SAR distance compressed domain ship target detection method based on anchor line dynamic generation

Abstract

The invention relates to the technical field of target detection, in particular to an SAR (Synthetic Aperture Radar) distance compressed domain ship target detection method based on anchor line dynamic generation. Comprising the following steps: step 1, acquiring spaceborne SAR echo data subjected to BAQ decompression processing; step 2, based on the spaceborne SAR echo data, executing matched filtering processing in a distance direction so as to obtain a corresponding distance compressed domain preliminary image; 3, in the distance compressed domain image, a 1024 * 1024 sliding window is used for window sliding; step 4, extracting four scale features of a target in the image by using Resnet101 as a backbone network; step 5, carrying out fusion processing on the feature maps # imgabs0 #, # imgabs1 # and # imgabs2 # to obtain multi-scale features # imgabs3 #, # imgabs4 # and # imgabs5 #, and carrying out fusion processing on the feature maps # imgabs0 #, # imgabs1 # and # imgabs2 #; step 6, embedding a large kernel attention between the feature maps # imgabs6 # and # imgabs7 # of the feature pyramid; 7, dynamically generating an anchor line; and step 8, accurately positioning the target based on the anchor line. And the detection performance is improved by using the amplitude gradient and geometric scale characteristics of the ship target in the distance compression domain.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and specifically relates to a ship target detection method in the SAR range compression domain based on dynamic generation of anchor lines. Background Art

[0002] Synthetic Aperture Radar (SAR) has the characteristics of all-weather, all-day operation and high-resolution imaging, and shows unique advantages in the field of ship positioning. At present, the ship target detection in SAR images generally adopts a two-stage processing mode of "full-scene imaging - image analysis": first, full-scene radar imaging is performed on the original echo signal to generate a focused image similar to the geometric characteristics of an optical image; then, traditional computer vision or deep learning algorithms are used to achieve the detection and positioning of ship targets.

[0003] However, this paradigm has obvious shortcomings in actual wide-area monitoring scenarios. On the one hand, the high complexity of the full-scene imaging algorithm leads to low processing efficiency and it is difficult to meet the real-time requirements; on the other hand, moving targets will cause defocus artifacts, and at the same time, the sidelobe interference effect significantly reduces the image signal-to-noise ratio, affecting the detection accuracy and reliability. Summary of the Invention

[0004] The present invention provides a ship target detection method in the SAR range compression domain based on dynamic generation of anchor lines, aiming to solve the problem of low efficiency caused by the processing of invalid data in the traditional "full-scene imaging - image analysis" paradigm through the cooperation mechanism of target specific area detection and local area fast imaging in the range compression domain, so as to significantly enhance the timeliness of the maritime ship target situation awareness.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] The present invention provides a ship target detection method in the SAR range compression domain based on dynamic generation of anchor lines, including the following steps:

[0007] Step 1, obtaining spaceborne SAR echo data that has undergone BAQ decompression processing;

[0008] Step 2, based on the above spaceborne SAR echo data, performing matched filtering processing in the range direction to obtain a corresponding preliminary imaging in the range compression domain;

[0009] Step 3, in the range compression domain image, performing sliding windowing using a sliding window of 1024 × 1024;

[0010] Step 4, using Resnet101 as the backbone network to extract 4 scale features of the target in the image, namely the features downsampled by 8, 16, and 32 times 、 , ;

[0011] Step 5, according to the generation rule of the feature pyramid, fuse the feature maps , , to obtain multi-scale features , , ;

[0012] Step 6, embed large kernel attention between the feature maps and of the feature pyramid to expand the receptive field of the detection model and effectively capture the global context information of slender and continuous ship curves;

[0013] Step 7, dynamic anchor line generation: Use the obtained multi-scale features , , , and simultaneously calculate the starting coordinates of the anchor line, the offset relative to the starting point, and the slope parameter through three independent prediction branches; Based on these calculation results, for each identified ship target instance, calculate and output a dynamic anchor line generation that can accurately locate and characterize the direction characteristics of this instance in real time;

[0014] Step 8, accurate target localization based on the anchor line: Use the anchor line generated in Step 7 as the spatial reference benchmark to predict the horizontal offset of the ship target contour in the range compression domain relative to this anchor line; Based on the predicted horizontal offset, achieve accurate calculation and positioning of the ship target position in the range compression domain.

[0015] Furthermore, the application of the large kernel attention in Step 6 is as follows:

[0016] The large kernel attention module is placed after the side layer of the FPN; In addition, a multi-scale aggregator (MSA) is used to quantify the correlation between input tokens, and its formula can be expressed as:

[0017] (1.1)

[0018] (1.2)

[0019] (1.3)

[0020] Among them, the four feed-forward paths are marked as , and are distinguished by different colors, where corresponds to the same forward path.

[0021] Furthermore, the dynamic generation of anchor lines described in step 7 specifically includes the following steps:

[0022] Step 7.1, starting point position probability heat map generation: build a key point detection branch and process the feature map through a two-level convolution layer , output resolution Probability heatmap The heat map represents the probability distribution of each pixel being the starting point of the anchor line;

[0023] During the training phase, based on the starting coordinates of the ship target curve Generate true value heat map; calculate downsampling coordinates ,in Corresponding feature map step size; then, Construct Gaussian distribution area for the center, true value heat map In coordinates The label value of is represented by the following function:

[0024] (1.4)

[0025] in, is the spatial index of the heatmap; standard deviation Control the width of the probability distribution;

[0026] In addition, in order to alleviate the sample imbalance problem between the starting area and the non-starting area, an improved focal loss function is used to process the heat map prediction; and Represents the position in the heat map The predicted probability value and the true value label at , the loss function is defined as:

[0027] (1.5)

[0028] in, and is an adjustable hyperparameter; Indicates the number of ship curves in the distance compression domain; Used to reduce penalties around true location;

[0029] Step 7.2, starting point offset estimation: In the starting point precise positioning stage, the initial position estimation comes from the peak detection of the probability distribution map, and its accuracy is affected by the feature map downsampling ratio. To eliminate positioning errors, this application introduces a dynamic offset correction mechanism, which uses a two-dimensional offset prediction map output in parallel. Achieve sub-pixel positioning;

[0030] During the model training process, a spatial constraint optimization strategy is adopted, which is to define the side length as a square effective area, within which a smooth loss function is applied for offset supervision. This constraint mechanism significantly improves the focusing of offset learning, and its mathematical definition is:

[0031] (1.6)

[0032] where and respectively represent the offsets of the effective position relative to in the axis and axis directions;

[0033] Step 7.3, Adaptive slope estimation mechanism: In the adaptive optimization link of the anchor line slope, this application adds a slope parameter prediction branch, which outputs a slope feature map with a dimension of This slope feature map directly determines the spatial inclination characteristics of the dynamic anchor line; through the slope self-adaptation mechanism, the generated anchor line can adaptively fit the morphological contour of the target ship; in addition, for the ship target curve , where represents the termination point index, and the key point geometric relationship modeling technology is adopted: taking the starting point as the reference point, calculating the average slope of the connection lines formed by the remaining effective points in the curve and the starting point as the true value label, and its mathematical representation is: (1.7)

[0034] (1.7)

[0035] In the training stage, the same effective training area as the offset estimation is selected, and loss is applied to constrain the slope prediction, and its definition is as follows:

[0036] (1.8)

[0037] where is the index of the ship target curve sequence;

[0038] Step 7.4, Anchor line decoding and spatial index optimization mechanism: This application innovatively adopts a two-stage peak confirmation strategy. First, a maximum pooling operation is performed on the starting point probability distribution map to generate a primary response map, and then a secondary verification is achieved through a logical AND operation to accurately extract the coordinates of each response peak in the distribution map ; The coordinate refinement formula is adopted in the anchor line generation process: , where the offset is taken from the predicted value at the position in the offset map , and the slope is obtained from the corresponding spatial position of the slope map .

[0039] Furthermore, the detection head described in step 8 specifically includes the following steps:

[0040] Step 8.1, Anchor line feature modeling stage: Based on the dynamic anchor line parameterized coordinates , uniformly sample spatial key points along the anchor line trajectory; extract the deep features of each sampling point in the feature map through the bilinear interpolation algorithm, and temporally splice all feature vectors to construct an RoI feature tensor representing the spatial distribution characteristics of the anchor line ;

[0041] Step 8.2, Target location parameter decoding method: Input the anchor line RoI feature into the fully connected regression network to synchronously generate three groups of location parameters: horizontal offset vector , starting index and extension length ; construct a vertical coordinate sequence based on the image height , and calculate the target horizontal coordinate through the anchor line calibration formula . This decoding mechanism realizes the joint modeling of the target position, scale and shape through a single forward propagation:

[0042] (1.9)

[0043] Based on the above analysis, the overall loss can be defined as:

[0044] (1.10)

[0045] where is the Smooth loss, which is used to supervise the prediction of the starting index and the curve length . is calculated based on the predicted curve and the true curve trajectory, and is used to constrain the offset regression ; , , , and are the weight coefficients of the corresponding losses respectively.

[0046] The beneficial effects achieved by the present invention are:

[0047] (1) Detection performance improvement: Utilize the amplitude gradient and geometric scale characteristics of ship targets in the range compression domain to improve detection performance. Use continuous and slender anchor lines to locate ship targets in the range compression domain, and force the feature learning network to capture the global context information of ships to improve the accuracy of detection.

[0048] (2) Processing efficiency optimization: Compared with the traditional SAR processing that first images and then detects, the method proposed in the present invention only completes the location of the suspected ship target area in the range compression domain of the SAR echo, eliminates most of the background areas, greatly reduces the data volume processing for subsequent imaging steps, meets the low-latency and fast extraction of information of interest such as SAR satellite echo to target slices, and lays a foundation for the on-orbit imaging of future SAR satellites, as well as the improvement of the intelligent level and on-orbit application efficiency. Brief Description of the Drawings

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0050] Figure 1 The following shows the schematic diagram of the target detection grid structure of the method of the present invention.

[0051] Figure 2 The following shows the structure diagram of the large kernel attention module of the method of the present invention; in the figure, (a) is the structure diagram of the attention mechanism module, and (b) is the structure diagram of the multi-scale aggregator (MSA) module. Detailed Embodiments

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0053] It should be noted that if there are directional indications (such as up, down, left, right, front, back...) involved in the embodiments of the present invention, the directional indications are only used to explain the relative position relationship and movement conditions between components in a specific posture (as shown in the drawings). If the specific posture changes, the directional indications will also change accordingly.

[0054] In addition, if the embodiments of the present invention involve descriptions such as "first" and "second", the descriptions of "first", "second", etc. are only for descriptive purposes and should not be construed as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In addition, the meaning of "and / or" appearing throughout the text is that it includes three parallel solutions. Taking "A and / or B" as an example, it includes solution A, or solution B, or the solution where A and B are satisfied simultaneously. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0055] Existing SAR image target detection frameworks usually adopt a two-stage processing mode of "full-scene imaging - image analysis", but this paradigm has significant limitations in actual wide-area monitoring applications: the algorithm complexity of full-scene imaging leads to insufficient processing timeliness; the defocus artifacts and sidelobe interference effects generated by moving targets will seriously reduce the image signal-to-noise ratio.

[0056] To address the bottlenecks of the existing technology, the present invention proposes a fast detection optimization strategy based on echo-domain feature analysis: first, a rough localization of potential target areas is achieved through fast echo signal feature extraction technology, and then dynamic adaptive imaging processing is implemented for key areas. This hierarchical processing mechanism can effectively avoid the computational redundancy of full-scene imaging, and through the collaborative optimization of local refinement processing and global fast scanning, significantly improve the overall processing efficiency of the SAR system from data acquisition to target recognition.

[0057] Specifically as Figure 1 shown in Fig. 2, the present invention provides a method for detecting ship targets in the SAR range compression domain based on dynamic generation of anchor lines, specifically including the following steps:

[0058] Step 1, obtain spaceborne SAR echo data that has undergone BAQ decompression processing;

[0059] Step 2, based on the above spaceborne SAR echo data, perform matched filtering processing in the range direction to obtain the corresponding preliminary imaging in the range compression domain;

[0060] Step 3, in the range compression domain image, perform sliding windowing using a 1024 × 1024 sliding window;

[0061] Step 4, use Resnet101 as the backbone network to extract 4 scale features of the targets in the image, namely the features downsampled by 8, 16, and 32 times 、 , ;

[0062] Step 5, as Figure 1 shown, according to the generation rule of the feature pyramid, fuse the feature maps , , to obtain multi-scale features , , ;

[0063] Step 6, as Figure 1 shown, embed large kernel attention between the feature maps and of the feature pyramid to expand the receptive field of the detection model and effectively capture the global context information of the slender and continuous ship curves;

[0064] Step 7, anchor line dynamic generation: Use the obtained multi-scale features , , , and simultaneously calculate the starting coordinates of the anchor line, the offset relative to the starting point, and the slope parameter through three independent prediction branches. Based on these calculation results, for each identified ship target instance, calculate and output an anchor line dynamic generation that can accurately locate and characterize the direction characteristics of this instance in real time;

[0065] Step 8, accurate target positioning based on the anchor line: Use the anchor line generated in Step 7 as the spatial reference benchmark to predict the horizontal offset of the ship target contour in the range compression domain relative to this anchor line. Based on the predicted horizontal offset, achieve accurate calculation and positioning of the ship target position in the range compression domain.

[0066] As a further preferred solution of the SAR ship target detection method based on anchor line dynamic generation of the present invention, the application of the large kernel attention in the said Step 6 is as follows:

[0067] As Figure 2 shown, the large kernel attention (LKA) module is placed after the side layer of the FPN to minimize the computational cost. Different from generating the similarity score between the query and the value output, the present invention uses a multi-scale aggregator (MSA) to quantify the correlation between the input tokens, and its formula can be expressed as:

[0068] (1.1)

[0069] (1.2)

[0070] (1.3)

[0071] Among them, represents the attention feature, which is calculated by the multi-scale aggregator (MSA) and is used to quantify the correlation between input tokens, highlight important features and suppress unimportant features, thereby improving the model's attention to key information. is a learnable weight matrix used to perform a linear transformation on the features after multi-channel depth convolution, adjust the scale and bias of the features, and enable the model to learn more discriminative feature representations. is another learnable weight matrix that performs a linear transformation on the input feature map and maps it to a suitable feature space for element-wise multiplication operation with the attention feature to achieve feature modulation. represents a multi-channel operation. The subscript i ranges from 0 to 3, meaning that different channels are processed separately to obtain multi-channel feature information and enhance the model's ability to capture multi-scale features. represents a 5×5 depth convolution operation used to perform convolution processing on the input feature map to extract features of different channels and capture spatially relevant information. represents the input feature map, which is the input data of the entire LKA module and contains feature information of the image, such as the number of channels, spatial dimensions, etc. is the feature map modulated by the attention mechanism, which fuses the attention information and the result of the linear transformation of the input feature map, and at the same time retains the original information of the input feature map (achieved through skip connections), which helps the model better learn rich feature representations. is the final output feature map, which fuses the features processed by the feed-forward network and 's features, providing a more advanced and discriminative feature representation for the subsequent lane detection task. FFN represents the feed-forward network, which is usually composed of two linear transformations and a non-linear activation function and is used to further process features and increase the model's expressive power and non-linear fitting ability.

[0072] Four feed-forward paths are labeled as and are distinguished by different colors, where corresponds to the same forward path. Compared with using 31×31 and 7×7 depth convolutions, the strip-shaped convolution more effectively reduces the computational cost while identifying ship curves. The linear layer is represented by Use the Hadamard product (denoted by ⊙ in Equation 1.2) instead of the matrix product to utilize the advantages of the large kernel in the MSA.

[0073] As a further preferred solution of the ship target detection method in the SAR range compression domain based on anchor line dynamic generation of the present invention, the specific steps of the anchor line dynamic generation in step 7 are as follows:

[0074] Step 7.1, generation of starting point position probability heatmap: Construct a key point detection branch, and process the feature map through two-level convolutional layers , and output a probability heatmap with a resolution of . The heatmap represents the probability distribution of each pixel point as the starting point of the anchor line. .

[0075] In the training stage, based on the starting point coordinates of the ship target curve , generate a ground truth heatmap; where is the abscissa of the starting point of the ship target curve, and is the ordinate of the starting point of the ship target curve. Calculate the downsampled coordinates , where corresponds to the feature map stride. Then, construct a Gaussian distribution region centered on , and the label value of the ground truth heatmap at the coordinate is represented by the following function:

[0076] (1.4)

[0077] where is the spatial index of the heatmap; is the original width of the input image; is the original height of the input image; the standard deviation controls the width of the probability distribution, and the standard deviation is related to the size of the input image; is the operation of taking the maximum exponential function value, that is, among all the Gaussian distributions corresponding to the starting points of the ship target curves, take the maximum value at the coordinate ; is the abscissa of the starting point of the downsampled ship target curve; is the ordinate of the starting point of the downsampled ship target curve.

[0078] To alleviate the sample imbalance problem between the starting point region and the non-starting point region, an improved focal loss function is used to process the heatmap prediction. Let and respectively represent the predicted probability value and the ground truth label at the position in the heatmap, and define the loss function as:

[0079] (1.5)

[0080] where and are adjustable hyperparameters. Their values are usually related to the ratio of positive and negative samples in the dataset, the difficulty of the task, and the requirements for prediction accuracy, and need to be determined according to the actual situation. This article does not make any restrictions; represents the number of ship curves in the distance compression domain; the term is used to reduce the penalty around the true position; is the height of the ground truth heatmap, that is, the dimension of the downsampled heatmap in the vertical direction; is the width of the ground truth heatmap, that is, the dimension of the downsampled heatmap in the horizontal direction.

[0081] Step 7.2, starting point offset estimation: In the stage of accurate starting point positioning, the initial position estimation comes from the peak detection of the probability distribution map, and its accuracy is restricted by the feature map downsampling ratio . To eliminate the positioning error, this patent introduces a dynamic offset correction mechanism to achieve sub-pixel positioning through the two-dimensional offset prediction map output in parallel. This offset map has a dual correction function: First, it compensates for the coordinate quantization error caused by the downsampling process (that is, the mapping deviation between the original coordinates and the downsampled coordinates ), and second, it corrects the initial position offset existing in the probability distribution prediction in real time.

[0082] During the model training process, a spatial constraint optimization strategy is adopted - only a square valid area with a side length of is delimited with the true starting point coordinates as the center, and a smooth loss function is applied within this limited area for offset supervision. This constraint mechanism significantly improves the focusing of offset learning, and its mathematical definition is:

[0083] (1.6)

[0084] where and respectively represent the offsets of the valid position relative to in the axis and axis directions; is the offset predicted by the model.

[0085] Step 7.3, adaptive slope estimation mechanism: In the link of adaptive optimization of the anchor line slope, this scheme adds a slope parameter prediction branch to output a slope feature map with a dimension of , and this slope feature map directly determines the spatial inclination characteristics of the dynamic anchor line. Through the slope self-adaptation mechanism, the generated anchor line can adaptively fit the contour of the target ship, significantly improving the fitting accuracy and recognition efficiency of subsequent target detection. For the ship target curve (where represents the end point index), the key point geometric relationship modeling technology is adopted: taking the starting point as the reference point, the average slope of the lines formed by the remaining valid points in the curve and the starting point is calculated as the true value label, and its mathematical representation is:

[0086] (1.7)

[0087] where is the abscissa of the starting point of the ship target curve; is the ordinate of the starting point of the ship target curve; is the abscissa of the th valid point on the ship target curve; is the ordinate of the th valid point on the ship target curve; is the index of the starting point of the ship target curve; is the number of valid points of the ship target curve.

[0088] In the training stage, the same effective training area as the offset estimation is selected, and loss is applied to constrain the slope prediction, and its definition is as follows:

[0089] (1.8)

[0090] where is the index of the ship target curve sequence; is the slope feature map predicted by the model at the position ; is the true value label of the average slope corresponding to the k-th curve of the ship target curve sequence.

[0091] Step 7.4, Anchor Line Decoding and Spatial Index Optimization Mechanism: This solution innovatively adopts a two-stage peak confirmation strategy: First, perform a max pooling operation on the starting point probability distribution map to generate a primary response map, and then perform a secondary verification through a logical AND operation to accurately extract the coordinates of each response peak in the distribution map. This collaborative processing mechanism shows significant advantages in eliminating redundant predictions - it is more efficient than traditional non-maximum suppression methods based on distance or IoU. The obtained peak coordinates have dual functions: they serve as both the initial estimate of the starting point position and the spatial index to associate the offset with the slope prediction branch. The anchor line generation process uses the coordinate refinement formula: , where the offset is taken from the predicted value at the position in the offset map , and the slope is taken from the slope map Obtaining the corresponding spatial position (note: the index position here should be unified as to ensure spatial consistency).

[0092] As a further preferred solution of the ship target detection method in the SAR range compression domain based on dynamic anchor line generation of the present invention, the detection head described in step 8 specifically includes the following steps:

[0093] Step 8.1, Anchor line feature modeling stage: Based on the dynamic anchor line parameterized coordinates , uniformly sample spatial key points along the anchor line trajectory. Extract the deep features of each sampling point in the feature map through the bilinear interpolation algorithm, and temporally splice all feature vectors to construct a RoI feature tensor representing the spatial distribution characteristics of the anchor line.

[0094] Step 8.2, Target location parameter decoding method: Input the anchor line RoI feature into the fully connected regression network to synchronously generate three groups of location parameters: horizontal offset vector (representing the spatial deviation between the true contour of the ship target and the anchor line), starting index and extension length ; construct a vertical coordinate sequence based on the image height , and calculate the target horizontal coordinate through the anchor line calibration formula (where realizes the geometric conversion from slope to horizontal displacement). This decoding mechanism realizes the joint modeling of target position, scale and shape through a single forward propagation:

[0095] (1.9)

[0096] Based on the above analysis, the overall loss can be defined as:

[0097] (1.10)

[0098] Among them, is the Smooth loss, which is used to supervise the prediction of the starting index and the curve length . is calculated based on the predicted curve and the true curve trajectory, and is used to constrain the offset regression . , , , and They are the weight coefficients corresponding to the respective losses. These coefficients are usually obtained through empirical values or experimental adjustments to achieve a balance between different loss terms. The specific values are related to factors such as loss balance, task objectives, and dataset characteristics during the training process.

[0099] Through the above steps, the suspected target area can be quickly located in the wide - area sea surface, effectively reducing the time and computational resource consumption required for subsequent focused imaging. Moreover, the present invention can learn the strong discriminative features between ship targets and sea clutter based on the amplitude characteristics and geometric features of ship echo energy in range - direction focusing and azimuth - direction defocusing, and quickly locate the suspected target area.

[0100] The above - mentioned are only optional embodiments of the present invention, and thus do not limit the patent scope of the present invention. Any equivalent structural transformation made by using the content of the specification and drawings of the present invention under the inventive concept of the present invention, or direct / indirect application in other related technical fields, is included in the patent protection scope of the present invention.

Claims

1. A method for detecting ship targets in the SAR range compression domain based on dynamic generation of anchor lines, characterized in that It includes the following steps: Step 1: Obtain the spaceborne SAR echo data after BAQ decompression processing; Step 2: Based on the above spaceborne SAR echo data, perform matched filtering processing in the range direction to obtain the corresponding preliminary imaging in the range compression domain; Step 3: In the range compression domain image, perform sliding windowing using a 1024×1024 sliding window; Step 4: Use Resnet101 as the backbone network to extract four-scale features of the target in the image, namely the features downsampled by 8, 16, and 32 times , , ; Step 5, according to the generation rules of the feature pyramid, for the feature maps , , perform fusion processing to obtain multi-scale features , , ; Step 6, embed large kernel attention between the feature maps of the feature pyramid and to expand the receptive field of the detection model and effectively capture the global context information of slender and continuous ship curves; Step 7, Anchor Line Dynamic Generation: Using the obtained multi-scale features , , , simultaneously calculate the starting coordinates of the anchor line, the offset relative to the starting point, and the slope parameter through three independent prediction branches; based on these calculation results, for each identified ship target instance, calculate and output in real time an anchor line dynamic generation that can accurately locate and characterize the direction characteristics of this instance; Step 8: Precise target positioning based on the anchor line: Use the anchor line generated in Step 7 as the spatial reference benchmark to predict the horizontal offset of the ship target contour in the range compression domain relative to this anchor line; Based on the predicted horizontal offset, achieve precise calculation and positioning of the ship target position in the range compression domain.

2. The ship target detection method in the SAR range compression domain based on dynamic generation of anchor lines according to claim 1, characterized in that, The application of the large kernel attention in Step 6 is as follows: The large kernel attention module is placed after the side layer of the FPN; In addition, a multi-scale aggregator is used to quantify the correlation between input tokens, and its formula is expressed as: (1.1) (1.2) （1.3） Among them, four feedforward paths are marked as and are distinguished by different colors, where corresponds to an identical forward path.

3. A method for detecting ship targets in the SAR range compression domain based on dynamic generation of anchor lines according to claim 1, characterized in that, The specific steps for dynamically generating the anchor line described in Step 7 specifically include the following steps: Step 7.1, Generation of starting point position probability heatmap: Construct a key point detection branch, process the feature map through two-level convolutional layers , and output a probability heatmap with a resolution of ; The heatmap represents the probability distribution of each pixel point being the starting point of the anchor line; ​ During the training phase, based on the starting coordinates of the ship target curve generate the ground truth heatmap; calculate the downsampled coordinates , where corresponds to the feature map stride; then, with as the center, construct a Gaussian distribution region, and the ground truth heatmap at the coordinate is represented by the following function: （1.4） Among them, is the spatial index of the heat map; the standard deviation controls the width of the probability distribution; In addition, to alleviate the problem of sample imbalance between the starting region and non-starting regions, an improved focal loss function is adopted to process heatmap prediction; let and represent the predicted probability value and the true label at the position in the heatmap respectively, and the loss function is defined as: （1.5） Among them, and are adjustable hyperparameters; represents the number of ship curves in the distance compression domain; the term is used to reduce the penalty around the true position; Step 7.2, starting point offset estimation: In the starting point precise positioning stage, the initial position estimation is derived from the peak detection of the probability distribution map, and its accuracy is restricted by the downsampling ratio of the feature map ; To eliminate the positioning error, a dynamic offset correction mechanism is introduced, and sub-pixel positioning is achieved through the two-dimensional offset prediction map output in parallel; During the model training process, a spatial constraint optimization strategy is adopted - only a square effective area with a side length of is demarcated with the real starting point coordinates as the center, and a smooth loss function is applied within this defined area for offset supervision. This constraint mechanism significantly improves the focusing of offset learning, and its mathematical definition is: (1.6) Among them, and respectively represent the offsets of the effective position relative to in axis and axis directions; Step 7.3, Adaptive slope estimation mechanism: In the adaptive optimization link of the anchor line slope, a slope parameter prediction branch is added, and the output dimension is slope feature map , and this slope feature map directly determines the spatial inclination characteristics of the dynamic anchor line; through the slope self-adaptation mechanism, the generated anchor line can adaptively fit the morphological contour of the target ship; in addition, for the ship target curve , where represents the termination point index, and the key point geometric relationship modeling technology is adopted: taking the starting point as the reference point, the average slope of the lines connecting the remaining valid points in the curve to the starting point is calculated as the true value label, and its mathematical representation is: (1.7) During the training phase, the same valid training region as that for offset estimation is selected, and a loss is applied to constrain the slope prediction, which is defined as follows: （1.8） Among them, is the index of the ship target curve sequence; Step 7.4, Anchor Line Decoding and Spatial Index Optimization Mechanism: Adopt a two-stage peak confirmation strategy. First, perform a max pooling operation on the starting probability distribution map to generate a primary response map, and then achieve secondary verification through a logical AND operation to accurately extract the coordinates of each response peak in the distribution map. ; The anchor line generation process adopts the coordinate refinement formula: , where the offset is taken from the predicted value at the position in the offset map , and the slope is obtained from the corresponding spatial position of the slope map . ​​ 4. A ship target detection method in the SAR range compression domain based on dynamic generation of anchor lines according to claim 1, characterized in that, The specific steps for the detection head described in Step 8 specifically include the following steps: Step 8.1, Anchor Line Feature Modeling Stage: Based on the dynamic anchor line parameterized coordinates , uniformly sample spatial key points along the anchor line trajectory; Through the bilinear interpolation algorithm, extract the deep features of each sampling point in the feature map, and temporally splice all feature vectors to construct an RoI feature tensor representing the spatial distribution characteristics of the anchor line; Step 8.2, Target Location Parameter Decoding Method: Input the RoI features of the anchor line into the fully-connected regression network to synchronously generate three groups of location parameters: the horizontal offset vector , the starting index , and the extension length ; Based on the image height , construct the vertical coordinate sequence , and calculate the target horizontal coordinate through the anchor line calibration formula . This decoding mechanism realizes the joint modeling of the target position, scale, and shape through a single forward propagation: （1.9） Based on the above analysis, the overall loss can be defined as: （1.10） Among them, is the Smooth loss, which is used to supervise the prediction of the starting index and the curve length ; is calculated based on the predicted curve and the true curve trajectory and is used to constrain the offset regression ; , , , and are the weight coefficients of the corresponding losses respectively.

Images