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

Abstract

The present application relates to the technical field of target detection, and particularly relates to a SAR distance compression domain ship target detection method based on anchor line dynamic generation; comprising the following steps: step 1, obtaining spaceborne SAR echo data after BAQ decompression processing; step 2, based on the spaceborne SAR echo data, performing matching filtering processing in the distance direction, thereby obtaining corresponding distance compression domain preliminary imaging; step 3, using a 1024x1024 sliding window for sliding window in the distance compression domain image; step 4, using Resnet101 as a backbone network to extract four scale features of the target in the image; step 5, performing fusion processing on the feature maps,, and, to obtain multi-scale features,, and ; step 6, embedding a large kernel attention between the feature maps and of the feature pyramid; step 7, anchor line dynamic generation; step 8, anchor line-based target accurate positioning. The amplitude gradient and geometric scale characteristics of the ship target in the distance compression domain are used to improve the detection performance.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of target detection, and particularly relates to a SAR distance compression domain ship target detection method based on anchor line dynamic generation. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) has the characteristics of all-weather operation and high-resolution imaging, and exhibits unique advantages in ship positioning. At present, SAR image ship target detection 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 features of an optical image; and then, traditional computer vision or deep learning algorithms are used to realize ship target detection and positioning.

[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, which is difficult to meet the real-time requirements; on the other hand, motion targets will cause defocus artifacts, and the sidelobe interference effect significantly reduces the image signal-to-noise ratio, affecting the detection accuracy and reliability. SUMMARY

[0004] The application provides a SAR distance compression domain ship target detection method based on anchor line dynamic generation, which aims to solve the problem of low efficiency caused by invalid data processing in the traditional "full scene imaging - image analysis" paradigm through the cooperative mechanism of target specific region detection and local region fast imaging in the distance compression domain, so as to significantly enhance the timeliness of sea ship target situation awareness.

[0005] To achieve the above-mentioned purpose, the technical scheme of the application is as follows:

[0006] The application provides a SAR distance compression domain ship target detection method based on anchor line dynamic generation, which includes the following steps:

[0007] Step 1: obtaining spaceborne SAR echo data processed by BAQ decompression;

[0008] Step 2: based on the spaceborne SAR echo data, performing matching filtering processing in the distance direction to obtain corresponding distance compression domain preliminary imaging;

[0009] Step 3: using a 1024x1024 sliding window for sliding window in the distance compression domain image;

[0010] Step 4: using Resnet101 as the backbone network to extract four scale features of the target in the image, i.e., 8, 16 and 32 times down-sampling features , 、 ;

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

[0012] Step 6, embedding large kernel attention between the feature maps and of the feature pyramid, expanding the receptive field of the detection model and effectively capturing the global context information of the slender and continuous ship curve;

[0013] Step 7, anchor line dynamic generation: using the obtained multi-scale features 、 、 , the starting point coordinates, the offset relative to the starting point and the slope parameters of the anchor line are calculated simultaneously through three independent prediction branches; based on these calculation results, for each recognized ship target instance, an anchor line is dynamically generated in real time, which can accurately locate and represent the direction characteristics of the instance;

[0014] Step 8, target accurate positioning based on anchor line: using the anchor line generated in step 7 as a spatial reference benchmark, predicting the horizontal offset of the ship target contour in the distance compression domain relative to the anchor line; based on the predicted horizontal offset, the accurate calculation and positioning of the ship target position in the distance compression domain are realized.

[0015] Further, 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 feedforward paths are marked as , and are distinguished by different colors, where corresponds to a same forward path.

[0021] Further, the anchor line dynamic generation in step 7 specifically comprises the following steps:

[0022] Step 7.1, starting point position probability heat map generation: a key point detection branch is constructed to process the feature map through two convolution layers , and output a probability heat map with a resolution of ; the heat map represents the probability distribution of each pixel point being the starting point of the anchor line;

[0023] In the training stage, the ground truth heat map is generated based on the coordinates of the starting point of the ship target curve ; the down-sampling coordinates are calculated, wherein corresponds to the feature map step size; then, a Gaussian distribution region is constructed with as the center, and the label value of the ground truth heat map at the coordinates is represented by the following function:

[0024] (1.4)

[0025] wherein is the spatial index of the heat map; the standard deviation controls the width of the probability distribution;

[0026] In addition, 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 heat map prediction; let and represent the predicted probability value and the true value label at position in the heat map, respectively, and the loss function is defined as:

[0027] (1.5)

[0028] wherein 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;

[0029] Step 7.2, starting point offset estimation: in the accurate positioning stage of the starting point, the initial position estimation is derived from the peak value detection of the probability distribution map, and the accuracy is restricted by the feature map down-sampling ratio ; to eliminate the positioning error, a dynamic offset correction mechanism is introduced, and sub-pixel level positioning is realized through the parallel output of the two-dimensional offset prediction map ;

[0030] In the model training process, a spatial constraint optimization strategy is adopted — only the true starting point coordinates are used to define a square with a side length of Square effective area, smoothing is applied within this limited area The loss function is used to supervise the offset. This constraint mechanism significantly improves the focus of offset learning. Its mathematical definition is:

[0031] (1.6)

[0032] in, and Respectively represent the effective position relative to exist Axis and Axis offset;

[0033] Step 7.3, Adaptive Slope Estimation Mechanism: In the anchor line slope adaptive optimization link, this application adds a slope parameter prediction branch, and the output dimension is Slope characteristic map The slope characteristic 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 target ship shape contour; in addition, for the ship target curve ,in Represents the end point index, using key point geometric relationship modeling technology: taking the starting point as the reference point, the average slope of the line connecting the remaining valid points in the curve and the starting point is calculated as the true value label. Its mathematical representation is:

[0034] (1.7)

[0035] During the training phase, the same valid training region as the offset estimation is selected and applied The loss is used to constrain the slope prediction, which is defined as follows:

[0036] (1.8)

[0037] in, 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 probability distribution map to generate a primary response map. Then, a secondary check is performed through a logical AND operation to accurately extract the coordinates of each response peak in the distribution map. ; The anchor line generation process uses the coordinate refinement formula: , where the offset Taken from the offset map The predicted value of the middle position ,slope From the slope diagram The corresponding spatial position is obtained.

[0039] Further, the detection head in step 8 specifically comprises the following steps:

[0040] Step 8.1, anchor line feature modeling stage: based on dynamic anchor line parameterized coordinates , uniformly sampling along the anchor line trajectory a spatial key point; through a bilinear interpolation algorithm, deep features of each sampling point are extracted in a feature map, and all feature vectors are sequentially spliced to construct a RoI feature tensor representing the spatial distribution characteristics of the anchor line ;

[0041] Step 8.2, target positioning parameter decoding method: inputting the anchor line RoI feature into a fully connected regression network, three sets of positioning parameters are simultaneously generated: horizontal offset vector , starting index and extension length ; based on the image height , a vertical coordinate sequence is constructed , and the target horizontal coordinate is calculated through an anchor line calibration formula ; this decoding mechanism realizes the joint modeling of target position, scale and morphology through single forward propagation:

[0042] (1.9)

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

[0044] (1.10)

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

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

[0047] ​(1) Detection performance improvement: The amplitude gradient and geometric scale characteristics of the ship target in the range compression domain are used to improve the detection performance. A continuous and slender anchor line is used to locate the ship target in the range compression domain, and the feature learning network is forced to capture the global context information of the ship to improve the accuracy of detection.

[0048] (2) Processing efficiency optimization: Compared with the traditional imaging-detection SAR processing, the method proposed in the present application only completes the suspected ship target region positioning in the range compression domain of the SAR echo, eliminates most of the background regions, greatly reduces the data processing for the subsequent imaging step, meets the low-delay fast extraction of the SAR satellite echo to the target slice and other information of interest, and lays a foundation for the improvement of the on-orbit imaging, intelligent level and on-orbit application efficiency of future SAR satellites. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description only represent some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings shown.

[0050] Figure 1 The figure shows a schematic diagram of the target detection grid structure of the method of the present application.

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

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely in the following description of the drawings of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] It should be noted that if the embodiments of the present application involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative position relationship, motion condition, etc. between the components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly.

[0054] In addition, if the description of "first", "second" and the like is involved in the embodiments of the present application, the description of "first", "second" and the like is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can be explicitly or implicitly included at least one of the features. In addition, the meaning of "and / or" appearing throughout the text is that three parallel schemes are included, for example, "A and / or B" includes A scheme, or B scheme, or A and B schemes are satisfied at the same time. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of the ordinary skilled in the art, when the combination of technical solutions appears contradictory or cannot be realized, it should be considered that the combination of technical solutions does not exist, and is not within the protection scope required by the present application.

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

[0056] In view of the technical bottlenecks, the present application proposes a fast detection optimization strategy based on echo domain feature analysis: first, the potential target area is roughly positioned through echo signal feature fast extraction technology, and then dynamic adaptive imaging processing is implemented for the key area. 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, the whole link processing efficiency of the SAR system from data acquisition to target recognition is significantly improved.

[0057] As shown in Figure 1 The present application provides a SAR distance compression domain ship target detection method based on anchor line dynamic generation, which specifically comprises the following steps:

[0058] Step 1, obtaining the spaceborne SAR echo data processed by BAQ decompression;

[0059] Step 2, based on the above spaceborne SAR echo data, performing matching filtering processing in the distance direction, thereby obtaining the corresponding distance compression domain preliminary imaging;

[0060] Step 3, using a sliding window of 1024x1024 in the distance compression domain image for sliding window;

[0061] Step 4, using Resnet101 as the backbone network to extract four scale features of the target in the image, i.e. 8, 16, 32 times down-sampling features 、 、 ;

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

[0063] Step 6, as shown in the figure, a large kernel attention is embedded between the feature maps Figure 1 of the feature pyramid and , so as to expand the receptive field of the detection model and effectively capture the global context information of the slender and continuous ship curve;

[0064] Step 7, anchor line dynamic generation: using the obtained multi-scale features 、 、 , the starting point coordinates, the offset relative to the starting point and the slope parameters of the anchor line are calculated simultaneously through three independent prediction branches. Based on these calculation results, for each recognized ship target instance, an anchor line is calculated and output in real time, which can accurately locate and represent the direction characteristics of the instance;

[0065] Step 8, target accurate positioning based on anchor line: using the anchor line generated in step 7 as a spatial reference datum, the horizontal offset of the ship target contour in the distance compression domain relative to the anchor line is predicted. Based on the predicted horizontal offset, the accurate calculation and positioning of the ship target position in the distance compression domain are realized.

[0066] As a further preferred scheme of the SAR distance compression domain ship target detection method based on anchor line dynamic generation of the application, the application of the large kernel attention in step 6 is as follows:

[0067] As shown in the figure, the large kernel attention (LKA) module is placed after the side layer of the FPN to minimize the calculation cost. Unlike the similarity score between the generated query and the value output, the application adopts a multi-scale aggregator (MSA) to quantify the correlation between input tokens, and its formula can be expressed as: Figure 2

[0068] (1.1)

[0069] (1.2)

[0070] ​(1.3)

[0071] in, It 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. It is a learnable weight matrix that is used to perform linear transformation on the features after multi-channel deep 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 applies to the input feature map Perform linear transformation and map it to a suitable feature space to facilitate integration with attention features Perform element-wise multiplication , to achieve feature modulation. Indicates multi-channel operation, the subscript i ranges from 0 to 3, which means processing different channels 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, which is used to perform convolution on the input feature map to extract features of different channels and capture spatially relevant information. The feature map representing the input is the input data of the entire LKA module and contains the characteristic information of the image, such as the number of channels, spatial dimensions, etc. It is a feature map modulated by the attention mechanism, which combines the attention information and the linear transformation result of the input feature map, while retaining the input feature map The original information of (achieved through skip connections) helps the model better learn rich feature representations. It is the final output feature map, which combines the features processed by the feedforward network and FFN stands for Feed-Forward Network and is typically composed of two linear transformations and a nonlinear activation function. It is used to further process features and increase the model's expressiveness and nonlinear fitting capabilities.

[0072] The four feedforward paths are labeled , and are distinguished by different colors, among which Corresponding to the same forward path. Compared with using 31 × 31 and 7 × 7 depthwise convolution, striped convolution is more effective in recognizing the curves of ships while reducing the computational cost. The linear layer consists of The Hadamard product (denoted by ⊙ in Equation 1.2) is used instead of the matrix product to take advantage of the large kernel in MSA.

[0073] As a further preferred scheme of the SAR distance compression domain ship target detection method based on anchor line dynamic generation of the application, the anchor line dynamic generation in step 7 specifically comprises the following steps:

[0074] Step 7.1, starting point position probability heat map generation: construct a key point detection branch, process the feature map through two convolution layers , and output a probability heat map with a resolution of . The heat map represents the probability distribution of each pixel point being an anchor line starting point.

[0075] In the training stage, a true value heat map is generated based on the coordinates of the starting points of the ship target curve ; wherein, is the horizontal coordinate of the starting point of the ship target curve, is the vertical coordinate of the starting point of the ship target curve. The down-sampling coordinates are calculated, wherein corresponds to the feature map step size. Then, a Gaussian distribution region is constructed with as the center, and the label value of the true value heat map at the coordinates is represented by the following function:

[0076] (1.4)

[0077] wherein, is the spatial index of the heat map; 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 maximum exponential function value operation, that is, among all the Gaussian distributions corresponding to the starting points of the ship target curve, the maximum value at the coordinates is taken; is the horizontal coordinate of the down-sampled starting point of the ship target curve; is the vertical coordinate of the down-sampled starting point of the ship target curve.

[0078] In order 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 heat map prediction. Let and represent the predicted probability value and the true value label at the position in the heat map respectively, and the loss function is defined as:

[0079] (1.5)

[0080] wherein, and​ is an adjustable hyper-parameter, whose value is usually related to the proportion of positive and negative samples in the dataset, the difficulty of the task, and the requirement for prediction accuracy, and needs to be determined according to the actual situation, which is not limited in this paper; represents the number of ship curves in the distance compression domain; the term is used to reduce the penalty around the real position; is the height of the true value heat map, that is, the dimension of the heat map in the vertical direction after downsampling; is the width of the true value heat map, that is, the dimension of the heat map in the horizontal direction after downsampling.

[0081] Step 7.2, starting point offset estimation: in the starting point accurate positioning stage, the initial position estimation is derived from the peak detection of the probability distribution map, and its accuracy is restricted by the feature map downsampling ratio . In order to eliminate the positioning error, this patent introduces a dynamic offset correction mechanism, which realizes sub-pixel level positioning through the two-dimensional offset prediction map output in parallel. This offset map has a double 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 downsampling coordinates ), and second, it real-time rectifies the initial position offset existing in the probability distribution prediction.

[0082] In the model training process, a spatial constraint optimization strategy is adopted —— only a square effective area with a side length of is defined with the real starting point coordinates as the center, and a smooth loss function is applied within this limited area to supervise the offset, which significantly improves the focus of offset learning, and its mathematical definition is:

[0083] (1.6)

[0084] wherein and respectively represent the offset of the effective position relative to in the axis and axis direction; is the offset predicted by the model.

[0085] Step 7.3, adaptive slope estimation mechanism: in the adaptive optimization of anchor line slope, this scheme adds a slope parameter prediction branch, which outputs a slope feature map with a dimension of , which directly determines the spatial inclination characteristics of the dynamic anchor line. Through the slope self-adaption mechanism, the generated anchor line can adapt to the target ship shape contour, significantly improving the fitting accuracy and recognition efficiency of subsequent target detection. For the ship target curve (in represents the end point index), using key point geometric relationship modeling technology: taking the starting point as the reference point, the average slope of the line connecting the remaining valid points in the curve and the starting point is calculated as the true value label, which is mathematically represented as:

[0086] (1.7)

[0087] in, is the abscissa of the starting point of the ship target curve; is the ordinate of the starting point of the ship target curve; The first The horizontal coordinate of the valid points; The first The vertical coordinate of a valid point; is the index of the starting point of the ship target curve; is the number of valid points of the ship target curve.

[0088] During the training phase, the same valid training region as the offset estimation is selected and applied The loss is used to constrain the slope prediction, which is defined as follows:

[0089] (1.8)

[0090] in, is the index of the ship target curve sequence; The model predicts the location Slope characteristic map; is the true value label of the slope mean corresponding to the kth 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, a maximum pooling operation is performed on the starting probability distribution map to generate a primary response map, and then a secondary check is performed 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 distance-based or IoU-based non-maximum suppression methods. The obtained peak coordinates have a dual function: they serve as the initial estimate of the starting position and as 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 Taken from the offset map The predicted value of the middle position ,slope From the slope diagram The corresponding spatial position acquisition (Note: the index position should be unified as to ensure spatial consistency).

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

[0093] Step 8.1, anchor line feature modeling stage: based on dynamic anchor line parameterized coordinates , uniformly sampling key points along the anchor line track. Through the bilinear interpolation algorithm, the deep features of each sampling point are extracted in the feature map, and all feature vectors are time-series spliced to construct the RoI feature tensor characterizing the spatial distribution characteristics of the anchor line.

[0094] Step 8.2, target positioning parameter decoding method: input the anchor line RoI feature into a fully connected regression network to simultaneously generate three sets of positioning parameters: horizontal offset vector (characterizing the spatial deviation of the real contour of the ship target and the anchor line), starting index and extension length ; based on the image height , the vertical coordinate sequence is constructed , and the target horizontal coordinate is calculated by the anchor line calibration formula (where achieves the geometric conversion of slope to horizontal displacement), and the decoding mechanism realizes the joint modeling of target position, scale and morphology through single forward propagation:

[0095] (1.9)

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

[0097] (1.10)

[0098] where is the Smooth loss used to supervise the prediction of the starting index and the curve length . is calculated based on the predicted curve and the real curve trajectory, and is used to constrain the offset regression . , , , and are weight coefficients corresponding to the losses, which are usually obtained by empirical values or experimental adjustment to achieve a balance between different loss terms, and the specific values are related to loss balance in the training process, task objectives, and data set characteristics, and other factors.

[0099] Through the above steps, the suspected target area can be quickly positioned on a wide sea surface, the time and computational resource consumption required for subsequent focused imaging are effectively reduced, and the present application can learn the strong distinguishing characteristics of the ship target and sea clutter according to the amplitude characteristics and geometric characteristics of the ship echo energy focusing in the range direction and defocusing in the azimuth direction, and quickly position the suspected target area.

[0100] The above only describes optional embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structural transformation made under the inventive concept of the present application, using the contents of the present application specification and drawings, or directly / indirectly applied in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A ship target detection method based on anchor line dynamic generation of SAR range compression domain, characterized in that, The method comprises the following steps: Step 1, obtaining the spaceborne SAR echo data after BAQ decompression processing; Step 2, based on the spaceborne SAR echo data, performing matching filtering processing in the range direction to obtain corresponding range compression domain preliminary imaging; Step 3, using a sliding window of 1024*1024 to slide in the range compression domain image; Step 4, using Resnet101 as the backbone network, extracting 4 scale features of the target in the image, i.e. features with 8, 16, 32 times of down-sampling , , ; Step 5, according to the generation rule of the feature pyramid, the feature map 、 、 is fused to obtain multi-scale features 、 、 ; Step 6, embedding large kernel attention between the feature maps of the feature pyramid and which expands the receptive field of the detection model and effectively captures the global context information of the slender and continuous ship curve. Step 7, anchor line dynamic generation: using the obtained multi-scale features 、 、 , simultaneously calculate the starting point coordinates, offset relative to the starting point and slope parameters of the anchor line through three independent prediction branches; based on these calculation results, for each recognized ship target instance, real-time calculation and output an anchor line dynamic generation that can accurately locate and represent the direction characteristics of the instance. Step 8, anchor line-based target accurate positioning: using the anchor line generated in step 7 as a spatial reference benchmark, predicting the horizontal offset of the ship target contour in the range compression domain relative to the anchor line; based on the predicted horizontal offset, the accurate calculation and positioning of the ship target position in the range compression domain are realized.

2. The SAR range cell compression domain ship target detection method 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 represented as: (1.1) (1.2) （1.3） Wherein the four feedforward paths are labeled as and distinguished by different colors, wherein corresponding to one same forward path.

3. The SAR range cell compression domain ship target detection method based on dynamic generation of anchor lines according to claim 1, characterized in that, The anchor line dynamic generation in step 7 specifically comprises the following steps: Step 7.1, starting point position probability heat map generation: construct a key point detection branch to process the feature map through two convolution layers , output a probability heat map with a resolution ; the heat map represents the probability distribution of each pixel being the starting point of the anchor line;​ In the training stage, based on the ship target curve starting point coordinates Generate the true value heat map; calculate the down-sampling coordinates Wherein The corresponding feature map step; then, a Gaussian distribution region is constructed with The true value heat map The label value of the coordinates is represented by the following function: （1.4） wherein, is the spatial index of the heat map; standard deviation controls the width of the probability distribution; 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; let and respectively represent the predicted probability value and the true value label of the position in the heat map, and the loss function is defined as: （1.5） wherein, and are adjustable hyperparameters; denotes the number of distances to the ship curve in the compressed domain; the term is used to reduce the penalty around the true position; Step 7.2, start point offset estimation: in the start point accurate positioning stage, the initial position estimation is derived from the peak detection of the probability distribution map, whose accuracy is restricted by the feature map down-sampling ratio; to eliminate the positioning error, a dynamic offset correction mechanism is introduced, and sub-pixel level positioning is realized through the two-dimensional offset prediction map output in parallel ;​ During the model training process, a spatial constraint optimization strategy is used, which is to define the side length of the model with the real starting point coordinate as the center. Square effective area, smoothing is applied within this limited area The loss function is used to supervise the offset. This constraint mechanism significantly improves the focus of offset learning. Its mathematical definition is: (1.6) wherein, and respectively represent the effective position relative to the offset in the direction of the axis and axis direction; Step 7.3, adaptive slope estimation mechanism: in the anchor line slope adaptive optimization link, add a slope parameter prediction branch, the output dimension is the slope feature map , which directly determines the spatial inclination characteristics of the dynamic anchor line; through the slope self-adaption mechanism, the generated anchor line can adapt to the target ship shape contour; in addition, for the ship target curve , where indicates the end point index, adopts key point geometric relationship modeling technology: taking the starting point as the reference point, calculating the slope average of the connecting line between the remaining effective points in the curve and the starting point as the true value label, which is mathematically represented as: (1.7) In the training phase, the same effective training region as for the bias estimation is selected, and the following loss is applied to constrain the slope prediction, which is defined as follows: （1.8） wherein, is an index of the sequence of ship target curves; Step 7.4, Anchor Line Decoding and Spatial Index Optimization Mechanism: A two-stage peak confirmation strategy is adopted. First, a maximum pooling operation is performed on the starting probability distribution map to generate a primary response map. Then, a secondary check is performed through a logical AND operation to accurately extract the coordinates of each response peak in the distribution map. ; The anchor line generation process uses the coordinate refinement formula: , where the offset Taken from the offset map The predicted value of the middle position ,slope From the slope diagram The corresponding spatial position is obtained.

4. The SAR range cell compression domain ship target detection method based on dynamic generation of anchor lines according to claim 1, characterized in that, The detection head in step 8 specifically comprises the following steps: Step 8.1, anchor line feature modeling stage: dynamic anchor line parameterized coordinates , uniform sampling along anchor line trajectory key points in space; through a bilinear interpolation algorithm in feature map to extract the deep features of each sampling point, and all feature vectors are time series spliced to construct a RoI feature tensor representing the spatial distribution characteristics of the anchor line ; Step 8.2, target localization parameter decoding method: anchor line RoI feature Input the fully connected regression network to synchronously generate three sets of localization parameters: horizontal offset vector , starting index , and 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 target position, scale and morphology through single forward propagation: （1.9） Based on the above analysis, the overall loss can be defined as: （1.10） where, is Smooth Loss, used to supervise the predicted start index and curve length ; is computed based on the predicted curve and the real curve trajectory, used to constrain the offset regression ; , , , and are the weight coefficients for the corresponding losses, respectively.

Images