A SAR image target detection method based on fusion of statistical characteristics and structural characteristics

Abstract

This invention discloses a SAR image target detection method based on the fusion of statistical and structural characteristics. It utilizes a CFAR detector and a Faster-RCNN detection network to iteratively identify targets in the SAR image under test until the alternating iteration termination condition is met, thus obtaining the target identification result of the SAR image. This invention fuses the CFAR detector and the Faster-RCNN detection network. Based on the large number of false alarm bounding boxes obtained from the CFAR detector, soft annotations of the bounding boxes are obtained through pixel-level to target-level fusion calculations. The Faster-RCNN detection network is used as the basic training framework, and the bounding boxes output by the Faster-RCNN detection network are used as the protected area size of the CFAR detector, reducing the mixing of target samples with clutter samples and making the parameter estimation of the statistical distribution model more accurate.

Description

Technical Field

[0001] This invention belongs to the field of SAR image target detection technology, and in particular relates to a SAR image target detection method based on the fusion of statistical and structural characteristics. Background Technology

[0002] Reliable marine monitoring capabilities can effectively safeguard a nation's maritime rights and interests, while also facilitating tasks such as maritime rescue, fisheries management, and maritime traffic control. However, when using spaceborne SAR to monitor and detect ship targets in unfamiliar waters, the labeling of these targets lacks the necessary expert knowledge (such as ship GPS information). Due to the wide coverage area of ​​spaceborne SAR and the sparse distribution of target locations, relying solely on manual visual labeling of ship data is time-consuming and labor-intensive. Furthermore, human judgment of targets is easily affected by factors such as sea clutter, making target labeling extremely difficult. Therefore, spaceborne SAR ship target detection often faces unsupervised information situations.

[0003] With the development of artificial intelligence technology, deep learning has been widely used in image object detection. However, fully supervised object detection algorithms based on deep learning (such as Faster-RCNN and YOLO-V3) require a large number of accurate instance annotations. Most current unsupervised object detection methods are based on extracting target location information from the differences between consecutive frames in a video. However, spaceborne SAR typically relies on single-scene image detection, making these methods unsuitable. Deep learning-based detectors offer fast detection speeds and can extract and utilize structural features from images, but they still require a large number of instance-level annotations.

[0004] Traditional SAR-based methods typically do not require annotation information. For example, methods based on SAR image polarization features rely on the fact that ship targets and ocean scattering mechanisms differ, resulting in different polarization characteristics, which can be used to distinguish targets from the background. However, these methods are not suitable for single-polarization data.

[0005] Methods based on the statistical characteristics of SAR play an important role in ship detection. These algorithms are mainly based on the Constant False Alarm Rate (CFAR) detection algorithm and a series of derived methods, such as bilateral CFAR (TP-CFAR) and mean-based CFAR (CA-CFAR). These methods first establish a mathematical statistical model of clutter power, and then use hypothesis testing methods to determine the threshold T for accepting or rejecting the hypothesis under a certain false alarm rate, thus determining whether a pixel is a target.

[0006] However, methods based on constant false alarm detection require traversing pixels using detection units, resulting in slow detection speed and a tendency to generate false alarms. Summary of the Invention

[0007] The purpose of this invention is to provide a SAR image target detection method based on the fusion of statistical and structural characteristics, so as to reduce the false alarm rate of CFAR and improve the target detection accuracy.

[0008] This invention adopts the following technical solution: a SAR image target detection method based on the fusion of statistical and structural characteristics, comprising the following steps:

[0009] The CFAR detector and Faster-RCNN detection network are used to perform target recognition on the SAR image to be identified in an alternating iterative manner until the alternating iteration termination condition is met, so as to obtain the target recognition result of the SAR image to be identified.

[0010] Specifically, the Faster-RCNN detection network is trained using the target recognition results output by the CFAR detector, and the protected area size of the CFAR detector is updated using the target recognition results output by the Faster-RCNN detection network.

[0011] Furthermore, the method for selecting the statistical distribution model of the CFAR detector is as follows:

[0012] Calculate the deviation between the SAR image to be identified and the statistical distribution model;

[0013] The statistical distribution model corresponding to the minimum deviation was selected as the statistical distribution model of the CFAR detector.

[0014] Furthermore, the shape and scale parameters of the statistical distribution model are calculated using the maximum likelihood estimation method.

[0015] Furthermore, training the Faster-RCNN detection network using the target recognition results output by the CFAR detector includes:

[0016] The target recognition results output by the CFAR detector are added as weights to the binary cross-entropy loss function of the Faster-RCNN detection network.

[0017] Furthermore, target identification using the CFAR detector in the SAR image to be identified includes:

[0018] The first target bounding box is obtained by combining connected target pixels in the SAR image to be identified;

[0019] The first target box is enlarged proportionally to obtain the second target box.

[0020] Furthermore, after obtaining the second target box, it also includes:

[0021] The confidence scores of all pixels within the second target bounding box are fused to obtain the confidence score of the second target bounding box.

[0022] Furthermore, the Faster-RCNN detection network is trained using the target recognition results output by the CFAR detector:

[0023] The confidence level of the second target bounding box is used as the target recognition result output by the CFAR detector.

[0024] Furthermore, the loss function of the Faster-RCNN detection network is:

[0025] L = L roiHead +L RPN ,

[0026] Among them, L RPN For RPN loss, L roiHead Head loss detection for Faster-RCNN.

[0027] Furthermore, the RPN loss is:

[0028]

[0029] in, For binary classification, cross-entropy loss, y j For the confidence level of the j-th target, use a standard label. Let N be the confidence prediction value of the RPN for the j-th target, and N be the number of targets. smooth L1 (T,P * ) represents the RPN target box loss, T represents the size of the second target box, and P represents the value of the second target box. * This is the predicted value for the candidate bounding box size in the RPN.

[0030] Furthermore, the loss of the Faster-RCNN detection head is:

[0031]

[0032] in, For the KL divergence of the Faster-RCNN detector head, To detect the head pair classification probability prediction output, smooth L1 (T,T * To detect the head target bounding box loss, T * This is the predicted value for the bounding box size.

[0033] The beneficial effects of this invention are as follows: This invention fuses the CFAR detector and the Faster-RCNN detection network. Based on the large number of false alarm bounding boxes (i.e., pixel-level bounding boxes) obtained by the CFAR detector, soft annotation of the bounding boxes is obtained through pixel-level to target-level fusion calculation. The Faster-RCNN detection network is used as the basic training framework, and the bounding boxes output by the Faster-RCNN detection network are used as the protected area size of the CFAR detector, which reduces the mixing of target samples into clutter samples and makes the parameter estimation of the statistical distribution model more accurate. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the principle framework of a SAR image target detection method based on the fusion of statistical and structural characteristics according to the present invention.

[0035] Figure 2 This is a grayscale histogram of a SAR image in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the signal data of the SAR image in two-dimensional coordinates in an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the detection unit of the CFAR detector in an embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram of the probability density distribution function in an embodiment of the present invention;

[0039] Figure 6 This is a comparison diagram of the CFAR detector detection unit before and after improvement in an embodiment of the present invention;

[0040] Figure 7 These are schematic diagrams illustrating different forms of SAR images to be identified in embodiments of the present invention;

[0041] Figure 8 This is a schematic diagram illustrating the appearance of the soft label in an embodiment of the present invention;

[0042] Figure 9 This is a schematic diagram of the Faster-RCNN structure used in the embodiments of the present invention;

[0043] Figure 10 This is a schematic diagram comparing one-hot labels and soft labels in an embodiment of the present invention;

[0044] Figure 11 This is a schematic diagram of the logic value output of the Faster-RCNN detection head in an embodiment of the present invention. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0046] When there is no prior information about the ship target in the SAR image, unsupervised algorithms based on deep learning typically use contextual information between consecutive frames within the video to locate the target. However, spaceborne SAR detection is usually limited to a single image, so this type of method is not applicable.

[0047] In unsupervised ship target detection methods for single-polarization SAR images, various statistical distribution-based constant virtual detection algorithms are commonly used. These methods mainly utilize the statistical characteristics of SAR images, treating each pixel in the SAR image as a single-point signal. Based on hypothesis testing and local clutter statistical distribution model estimation, a segmentation threshold for test points is calculated under a certain significance level, thereby dividing image pixels into two main categories: foreground and background.

[0048] Because these algorithms perform pixel-level surface detection, they do not fully utilize the structural characteristics between pixels in an image, easily resulting in a large number of false alarms. Furthermore, constant false alarm rate (CFAR) detection algorithms often traverse image pixels using a sliding window approach, leading to low detection efficiency and slow speed. In contrast, convolutional neural network-based detectors offer faster detection speeds and can perform target-level detection tasks. However, training deep learning-based detectors often requires a large number of accurate instance-level labels, making optimization training difficult under conditions of severely limited annotation information.

[0049] To detect ship targets in unlabeled data, this invention proposes an unsupervised SAR ship target detection method based on the fusion of statistical and structural characteristics of SAR images. First, a gamma distribution statistical model is established based on the statistical characteristics of SAR images. Then, gamma distribution parameters are estimated using pixel points to calculate a segmentation threshold, and constant false alarm rate (CFAR) detection is performed on the SAR image to obtain high-false alarm rate pixel detection results. Next, the confidence score of the detection box is obtained using the mean of the differences within the detection box. Furthermore, using the concept of knowledge distillation, a Faster-RCNN is trained with soft labels to amplify the inter-class discrepancy between clutter and the target. Based on the optimization results of the detection network, the target region is set as a protected area to further reduce the number of non-noise samples when estimating the gamma distribution model parameters, gradually optimizing the CFAR detector performance. Finally, extensive experiments on Sentinel-1 data demonstrate that the method of this invention outperforms traditional detectors, improving the detection rate, reducing the false alarm rate, and increasing detection efficiency.

[0050] This invention proposes a ship target detection method based on the fusion of statistical and structural characteristics of SAR images. For example... Figure 1The diagram illustrates the overall framework and steps of the method of this invention. First, by statistically analyzing the pixel frequency data of unlabeled remote sensing images, a statistical distribution model matching the SAR image to be detected is selected. Then, a false alarm rate is set, and the image is detected using a CFAR detector to obtain labels for training Faster-RCNN.

[0051] Similar to the knowledge distillation method, in the method of this invention, the CFAR detector is used instead of the Teacher network, the Faster-RCNN is used instead of the Student network, and training is guided by soft labels provided by the Teacher network.

[0052] Faster-RCNN can filter out clutter and false alarms in CFAR detection results that have significant differences in morphology and structure from the target object. Based on the similarity assessment between targets using Faster-RCNN, to optimize gamma distribution parameter estimation, during maximum likelihood estimation, the training area samples for any detection point are no longer limited to the surrounding neighborhood, but instead find pixels around the target with high similarity globally. The structural characteristics are used to optimize the estimated sample images, increasing the number of estimated sample images and improving the accuracy of the distribution estimation parameters, thus iteratively optimizing the detection results. The key to this invention lies in the fusion of statistical and structural characteristics; therefore, the focus of this invention is on the information interaction between the two detectors.

[0053] Specifically, this invention discloses a SAR image target detection method based on the fusion of statistical and structural characteristics, comprising the following steps: using a CFAR detector and a Faster-RCNN detection network to iteratively identify targets in the SAR image to be identified until the alternating iteration termination condition is met, thereby obtaining the target identification result of the SAR image to be identified; wherein, the target identification result output by the CFAR detector is used to train the Faster-RCNN detection network, and the target identification result output by the Faster-RCNN detection network is used to update the protected area size of the CFAR detector.

[0054] This invention fuses the CFAR detector and the Faster-RCNN detection network. Based on the large number of false alarm bounding boxes (i.e., pixel-level bounding boxes) obtained by the CFAR detector, soft annotation of the bounding boxes is obtained through pixel-level to target-level fusion calculation. The Faster-RCNN detection network is used as the basic training framework, and the bounding boxes output by the Faster-RCNN detection network are used as the protected area size of the CFAR detector to reduce the mixing of target samples in clutter samples and make the parameter estimation of the statistical distribution model more accurate.

[0055] Regarding the termination condition for alternating iterations, it can be designed according to requirements, such as reaching a certain number of iterations or a certain accuracy requirement. In this embodiment of the invention, the termination condition for alternating iterations is that the percentage of detected target pixels is less than the constant false alarm rate for five consecutive iterations. The percentage of detected target pixels is the number of pixels in the target box divided by the total number of pixels in the entire image.

[0056] The following sections first introduce the basis for selecting the statistical distribution model, then describe the specific process from statistical characteristics to structural characteristics, which mainly realizes the fusion of target objects from the pixel level to the instance level. Finally, the basic principles of converting structural characteristics to statistical characteristics are explained in detail, which mainly realizes the feedback of structural characteristics to the statistical model.

[0057] In the process of selecting statistical distribution models, from a statistical perspective, SAR image statistical models can be divided into two main categories according to their model origin: statistical models based on the prior assumption of speckle and empirical distribution models based on experimental data.

[0058] Based on the prior assumption of speckle, statistical models were established, including those showing speckle following a Rayleigh distribution and intensity following a negative exponential distribution. For high-resolution SAR images, in addition to speckle, there is a large amount of texture information; therefore, statistical models such as the K-distribution and Pearson system distribution were also established. Empirical distribution models based on experimental data were obtained through statistical modeling experiments on real SAR image data, mainly including normal distribution, gamma distribution, Weibull distribution, Fisher distribution, etc.

[0059] As can be seen from the above, different types of SAR data conform to different statistical distributions. In order to find a statistical distribution model that matches the SAR image to be identified, it is necessary to use experimental data to fit the parameters of multiple distributions and find the distribution with the highest fit as the statistical distribution model of the SAR image to be identified.

[0060] In this embodiment of the invention, the method for selecting the statistical distribution model of the CFAR detector is as follows: calculate the deviation between the SAR image to be identified and the statistical distribution model; select the statistical distribution model corresponding to the minimum deviation as the statistical distribution model of the CFAR detector.

[0061] like Figure 2 As shown in the figure, the bar chart is the gray-level histogram of the SAR image to be identified in this embodiment of the invention, and the curves are the results of fitting each distribution. After the fitting is completed, the deviation between the sample image and the fitted distribution is calculated, and the calculation formula is as follows:

[0062]

[0063] Where Bias is the deviation between the sample image and the fitted distribution, n = 256, and n is the pixel grayscale value of the SAR image to be identified. x i Let y be the i-th normalized gray value. i For x i The frequency of occurrence, f() is the fitted statistical distribution probability density function.

[0064] For example, by counting the number of pixels with gray values ​​from 0 to 256 in a SAR image, when i = 0, x0 = 0 and y0 = 6, it means that there are 6 pixels with a gray value of 0 in the SAR image.

[0065] As shown in Table 1, the deviation between the data and the distribution shows that the fitted gamma distribution has the smallest deviation from the statistical distribution of the experimental data. Similarly, visualization (such as...) Figure 2 As shown in the figure, the gamma statistical distribution model best matches the data distribution. Therefore, in this embodiment of the invention, the gamma distribution is used as the basic statistical model for the constant virtual detection method.

[0066] Table 1

[0067] distributed Gamma Rayleigh Log-Normal Weibull average deviation 0.1636 0.4947 1.5000 0.4467

[0068] Based on the statistical distribution matching of the data described above, a two-parameter gamma distribution is selected as the mathematical model to support the CFAR detector. Next, soft labels for the SAR image to be identified are generated based on the gamma distribution and the CFAR detector. This process will first explain how to estimate the parameters of the gamma distribution using the SAR image to be identified, then introduce the working principle of the CFAR detector. Next, the deformable CFAR of the training area will be introduced, and finally, the transformation and transfer process from statistical characteristic information to structural information will be explained.

[0069] Specifically, the probability density function of the gamma distribution is shown below:

[0070]

[0071] Where α and λ represent the shape and scale parameters of the gamma distribution, respectively, and Γ(α) is the improper integral defined by an infinite product function containing parameters.

[0072] In this embodiment of the invention, the two unknown parameters in Equation 2, namely the shape parameter and scale parameter of the statistical distribution model, are estimated using the maximum likelihood estimation method (MLE).

[0073] First, its maximum likelihood function is constructed as follows:

[0074]

[0075] To obtain the maximum likelihood estimate, the maximum likelihood function is simplified as follows:

[0076]

[0077] Then, setting the partial derivatives with respect to the estimated parameters to zero, we can obtain:

[0078]

[0079] Solving this system of equations yields:

[0080]

[0081] in, For the optimized shape parameters, The optimized scale parameters, This is the normalized mean gray level.

[0082] Substituting equation 6 into the first expression of the system of equations, we get:

[0083]

[0084] According to this formula, by setting the function ψ(α) to zero, the optimized shape and scale parameters can be obtained by substituting specific numerical values.

[0085] The working principle of the CFAR detector is mainly based on hypothesis testing. In the detection process, such as... Figure 3 As shown in the left figure, if we consider the SAR image pixel x(t) as a set of signal data on a two-dimensional coordinate system, we can make the following two assumptions to determine whether the target signal exists in the collected signal:

[0086]

[0087] The null hypothesis H0 states that the signal x(t) contains both the target signal s(t) and the noise signal n(t), while the alternative hypothesis H1 states that the signal x(t) contains only the noise signal. Under a certain significance level, the acceptance or rejection region is calculated using the sample distribution to determine whether to accept the null hypothesis at that significance level. If accepted, it indicates that the target has been detected; otherwise, it indicates that the signal is background information.

[0088] like Figure 4The diagram shows a schematic of the CFAR detector's detection unit, which is mainly divided into three areas: the test area, the guard band, and the training band. As can be seen from the diagram, the test area, guard band, and training band are arranged sequentially from the inside out. The test area represents the signal at the test point and typically contains only one pixel. The sizes of the guard band and training band can be set according to requirements. The guard band's function is to separate the training and test areas. The constant virtual signal detection mathematical statistical model is established based on clutter. It is generally assumed that the pixel values ​​around the test point may be correlated with the test point. For example, if the test point is a target signal, the surrounding pixels are likely also signals. Therefore, it is necessary to avoid these signal samples from participating in parameter estimation. Thus, when estimating parameters based on clutter statistics, the data from the guard band and test area are removed, and only the data from the training band is used.

[0089] During the detection process, the detection unit traverses each pixel in the form of a sliding window. Each time the window slides, a set of training area, protection area and test data can be obtained. Using the pixels in the training area as samples, the maximum likelihood estimation mentioned above is used to estimate the shape parameter α and scale parameter λ of the corresponding gamma distribution, and obtain the probability density function of the clutter distribution model in the current window.

[0090] Under the significance level, the rejection region is calculated to obtain the segmentation threshold used to distinguish between foreground and background. The null and alternative hypotheses are then selected by comparing the detection points with the segmentation threshold. Throughout this process, the estimation of the statistical distribution depends primarily on the samples included in the training region. However, in traditional constant imaginary detection (CFAR), the sizes of the protection and training regions are fixed, and the parameters are estimated using only the pixel values ​​around the detection points. This setup means that the selected samples may contain target signals, leading to inaccurate estimation of the gamma distribution parameters and thus affecting the size of the foreground / background segmentation threshold.

[0091] To improve the accuracy of statistical distribution model estimation, two aspects can be considered: first, finding a mathematical model that better matches the experimental data, which has already been addressed above; and second, obtaining more representative samples to complete parameter estimation, which is also one of the core aspects. According to Bernoulli's law of large numbers:

[0092]

[0093] As shown in the above equation, the more random samples (n) used for estimating the statistical distribution parameters, the closer the frequency of occurrence of the statistical samples is to the probability of event occurrence, and the more accurate the estimated statistical distribution parameters and the closer the estimated distribution is to the actual distribution. Therefore, in deformable CFAR training area, the samples used to estimate the statistical distribution of local clutter in SAR images are no longer limited to surrounding neighborhood pixels. Samples with statistical distributions similar to those of detection points can also be included in the calculation to increase the sample size and thus make the statistical distribution model closer to the true distribution.

[0094] After obtaining a large number of clutter samples, following the steps described above, the maximum likelihood estimation yields two parameters α and λ of the gamma distribution. After the local clutter model parameters are estimated, substituting the estimated parameters into the statistical distribution model, the probability density function is obtained as follows:

[0095]

[0096] Then, given the constant false alarm rate P of the image fa Given the significance level of the hypothesis test, the splitting threshold (acceptance region or rejection region) is calculated using the following formula:

[0097]

[0098] Among them, P fa is the constant false alarm rate, is the significance level in hypothesis testing, and represents a confidence level. In this embodiment, the value is 0.04, which means that under this hypothesis, the confidence level of the detected signal being the target signal is 96%. F() represents the probability distribution function, and the segmentation threshold corresponding to the significance level can be obtained through the probability distribution function.

[0099] like Figure 5 As shown, the bar chart in the left image represents the pixel frequency statistics of the training area (after normalization), the curve represents the fitted gamma distribution probability density function, and the vertical line with a triangle represents the detection unit segmentation threshold point T. The area enclosed by the probability density function to its right is equal to P. fa The right figure shows the fitted gamma probability distribution function, used to calculate the threshold T.

[0100] Once the segmentation threshold is obtained, the hypothesis can be evaluated. Figure 5 In the null hypothesis H0, when T > Test, the null hypothesis H0 is rejected, indicating that no target was detected; when T ≤ Test, the null hypothesis H0 is accepted, indicating that the target signal was detected.

[0101] To obtain more clutter samples with similar statistical distributions around a certain detection point, SAR image structural characteristics are introduced. By judging the structural similarity between pixel blocks, more samples are provided for optimizing the estimation of statistical distribution parameters.

[0102] like Figure 6The image shows a comparison of the CFAR detector unit before and after improvement. In the CFAR detector, the detection unit traverses each detection object in a sliding window manner. Since no prior knowledge is involved, when the detection unit approaches the edge of the target object, the protection zone of the detection unit cannot protect the target signal, resulting in a large amount of target signal being mixed into the training area. When using the pixels in the training area to estimate the clutter statistical model, the gamma distribution parameter estimation becomes inaccurate. The mixed target signal raises the threshold, causing the target signal at the test point to be missed, thus leading to the target shielding effect. To address this, the network detector can provide a more precise protection zone for the detection point, protecting the target signal. As can be seen from the image, the improved detection unit experiences less interference from clutter samples in the training area, thus resulting in more accurate parameter estimation when estimating the statistical distribution using maximum likelihood estimation.

[0103] CFAR detectors treat the detection of ship targets in SAR images as the detection of discrete signals, with each detection unit focusing on individual pixels, thus ignoring the structural characteristics of the image to some extent. Therefore, CFAR's detection result for SAR images is a binary pixel segmentation map, requiring manual determination of which pixels constitute the entire target. In contrast, detectors trained on deep networks can extract features related to the target image structure. Their target detection results are typically labeled in the form of target bounding boxes, providing a clearer definition of the overall structure of target instances. However, these detectors require instance annotation to guide training; therefore, it is necessary to establish a bridge between statistical detection and structural detection, using statistical detection to provide guidance for structural detection.

[0104] In this embodiment of the invention, the basis for determining whether a pixel is a target pixel is the comparison between the test pixel Test and the threshold T. Therefore, the difference between the two can be used to represent the degree to which it belongs to the target pixel. The larger the value, the greater the probability that it represents the target pixel. For ease of calculation, the difference is normalized to the range of 0 to 1 using the Sigmoid function. The normalization function is shown in the following formula:

[0105]

[0106] In the formula, 'a' represents the normalization parameter, which is used to adjust the intra-class and inter-class differences of pixels during the normalization process.

[0107] like Figure 7The images shown are schematic diagrams of different forms of SAR images for a certain scene. The threshold map T represents the visualization of the threshold corresponding to all pixels in this scene. The pixel confidence map represents the result obtained by subtracting the SAR image from the threshold map and normalizing it using the above formula. From the pixel confidence map, it can be seen that the confidence of the target object's location is relatively high, appearing as a bright area. The segmentation map is the CFAR detection result. From this result, it can be seen that in addition to the target object being detected, a large amount of clutter is detected on both the sea surface and land. The clutter and the target object are similar in mathematical statistical distribution, but their morphological structures differ.

[0108] To obtain bounding boxes suitable for training Faster R-CNN, CFAR needs to be transformed from a binary segmentation image to target boxes. This process involves pixel-level to target-level fusion, and the challenge lies in grouping scattered pixels into the same instance. In this embodiment, the boundaries of connected components within the binary segmentation image are used as target boxes. Target pixels are marked in white, and the background is marked in black. The white blocks formed by multiple adjacent target pixels constitute the connected components. Figure 7 In a segmented image, a white spot can be considered a connected component; the method for determining the connectivity of a binary image is to check if there are any identical values ​​among its eight neighboring pixels.

[0109] In other words, connected target pixels in the SAR image to be identified are combined to obtain a first target bounding box; then, the first target bounding box is enlarged proportionally to obtain a second target bounding box. Next, the confidence scores of all pixels within the second target bounding box are fused to obtain the confidence score of the second target bounding box. Finally, the confidence score of the second target bounding box is used as the target recognition result output by the CFAR detector.

[0110] In the sea, apart from a few ships anchored in harbors, most ships sailing on the surface are relatively far apart, resulting in minimal pixel overlap between different targets in SAR images. Regions defined by connected components can encompass most of the target's pixels while filtering out smaller connected components. However, during CFAR detection, segments of the same target may not belong to the same connected component. Therefore, the size of the bounding rectangles of all connected components is enlarged proportionally to ensure that targets are within the bounding boxes as much as possible. Secondly, the false alarm rate P... fa The size of the area is also related to the size of the connected components. A higher false alarm rate results in a higher target detection rate and a larger connected component area. Therefore, the false alarm rate should not be set too low.

[0111] To represent the differences between objects on the bounding box, it is necessary to calculate the confidence score for each object. Since the CFAR detector can only obtain the confidence score of a single pixel, the process from statistics to structure involves pixel-level to object-level fusion. Therefore, this embodiment fuses the confidence scores of pixels within the bounding box based on the bounding box and pixel confidence scores to obtain the bounding box confidence score.

[0112] like Figure 7 As shown in the bounding box confidence calculation, the overall confidence of the bounding box is calculated based on the size of the bounding box and all pixels within the target box. By calculating the confidence, the statistical difference between the target and the false alarm background objects is obtained. This difference in statistical characteristics between the target and the clutter background provides guidance for structural feature extraction, thereby expanding the inter-class difference between the target and the clutter.

[0113] The pixel confidence score (s(x)) is obtained through a CFAR detector, and then the bounding box confidence score (i.e., the combined confidence scores of all pixels within the bounding box) is obtained. These post-processing steps enable a simple transformation from statistical to structural methods for SAR image ship target detection, achieving a leap from pixel-level to target-level detection. However, detecting the overall target structure solely through pixel connected components is clearly insufficient.

[0114] like Figure 7 As shown in the bounding boxes, although they are detected by the CFAR detector, the contents contained within the bounding boxes vary greatly in terms of image structural characteristics, such as shape. During training, convolutional neural networks can obtain higher-level structural features through convolutional downsampling.

[0115] To this end, the Faster-RCNN object detection model is trained using the detection bounding boxes obtained from statistical properties. The advantages of deep convolutional networks in extracting image structures are utilized to provide optimized feedback for CFAR detection, thereby reducing the false alarm rate of CFAR detection results. At the same time, through iterative training, a deep network detection model with good robustness, low false alarm rate, and high real-time performance is obtained.

[0116] In the knowledge distillation model, two networks are constructed: a teacher network and a student network. The teacher network is typically more complex with a larger number of parameters, while the student model has fewer parameters. During training, the teacher network is trained using labels, and the student network is trained using the teacher network's output. The soft labels obtained from the teacher network's output, compared to standard labels, can reflect intra-class differences to some extent. Therefore, the student network trained using these soft labels also possesses strong generalization performance.

[0117] In this invention, a CFAR detector is used to replace the teacher network, and Faster-RCNN is used as the student model. The bounding boxes obtained by CFAR detection and the confidence of the bounding boxes are used to train Faster-RCNN in the hope of reducing the false alarm rate and improving detection performance by further widening the confidence gap between clutter and target.

[0118] In the Faster-RCNN architecture (which is as follows) Figure 9 As shown), the role of the RPN (Region Proposal Network) is to provide candidate regions for the second stage of training. Figure 8 As shown, in RPN, the feature extraction network undergoes convolution to obtain two types of outputs: one is a logistic value representing the target box confidence, and the other is an output representing the target box position and size, respectively. During the knowledge distillation training process, the bounding box regression part of RPN uses a smooth L1 loss function to calculate the backpropagation parameters. Meanwhile, soft labels are used to guide the training of candidate bounding box confidence.

[0119] In soft labeling, positive sample labels are replaced by the corresponding bounding boxes, and negative sample labels are replaced by 0 (because CFAR has a high false alarm rate, the confidence score of targets in the background is set to 0 to reduce false alarms). In the original FPN, the confidence loss value is obtained by the binary cross-entropy loss function. In this embodiment, the ordinary label is the category, i.e., whether it is a target or not, with 1 representing a target and 0 representing the background. The soft label refers to the confidence score, which is a probability. For example, if two targets are detected in CFAR detection, their ordinary labels are both 1, while their soft labels may be 0.71 and 0.95, respectively. Soft labels can reflect intra-class differences, while ordinary labels cannot.

[0120] During RPN training, the target recognition results output by the CFAR detector are added as weights to the binary cross-entropy loss function of the Faster-RCNN detection network. In other words, soft labels are added to the binary cross-entropy loss function as weights, resulting in the binary cross-entropy loss:

[0121]

[0122] In the formula, y j For the confidence level of the j-th target, use a standard label. Let w be the confidence prediction value of the RPN for the j-th target, i.e., the confidence prediction value of the RPN, where N is the number of targets in the SAR image. j The weight corresponding to each label is ε, where ε is an infinitesimal quantity and j represents the index of the target object.

[0123]

[0124] In the above formula, smooth L1 (T,P * ) represents the RPN bounding box loss, T represents the size of the second bounding box, and P represents the second bounding box size. * Given the predicted size of the candidate bounding box in the RPN, the RPN loss can be expressed as:

[0125]

[0126] like Figure 11 As shown, the logistic output of the Faster-RCNN detector head differs from that of the RPN. After ROI pooling, each recommended region is unified into a feature map of the same size, and then fully connected layers are used for classification. The final output is processed by a Softmax function to obtain the probability value corresponding to each category. When the detector head is trained using knowledge distillation, as shown... Figure 10 As shown, the bounding box confidence obtained from the CFAR detection part is encoded into soft labels corresponding to the one-hot vectors, replacing the one-hot labels to guide training.

[0127] In terms of the loss function, KL divergence is used instead of the multi-class cross-entropy loss function, i.e., the KL divergence of the Faster-RCNN detector head. The loss is calculated as follows:

[0128]

[0129] in, To detect the head pair classification probability prediction output, Y i Let the classification probability of the i-th target object be a soft label. Let be the predicted label for the i-th target.

[0130] The diagram only shows the training structure of the classification part. For the bounding box regression part, the smooth L1 loss function is used for calculation, just like RPN. The following formula represents the loss of the Faster-RCNN detection head:

[0131]

[0132] Among them, smooth L1 (T,T * ) represents the loss between the target box and the predicted box.

[0133] The total network loss function is the sum of the RPN loss and the Faster-RCNN detector head loss:

[0134] L = L roiHead +L RPN (18)

[0135] Training the Faster R-CNN neural network requires labels, while CFAR detection does not require labels but has a high false alarm rate. CFAR detection results can provide soft labels for Faster R-CNN. Furthermore, the iterative training process of Faster R-CNN can amplify intra-class discrepancies and reduce false alarms.

[0136] Furthermore, the method of this invention is compared with other unsupervised detection methods to illustrate its superiority in unsupervised SAR ship target detection tasks. Most unsupervised ship target detection methods are based on polarization or statistical characteristics. This invention uses single-polarization SAR images, and the constant false alarm rate (CFAR) detection module employs CFAR detection under the Gamma statistical distribution. The comparison methods include CA-CFAR, GOCA-CFAR, and SOCA-CFAR, in addition to CFAR-Gamma. A constant false alarm rate P is set. fa =0.04.

[0137] Table 2

[0138]

[0139] Table 3

[0140]

[0141]

[0142] Table 4

[0143]

[0144] The comparison results under the same conditions are shown in Tables 2, 3, and 4. It can be seen that the proposed method (USOD) exhibits good performance in Sentinel-1 SAR ship target detection data. Regarding the number of correctly detected targets, the higher the IOU value, the better the proposed method is compared to statistical detection methods. In terms of the number of false alarms, the proposed method has the fewest false alarms. Furthermore, in terms of the overall evaluation index, the Quality Factor (FoM), the proposed method also achieves the highest score.

[0145] Compared to statistical detection methods, USOD offers a significant performance improvement, greatly reducing the number of false alarms. Furthermore, in terms of detector speed, USOD relies on convolutional networks for forward propagation, while statistical methods use sliding windows, making USOD far faster than statistical detection.

[0146] This invention addresses the scenario where the data source contains only unlabeled samples, proposing an unsupervised ship target detection method based on the fusion of SAR image statistics and structural characteristics. The innovations are mainly twofold: First, it utilizes a constant false alarm rate (CFAR) method to obtain soft annotations for training a deep detection network. This method is driven by statistical distribution, obtaining a large number of false alarm bounding boxes through CFAR detection. Pixel-level to target-level fusion calculations yield the soft annotations for these bounding boxes. A knowledge distillation training model is constructed based on the Faster-RCNN detector as the fundamental training framework. Second, during the iterative process, an adaptive detection unit is proposed, replacing the protected area of ​​the detection unit in CFAR detection with a detection box, optimizing the estimation of statistical distribution parameters, thereby achieving mutual optimization between the two detectors.

[0147] To address the challenge of training unsupervised detectors without any supervised information, and considering that traditional detection methods based on statistical or polarization characteristics are mostly pixel-level and lack a comprehensive description of the overall structure of the target object, this invention combines the advantages of traditional unsupervised methods with those of current deep learning models. It utilizes the statistical characteristics of SAR images to obtain high false alarm results through CFAR statistical detection. Then, a knowledge distillation model is used to widen the gap between false alarm clutter and the target object. Based on the bounding boxes predicted by the network, adaptive detection units are provided for the CFAR detector, reducing the mixing of target samples with clutter samples and making the parameter estimation of the statistical distribution model more accurate. Iterative iteration promotes mutual optimization between the two detectors. This method can fully exploit data information under unlabeled conditions, possessing strong theoretical research significance and practical value.

[0148] Compared to fully supervised detectors, the method proposed in this invention does not require annotation information, reducing data annotation costs and saving manpower and resources for data annotation. Compared to traditional unsupervised ship inspection methods, the method proposed in this invention has higher detection performance and higher detection accuracy, while also significantly improving detection efficiency.

[0149] The present invention also discloses a SAR image target detection device based on the fusion of statistical and structural characteristics, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned SAR image target detection method based on the fusion of statistical and structural characteristics.

[0150] It should be noted that the information interaction and execution process between the above-mentioned devices are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0151] The device can be a computing device such as a desktop computer, laptop, handheld computer, radar, or cloud server. The device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that it may include more or fewer components, or a combination of certain components, or different components; for example, it may also include input / output devices, network access devices, etc.

[0152] The processor referred to can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0153] In some embodiments, the memory may be an internal storage unit of the extraction device, such as the hard drive or memory of the extraction device. In other embodiments, the memory may be an external storage device of the extraction device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the extraction device. Furthermore, the memory may include both internal storage units and external storage devices of the extraction device. The memory is used to store operating systems, applications, bootloaders, data, and other programs, such as the program code of the computer program. The memory can also be used to temporarily store data that has been output or will be output.

Claims

1. A SAR image target detection method based on the fusion of statistical and structural characteristics, characterized in that, Includes the following steps: The CFAR detector and Faster-RCNN detection network are used to perform target recognition on the SAR image to be identified in an alternating iterative manner until the alternating iteration termination condition is met, thereby obtaining the target recognition result of the SAR image to be identified. Specifically, the Faster-RCNN detection network is trained using the target recognition results output by the CFAR detector, and the bounding box output by the Faster-RCNN detection network is used as the protected area size of the CFAR detector. Training the Faster-RCNN detection network using the target recognition results output by the CFAR detector includes: The target recognition results output by the CFAR detector are added as weights to the binary classification cross-entropy loss function of the Faster-RCNN detection network; Target identification using a CFAR detector in a SAR image includes: The first target bounding box is obtained by combining connected target pixels in the SAR image to be identified; The first target box is enlarged proportionally to obtain the second target box; After obtaining the second target bounding box, it also includes: The confidence scores of all pixels within the second target bounding box are fused to obtain the confidence score of the second target bounding box. The Faster-RCNN detection network is trained using the target recognition results output by the CFAR detector: The confidence level of the second target bounding box is used as the target recognition result output by the CFAR detector.

2. The SAR image target detection method based on the fusion of statistical and structural characteristics as described in claim 1, characterized in that, The method for selecting the statistical distribution model of the CFAR detector is as follows: Calculate the deviation between the SAR image to be identified and the statistical distribution model; The statistical distribution model corresponding to the minimum deviation is selected as the statistical distribution model of the CFAR detector.

3. The SAR image target detection method based on the fusion of statistical and structural characteristics as described in claim 2, characterized in that, The shape and scale parameters of the statistical distribution model are calculated using the maximum likelihood estimation method.

4. The SAR image target detection method based on the fusion of statistical and structural characteristics as described in claim 3, characterized in that, The loss function of the Faster-RCNN detection network is: ， in, For RPN loss, Head loss detection for Faster-RCNN.

5. The SAR image target detection method based on the fusion of statistical and structural characteristics as described in claim 4, characterized in that, The RPN loss is: ， in, For binary classification, cross-entropy loss, , , It is an infinitesimal quantity. For the confidence level of the j-th target, use a standard label. Let N be the confidence prediction value of the RPN for the j-th target, and N be the number of targets. The weight corresponding to each label. , For RPN target box loss, The size of the second target box. This is the predicted value for the candidate bounding box size in the RPN.

6. The SAR image target detection method based on the fusion of statistical and structural characteristics as described in claim 5, characterized in that, The loss of the Faster-RCNN detector head is: ， in, For the KL divergence of the Faster-RCNN detector head, The output is the classification probability prediction for the detection head pair. To detect head bounding box loss, This is the predicted value for the bounding box size.

Images