with a third threshold. If the shape confidence is less than the third threshold, it is a target; if the shape confidence is greater than the third threshold, it is clutter, thus completing the target detection of the SAR image.

5. The method for detecting low false alarm targets in wide-swath SAR images according to claim 4, characterized in that, Step 5.3 includes: The relationship between each pixel value of the transformed image and a comparison condition is determined. If a pixel value of the transformed image is less than the comparison condition, the pixel is considered clutter, and the clutter is removed. If a pixel value of the transformed image is greater than or equal to the comparison condition, the pixel is considered the target. The segmented image is obtained by comparing all pixel values ​​of the transformed image with the comparison condition. ,in, The average pixel value of the transformed image. It is a constant. denoted as the standard deviation of the transformed image.

6. The method for detecting low false alarm targets in wide-swath SAR images according to claim 5, characterized in that, The maximum distance feature is the maximum distance from all target pixels in the discriminative slice image to the centroid of the target region. The target diameter feature is the diagonal length of the smallest bounding rectangle of the target region in the identification slice image; The counting feature is the proportion of the number of pixels with a pixel intensity greater than a second threshold in the target region of the identification slice image to the total number of pixels in the entire target region.

7. The method for detecting low false alarm targets in wide-swath SAR images according to claim 5, characterized in that, The formula for calculating the discrimination fusion feature value is as follows: in, To identify fusion feature values, For the maximum distance feature, For the target diameter feature, This is a counting characteristic.

8. The method for detecting low false alarm targets in wide-swath SAR images according to claim 5, characterized in that, The formula for calculating the shape confidence level is: in, For shape confidence.

Images