SAR ship detection method based on distance compression domain characteristic guidance

Abstract

The invention relates to the technical field of SAR image target detection, in particular to an SAR ship detection method based on distance compression domain characteristic guidance. Comprising the following steps: step 1, acquiring spaceborne SAR echo data subjected to BAQ decompression processing; 2, matched filtering processing is executed in the distance direction, so that a corresponding distance compressed domain preliminary image is obtained; 3, a 1024 * 1024 sliding window is used for window sliding; step 4, adding a strip attention mechanism module into a backbone network ResNet50 of the RT-DETR reference model; step 5, adding a SimVSS module into a hybrid encoder of the RT-DETR reference model; step 6, inputting the features into an IoU perception query selection of the RT-DETR; step 7, performing multiple rounds of optimization on the screened candidate boxes; and step 8, adding shape-perceived IoU regression positioning loss. The ship target amplitude gradient and geometric scale characteristics in the distance compression domain are utilized, the detection capability of the ship target is effectively enhanced, the detection performance is remarkably improved, and the ship target can be recognized more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of SAR image target detection, and in particular to a SAR ship detection method based on range compression domain characteristic guidance. Background Art

[0002] The ocean covers 71.8% of the Earth's surface, constituting humanity's second living space after land. Ships, as the core means of water transportation, carry out multiple functions, including transportation, fishing activities, and national defense security. Therefore, effective monitoring of ships is particularly important. Currently, synthetic aperture radar (SAR) technology plays a key role in the monitoring and identification of ship targets, mainly due to its ability to generate high-resolution imagery with global coverage and achieve continuous observation under all-weather and all-time conditions. In view of this, the development of a reliable and efficient SAR ship detection algorithm is of critical strategic significance for enhancing maritime safety management and monitoring effectiveness.

[0003] Traditional SAR image target detection first performs two-dimensional imaging processing on the collected raw SAR echo domain data to generate a focused image similar to a visible light image. Subsequently, traditional or deep learning target detection techniques are used to automatically detect and locate ship targets from the focused image. However, SAR ship detection methods based on this imaging paradigm face challenges in practical applications of wide-area maritime ship detection, including computationally intensive full-scene imaging and poor image quality caused by defocusing moving targets and sidelobe interference. Summary of the Invention

[0004] The present invention provides a SAR ship detection method based on range compression domain characteristic guidance. Its purpose is to first complete the detection of a specific area with a target in the range compression domain, and then quickly image only the specific local area, thereby solving the problem of low efficiency caused by invalid data processing in traditional technologies, and improving the timeliness of maritime ship target situation awareness.

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

[0006] The present invention provides a SAR ship detection method based on range compression domain characteristic guidance, comprising the following steps:

[0007] Step 1: Obtain spaceborne SAR echo data after BAQ decompression processing;

[0008] Step 2: Based on the above-mentioned spaceborne SAR echo data, matched filtering is performed in the range direction to obtain the corresponding preliminary imaging in the range compression domain;

[0009] Step 3: Slide the image in the range compression domain using a 1024×1024 sliding window.

[0010] Step 4: Add a strip attention mechanism module to the ResNet50 backbone network of the RT-DETR benchmark model and extract three scale features of the target in the 1024×1024 range compression domain image, that is, sample the multi-scale feature maps {S3, S4, S5} at 8, 16, and 32 times;

[0011] Step 5: Add the SimVSS module to the hybrid encoder of the RT-DETR benchmark model, and feed the above 8-, 16-, and 32-fold multi-scale feature maps {S3, S4, S5} into the hybrid encoder based on the SimVSS module to extract multi-scale features, and then perform cross-scale fusion of the multi-scale features.

[0012] Step 6: Input the features output by the hybrid encoder into the IoU-aware query selection of RT-DETR to select the areas with high overlap (IoU) with the true target as the initial candidate boxes;

[0013] In step 7, RT-DETR's Transformer decoder performs multiple rounds of optimization on the selected candidate boxes, analyzes the relationship between the objects through the self-attention mechanism, gradually adjusts the position and size of the candidate boxes, and finally outputs accurate detection results;

[0014] Step 8: In the RT-DETR model, in order to accurately locate the slender ship targets in the range compression domain, a shape-aware IoU regression positioning loss is added to the auxiliary prediction head structure.

[0015] Furthermore, the strip attention mechanism module described in step 4 specifically includes the following steps:

[0016] Step 4.1, define the strip attention mechanism module: given the input feature F∈R C×H×W , first perform three 1×1 convolution operations on it, and get and V∈R C×H×W These three feature maps; perform matrix dimension transformation and transposition operation on the feature map Q to generate N is H×W, and then the StripPooling operation is applied to the feature map K, that is, the average value of each row of pixels in the feature map is calculated to generate Then, the feature map and feature maps Perform product operation and finally obtain the position attention mapping result A∈R through Softmax operation N×W ; In essence, the feature map and The matrix multiplication operation is the product of the eigenvector corresponding to the position and the eigenvectors of all other positions, so the similarity calculation formula between positions is:

[0017]

[0018] Among them, Q i Represents the i-th position element of the feature map Q; K j Represents the j-th position element of feature map K, A j，i Represents Q i With K j the degree of correlation between them;

[0019] Furthermore, the feature map V is stripped and pooled horizontally, and then the dimension is converted to After matrix multiplication with matrix A, the dimension is converted to Final feature map Add the corresponding elements of the input feature F to get the final feature map

[0020] Step 4.2, embedding position of the SAM module: In the BasicBlock or Bottleneck structure of ResNet, insert the SAM module after the last convolutional layer of the residual path and perform channel weighting operation before adding the skip connection;

[0021] Step 4.3, multi-level SAM module configuration: Add a SAM module at the output of each stage of ResNet to form a global-local joint attention mechanism.

[0022] Furthermore, the SimVSS module described in step 5 specifically includes the following steps:

[0023] Step 5.1, define the SimVSS module: its design is mainly based on SSM and a feedforward network; the structure contains residual connections and normalization layers, which are designed to stabilize the gradient training process and accelerate model convergence; the traditional SSM is a linear time-invariant system function, which is implemented through intermediate variables The univariate series Mapping to output sequence Specifically, given the state transition matrix As the evolution factor, weight matrix and observation matrix B, P∈C N As projection factors, and defined skip connections Its mathematical expression is as follows:

[0024]

[0025] In addition, the system function is implemented by introducing a time scale parameter To discretize it to process discrete time series data; this transformation is defined as follows:

[0026]

[0027] in, represents the mapping from state space to output space; in order to refine the approximation of B using a first-order Taylor series expansion:

[0028]

[0029] For input The processing steps within the SimVSS block are described as follows:

[0030]

[0031] Among them, Z l-1 、Z l 、Z l+1 They represent the output states of different layers l of the SimVSS block;

[0032] In step 5.2, the SimVSS module is embedded in the middle of the attention-based intra-scale feature interaction and the convolutional neural network-based cross-scale feature fusion.

[0033] Furthermore, the shape-aware IoU regression loss described in step 8 is specifically:

[0034] The shape-aware IoU regression loss calculates the loss by focusing on the shape and scale of the bounding box itself, thereby enabling more accurate bounding box regression. Its calculation formula is as follows:

[0035]

[0036] Among them, scale is the scale factor, B and B gt The predicted box and the true value box, w and w respectively gt They are the predicted box B and the true value box B gt The width, h and h gt They are the predicted box B and the true value box B gt The height of , ww and hh are the weight coefficients in the horizontal and vertical directions respectively.

[0037] The beneficial effects achieved by the present invention are:

[0038] 1) Performance improvement: By utilizing the amplitude gradient and geometric scale characteristics of ship targets in the range compression domain, the detection capability of ship targets is effectively enhanced, the detection performance is significantly improved, and ship targets can be identified more accurately.

[0039] 2) Feature Expression: The design of the strip attention mechanism module is consistent with the long strip shape characteristics of ship targets and can effectively express their shape characteristics. This helps to accurately distinguish ship targets in complex backgrounds and avoid detection errors caused by shape misjudgment.

[0040] 3) Receptive field expansion and global information learning: Incorporating feature fusion based on the SimVSS module into the hybrid encoder expands the receptive field of the detection network, enabling the network to learn the global contextual information between the ship target and other targets and the background, thereby better understanding the target's environment and improving the reliability and accuracy of detection.

[0041] 4) Enhanced positioning capability: The introduction of shape-aware IoU regression loss optimizes the detection network’s prediction of target positions, significantly enhancing the network’s positioning capability and ensuring more accurate detection of ship target positions.

[0042] 5) Efficient data processing: Compared with the traditional SAR processing method of imaging first and then detecting, the present invention can complete the positioning of the suspected ship target area only in the range compression domain of the SAR echo, eliminate a large amount of background area in advance, greatly reduce the data processing amount of the subsequent imaging steps, and realize low-latency and fast extraction of interesting information such as SAR satellite echo to target slice.

[0043] 6) Broad application prospects: This method lays a solid foundation for the future on-orbit imaging, intelligence level improvement and on-orbit application efficiency improvement of SAR satellites, and helps promote more efficient and intelligent application of SAR satellites in fields such as ocean monitoring and target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0045] Figure 1 The figure shows a schematic diagram of the target detection grid structure of the method of the present invention.

[0046] Figure 2 Shown is the module structure diagram of the strip attention mechanism of the method of the present invention.

[0047] Figure 3 Shown is the structural diagram of the SimVSS module of the method of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0049] It should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications will also change accordingly.

[0050] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, if the meaning of "and / or" appearing in the full text is to include three parallel schemes, taking "A and / or B" as an example, it includes scheme A, or scheme B, or a scheme in which A and B are satisfied at the same time. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in this field to implement. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0051] like Figures 1 to 3 As shown, the present invention provides a SAR ship detection method based on range compression domain characteristic guidance, comprising the following steps:

[0052] Step 1: Obtain spaceborne SAR (Synthetic Aperture Radar) echo data that has undergone BAQ (Block Adaptive Quantization) decompression processing;

[0053] Step 2: Based on the above-mentioned spaceborne SAR echo data, matched filtering is performed in the range direction to obtain the corresponding preliminary imaging in the range compression domain;

[0054] Step 3: Slide the image in the range compression domain using a 1024×1024 sliding window.

[0055] Step 4, such as Figure 1As shown in Figure 1, a strip attention mechanism module is added to the backbone network ResNet50 of the RT-DETR benchmark model, and three scale features are extracted from the target in the 1024×1024 range compression domain image, that is, multi-scale feature maps {S3, S4, S5} are sampled 8, 16, and 32 times;

[0056] Step 5, such as Figure 1 As shown in the figure, the SimVSS module is added to the hybrid encoder of the RT-DETR benchmark model, and the above 8, 16, and 32 times multi-scale feature maps {S3, S4, S5} are fed into the hybrid encoder based on the SimVSS module to extract multi-scale features, and then the multi-scale features are fused across scales;

[0057] Step 6, such as Figure 1 As shown in the figure, the features output by the hybrid encoder are input into the IoU-aware query selection of RT-DETR, and regions with high overlap (IoU) with the ground-truth object are selected as initial candidate boxes. Generally speaking, in object detection tasks, an IoU greater than 0.5 can be considered a relatively high overlap, but the specific threshold varies depending on the dataset and task requirements.

[0058] Step 7, such as Figure 1 As shown in the figure, RT-DETR's Transformer decoder performs multiple rounds of optimization on the selected candidate boxes, analyzes the relationship between the objects through the self-attention mechanism, gradually adjusts the position and size of the candidate boxes, and finally outputs accurate detection results;

[0059] Step 8, such as Figure 1 As shown in the figure, in the RT-DETR model, to accurately locate slender ship targets in the range compression domain, a shape-aware IoU regression positioning loss is added to the auxiliary prediction head structure. The specific process is as follows: In step 7, the Transformer decoder optimizes the position and size information of the candidate boxes to obtain preliminary optimized candidate boxes; then, in step 8, the auxiliary prediction head structure uses these optimized candidate boxes as the basis for calculating the shape-aware IoU regression positioning loss; based on this, the auxiliary prediction head performs more refined positioning adjustments, thereby achieving accurate detection and positioning of slender ship targets in the range compression domain.

[0060] As a further preferred embodiment of the SAR ship detection method based on distance compression domain characteristics guidance of the present invention, the strip attention mechanism module described in step 4 specifically includes the following steps:

[0061] Step 4.1, define the Strip Attention Module (SAM). Figure 2As shown in the figure, the SAM module can make full use of the Strip Pooling technology to deeply explore features along the same spatial dimension, thereby significantly enhancing the strip context feature expression capability of ship targets in the range compression domain.

[0062] Given input features F∈R C×H×W (H and W are the height and width respectively, C is the number of channels, and R represents the real number domain. First, three 1×1 convolution operations are performed on it to obtain and V∈R C×H×W These three feature maps. Perform matrix dimension transformation and transposition operation on the feature map Q to generate N is H×W, and then the Strip Pooling operation is applied to the feature map K, that is, the average value of each row of pixels in the feature map is calculated to generate Then, the feature map and feature maps Perform product operation and finally obtain the position attention mapping result A∈R through Softmax operation N×W In essence, the feature map and The matrix multiplication operation is the product of the eigenvector corresponding to the position and the eigenvectors of all other positions, so the similarity calculation formula between positions is:

[0063]

[0064] Among them, Q i Represents the i-th position element of the feature map Q; K j Represents the j-th position element of feature map K, A j,i Represents Q i With K j The degree of correlation between them.

[0065] Furthermore, the feature map V is stripped and pooled horizontally, and then the dimension is converted to After matrix multiplication with matrix A, the dimension is converted to Final feature map Add the corresponding elements of the input feature F to get the final feature map

[0066] Step 4.2, embedding position of the SAM module: In the BasicBlock or Bottleneck structure of ResNet, the SAM module is inserted after the last convolutional layer of the residual path, and the channel weighting operation is performed before the addition operation with the skip connection.

[0067] Step 4.3, multi-level SAM module configuration: Add SAM modules to the output of each stage of ResNet to form a global-local joint attention mechanism. For example, in the layers 2 to 4 of ResNet50, set SAM modules with dimensionality reduction ratios of r = 16, 8, and 4 respectively.

[0068] As a further preferred embodiment of the SAR ship detection method based on range compression domain characteristic guidance of the present invention, the SimVSS module described in step 5 specifically includes the following steps:

[0069] Step 5.1, define the SimVSS module. Figure 3 As shown in Figure 2, the SimVSS module is used in the hybrid encoder to expand the receptive field of the detection network so that it can learn the global contextual information between the ship target and other targets and background.

[0070] Its design is mainly based on SSM (State Space Model) and a feedforward network. The structure contains residual connections and normalization layers, which are designed to stabilize the gradient training process and accelerate model convergence. The traditional SSM can be regarded as a linear time-invariant system function, which is The univariate series Mapping to output sequence Specifically, given the state transition matrix As the evolution factor, weight matrix and observation matrix B, P∈C N As projection factors, and defined skip connections Its mathematical expression is as follows:

[0071]

[0072] Where N is the dimension of the state variable; t is the time variable; h(t) is the state variable that describes the internal state of the system at time t

[0073] In addition, the system function can be improved by introducing the time scale parameter To discretize it to process discrete time series data. This transformation can be defined as follows:

[0074]

[0075] Among them, h t represents the discrete-time state vector, the state variable at time step t; Represents the discretized state transfer matrix, which is the continuous-time state matrix A through the matrix index e ΔA It is converted to describe the evolution of the state in a discrete time system. k-2 represents the state vector at time k-2; Represents the discretized input matrix, which is determined by the continuous time input matrix B, the time scale parameter Δ, the parameter u and A, and reflects the impact of the input on the state. k is the input vector at time step k, the external signal that drives the system state change; t is the output vector at time step t, the observable quantity of the system; P is the output matrix, which converts the state vector h k Mapped to output y t , which is the same as the output matrix in the continuous-time system Represents the discretized skip connection matrix, which is used to represent the direct impact of input on output. Δ∈R Q is the time scale parameter (sampling period), which is used to discretize the continuous-time system and determine the discrete time step size. A is the continuous-time state matrix, which describes the continuous-time dynamic changes of the system state. B is the continuous-time input matrix, which reflects the impact of the input signal on the continuous-time state. μ is a coefficient used to adjust the input term during the discretization process. e ΔA is the matrix exponential, the core operation used to convert a continuous-time system to a discrete-time system.

[0076] In the process of discretization of system functions, Represents the mapping from state space to output space, where D represents the number of output features and N represents the dimension of the state space. In order to refine the approximation of B using a first-order Taylor series expansion:

[0077]

[0078] Where ΔB is the discretized approximate version of B. When a continuous-time system is discretized, it is usually necessary to approximate it. Here, a first-order Taylor expansion is used to approximate B. Specifically, It can be approximated by Taylor expansion as This is the effect of a change in state on the output under the influence of ΔA. This approximation can help simplify calculations and make the model more easily handle discrete time series data. In some cases, ΔB can also be understood as the incremental change introduced by the input signal during the discretization process.

[0079] For input The processing steps within the SimVSS block can be described as follows:

[0080]

[0081] Among them, Z l-1 、Z l 、Z l+1 They represent the output states of different layers l of the SimVSS block.

[0082] Step 5.2, such as Figure 1 As shown in Figure 3, the SimVSS module is embedded in the middle of attention-based intra-scale feature interaction (AIFI) and convolutional neural network-based cross-scale feature fusion (CCFM).

[0083] As a further preferred embodiment of the SAR ship detection method based on distance compression domain feature guidance of the present invention, the shape-aware IoU regression loss in step eight is specifically:

[0084] Shape-IoU regression loss (Shape-IoU) calculates the loss by focusing on the shape and scale of the bounding box itself, thereby achieving more accurate bounding box regression. Its calculation formula is as follows:

[0085]

[0086] Among them, L Shape-IOU It is the shape-aware IoU regression loss (Shape-IoU Loss), which is used to evaluate the predicted bounding box (B) and the true bounding box (B gt ) is a loss function that measures the difference in shape and position. The smaller its value, the higher the degree of match between the predicted box and the true box. t is the weight coefficient used to measure the mismatch degree of the bounding box in the horizontal (w) and vertical (h) directions; c is the central horizontal coordinate of the predicted bounding box; is the horizontal coordinate of the center of the real bounding box; c is the central ordinate of the predicted bounding box; is the central ordinate of the true bounding box; c is a constant used to control the scaling of the distance function (to prevent underflow or overflow problems in numerical calculations), which determines the weight of the distance in the loss calculation; scale is the scale factor, B and B gt The predicted box and the true value box, w and w respectively gt They are the predicted box B and the true value box B gt The width, h and h gt They are the predicted box B and the true value box B gt The height of , ww and hh are the weight coefficients in the horizontal and vertical directions respectively.

[0087] Through the above steps, the suspected target area can be quickly located on the wide sea surface, effectively reducing the time and computing resource consumption required for subsequent focused imaging. In addition, the invention can learn the strong distinguishing characteristics of ship targets and sea clutter based on the amplitude characteristics and geometric characteristics of the ship echo energy in the range focusing and azimuth defocusing, and quickly locate the suspected target area.

[0088] The present invention provides a new method and technical approach for wide-area marine ship target detection by spaceborne SAR, and has great practical application value.

[0089] The above descriptions are merely optional embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present description and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included in the patent protection scope of the present invention.

Claims

1. A SAR ship detection method based on range compression domain characteristics, characterized in that: The following steps are involved: Step 1: Obtain spaceborne SAR echo data after BAQ decompression processing; Step 2: Based on the above-mentioned spaceborne SAR echo data, matched filtering is performed in the range direction to obtain the corresponding preliminary imaging in the range compression domain; Step 3: Slide the image in the range compression domain using a 1024×1024 sliding window. Step 4: Add a strip attention mechanism module to the ResNet50 backbone network of the RT-DETR benchmark model and extract three scale features of the target in the 1024×1024 range compression domain image, that is, sample the multi-scale feature maps {S3, S4, S5} at 8, 16, and 32 times; Step 5: Add the SimVSS module to the hybrid encoder of the RT-DETR benchmark model, and feed the above 8-, 16-, and 32-fold multi-scale feature maps {S3, S4, S5} into the hybrid encoder based on the SimVSS module to extract multi-scale features, and then perform cross-scale fusion of the multi-scale features. Step 6: Input the features output by the hybrid encoder into the IoU-aware query selection of RT-DETR to select the area with high overlap with the real target as the initial candidate box; In step 7, RT-DETR's Transformer decoder performs multiple rounds of optimization on the selected candidate boxes, analyzes the relationship between the objects through the self-attention mechanism, gradually adjusts the position and size of the candidate boxes, and finally outputs accurate detection results; Step 8: In the RT-DETR model, in order to accurately locate the slender ship targets in the range compression domain, a shape-aware IoU regression positioning loss is added to the auxiliary prediction head structure.

2. The SAR ship detection method based on range compression domain characteristic guidance according to claim 1 is characterized in that: The strip attention mechanism module described in step 4 specifically includes the following steps: Step 4.1, define the strip attention mechanism module: given the input feature F∈R C×H×W , first perform three 1×1 convolution operations on it, and get and V∈R C×H×W These three feature maps; perform matrix dimension transformation and transposition operation on the feature map Q to generate N is H×W, and then the StripPooling operation is applied to the feature map K, that is, the average value of each row of pixels in the feature map is calculated to generate Then, the feature map Perform product operation with feature map Q, and finally obtain the position attention mapping result A∈R through Softmax operation N×W ; In essence, the feature map and The matrix multiplication operation is the product of the eigenvector corresponding to the position and the eigenvectors of all other positions, so the similarity calculation formula between positions is: Among them, Q i Represents the i-th position element of the feature map Q; K j Represents the j-th position element of feature map K, A j，i Represents Q i With K j the degree of correlation between them; Furthermore, the feature map V is stripped and pooled horizontally, and then the dimension is converted to After matrix multiplication with matrix A, the dimension is converted to Final feature map Add the corresponding elements of the input feature F to get the final feature map Step 4.2, embedding position of the SAM module: In the BasicBlock or Bottleneck structure of ResNet, insert the SAM module after the last convolutional layer of the residual path and perform channel weighting operation before adding the skip connection; Step 4.3, multi-level SAM module configuration: Add a SAM module at the output of each stage of ResNet to form a global-local joint attention mechanism.

3. The SAR ship detection method based on range compression domain characteristic guidance according to claim 1 is characterized in that: The SimVSS module described in step 5 specifically includes the following steps: Step 5.1, define the SimVSS module: its design is mainly based on SSM and a feedforward network; the structure contains residual connections and normalization layers, which are designed to stabilize the gradient training process and accelerate model convergence; the traditional SSM is a linear time-invariant system function, which is implemented through intermediate variables The univariate series Mapping to output sequence Specifically, given the state transition matrix As the evolution factor, weight matrix and observation matrix B, P∈C N As projection factors, and defined skip connections Its mathematical expression is as follows: In addition, the system function is implemented by introducing a time scale parameter To discretize it to process discrete time series data; this transformation is defined as follows: Among them, B, represents the mapping from state space to output space; in order to refine the approximation of B using a first-order Taylor series expansion: For input The processing steps within the SimVSS block are described as follows: Among them, Z l-1 、Z l 、Z l+1 They represent the output states of different layers l of the SimVSS block; In step 5.2, the SimVSS module is embedded in the middle of the attention-based intra-scale feature interaction and the convolutional neural network-based cross-scale feature fusion.

4. The SAR ship detection method based on range compression domain characteristic guidance according to claim 1 is characterized in that: The shape-aware IoU regression loss described in step 8 is specifically: The shape-aware IoU regression loss calculates the loss by focusing on the shape and scale of the bounding box itself, thereby enabling more accurate bounding box regression. Its calculation formula is as follows: Among them, scale is the scale factor, B and B gt The predicted box and the true value box, w and w respectively gt They are the predicted box B and the true value box B gt The width, h and h gt They are the predicted box B and the true value box B gt The height of , ww and hh are the weight coefficients in the horizontal and vertical directions respectively.

Images