SAR image ship detection method based on multi-scale feature collaborative modeling

Abstract

This invention discloses a SAR image ship detection method based on multi-scale feature collaborative modeling, belonging to the fields of smart ocean and computer vision technology. The method first constructs a SAR image dataset, then builds a detection model consisting of a hierarchical global structure modeling backbone network, a saliency-based structured gated feature pyramid network, and a decoupled rotating box structured regression loss module. The backbone network extracts and enhances multi-scale features hierarchically through multiple types of coding units. The feature pyramid network combines saliency priors, background noise reduction, and spatial gating to obtain multi-scale fused detection features. The regression loss module decouples the rotating box parameters and introduces angle correction and scale adaptation constraints. After training the optimal model on the dataset, the input SAR image to be detected outputs the ship's rotating box. This invention effectively solves problems such as feature fragmentation, background interference accumulation, and unstable rotating box regression in complex scenes, and can be applied to SAR image ship detection tasks.

Description

Technical Field

[0001] This invention belongs to the fields of smart ocean and computer vision technology, and specifically relates to a SAR image ship detection method based on multi-scale feature collaborative modeling. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an all-weather, all-day active remote sensing imaging method, can stably acquire sea surface observation data under adverse conditions such as insufficient light, cloud cover, extreme weather, and complex sea states. It plays a crucial role in marine environmental monitoring, maritime situational awareness, and maritime security protection systems. Ship target detection technology based on SAR imagery directly serves key application scenarios such as maritime traffic supervision, fisheries law enforcement, offshore rescue, waterway navigation assurance, and maritime defense security management. It provides technical support for achieving refined governance of marine space and building a smart and digital ocean, and has significant practical implications for improving the efficiency of marine management and safeguarding maritime traffic safety and national maritime rights.

[0003] Constrained by the inherent characteristics of SAR coherent imaging mechanisms, marine SAR images generally suffer from problems such as high-intensity coherent speckle noise, significant differences in target backscattering characteristics, strong sea clutter interference, and obvious multipath propagation effects. This results in ship targets exhibiting typical characteristics in images, including a large range of target scales, a high proportion of weak targets, significant morphological distortion, blurred structural edges, and low target contrast. Furthermore, complex nearshore shoreline structures, irregular wave textures, dense target distribution, and island / reef obstruction further exacerbate the confusion between ship targets and the strongly scattering background, significantly increasing the difficulty of detection tasks.

[0004] Traditional ship detection methods primarily rely on manually designed features and prior knowledge. However, in complex marine environments, with high noise levels and multi-scale variations, they suffer from poor robustness, weak generalization ability, and insufficient detection accuracy, failing to meet the diverse scenarios and high-precision detection requirements. With the rapid development of deep learning technology, ship detection methods based on convolutional neural networks (CNNs), with their powerful feature self-learning and representation capabilities, have become the mainstream technology in the SAR ship detection field. However, for the unique challenges of SAR image ship detection tasks, such as the large target scale span, complex background interference, and arbitrary target distribution, existing deep learning detection methods still face significant technical bottlenecks, making it difficult to achieve stable detection with high accuracy and robustness.

[0005] Existing multi-scale target detection methods generally presuppose a fixed pyramid hierarchical structure for feature organization and only make local improvements to feature fusion strategies under this premise, without conducting systematic analysis and design of the organization patterns, interaction methods, and collaborative mechanisms of multi-scale features. In SAR ship detection tasks, due to significant differences in target scale and complex background interference, existing methods do not clearly distinguish whether features of different scales are used as intermediate results or as long-term parallel semantic carriers participating in global modeling. This results in rigid feature organization, a single information transmission path between scales, and insufficient semantic collaborative capabilities.

[0006] Meanwhile, existing methods largely rely on implicit empirical assumptions or fixed network structures to realize information interaction and collaboration among multi-scale features, lacking explicit modeling and structured design of inter-scale interaction mechanisms. In scenarios with drastic changes in target scale and complex background interference, inter-scale information flow is easily hindered, feature utilization efficiency is low, and cross-scale semantics are difficult to effectively align and fuse, severely restricting the overall expressive power of multi-scale features.

[0007] Furthermore, in the target localization stage, existing methods generally use a uniform loss function to regress and optimize the bounding box, which has insurmountable structural constraints: First, the error sensitivity and geometric properties of targets at different scales vary significantly, and a uniform metric cannot take into account the localization needs of targets of different sizes; Second, the rotation box regression includes regression variables with structural differences such as position, scale, and orientation, and a uniform loss function cannot achieve differentiated constraints, which can easily lead to unbalanced optimization and weakening of some variables; Third, uniform regression under large-scale spans can easily lead to unstable gradient responses, affecting model convergence and localization accuracy.

[0008] The aforementioned shortcomings manifest themselves in SAR ship detection tasks as follows: insufficient positioning accuracy for small-scale ship targets, difficulty in ensuring geometric consistency of large-scale targets, unstable regression of targets in arbitrary directions, and high false detection and false alarm rates in near-shore and strong clutter scenarios. Existing technologies cannot simultaneously achieve fine-grained feature extraction, multi-scale feature collaborative modeling, background interference suppression, and structured regression optimization, making it difficult to meet the high-precision ship detection requirements in complex marine scenarios. Therefore, constructing a SAR image ship detection method with fine-grained feature representation, strong semantic association modeling, effective fusion of prior knowledge, multi-scale feature collaborative interaction, and structured regression constraints has become an urgent technical problem to be solved in this field, possessing significant theoretical research value and engineering application value. Summary of the Invention

[0009] To address the shortcomings of existing SAR image ship detection methods, such as susceptibility to sea clutter and shoreline background interference, poor multi-scale feature fusion performance, and insufficient accuracy in rotated bounding box regression, this invention proposes a detection method based on multi-scale feature collaborative modeling. This method sequentially completes feature extraction, fusion, and regression optimization through a hierarchical global structure modeling backbone network, a salient structured gated feature pyramid network, and a decoupled rotated bounding box structured regression loss module. This invention effectively suppresses background noise propagation, enhances multi-scale feature representation capabilities, improves the localization effect of rotated targets, and significantly improves the overall performance of SAR ship detection in complex scenes.

[0010] The technical solution adopted in this invention: A SAR image ship detection method based on multi-scale feature collaborative modeling, the specific process of the method of the present invention is as follows: S1. Collect SAR images and construct a SAR image dataset; extract features from the SAR images to obtain the original input features; S2. Construct a SAR image ship detection model based on multi-scale feature collaborative modeling. The model includes a hierarchical global structure modeling backbone network module, a salient structured gated feature pyramid network module, and a decoupled rotating box structured regression loss module. The hierarchical global structure modeling backbone network module, based on the original input features, includes four types of phased hierarchical units: a lightweight spatial-channel co-coding unit, a direction-aware structure modeling unit, a nearshore scene-aware background suppression unit, and a globally consistent guided feature recalibration unit. It completes high-dimensional feature extraction and structure modeling in stages. Specifically, the lightweight spatial-channel co-coding unit performs channel convolution operations on the original input features to achieve local texture layering encoding; the direction-aware structure modeling unit extracts horizontal and vertical structural response features based on anisotropic convolution kernels to achieve directional structure layering encoding; and the nearshore scene-aware background suppression unit obtains background similarity weights through global context information and weakens the interference of strong scattering background in the nearshore region on feature expression, achieving nearshore background layering. The encoding process involves: a global consistency-guided feature recalibration unit generating a global semantic weight self-vector and adaptively enhancing it to achieve hierarchical global semantic encoding; sequentially processing the output features of four types of units through staged hierarchical serial processing, then uniformly processing them to generate four sets of multi-scale features with different resolutions and outputting them; the saliency structured gating feature pyramid network module takes the multi-scale features as input, and through saliency decomposition, background noise reduction, spatial gating control, and cross-scale saliency propagation, achieves foreground feature enhancement and background interference suppression, outputting multi-scale fused detection features; the decoupled rotating box structured regression loss module takes the multi-scale fused detection features as input, performs structured decoupling modeling and scale adaptive constraints on the center, size, and angle terms of the rotating box, and outputs the ship's rotating box position; S3. Train the model by inputting the sample data from the dataset into the model, calculating the total loss function value, performing backpropagation, optimizing the connection weights through the selected optimizer and corresponding parameters, and obtaining the final SAR image ship detection model after multiple training rounds. S4. Based on the trained SAR image ship detection model, input the SAR image to be detected and output the ship rotation box position.

[0011] Preferably, in step S2, the hierarchical global structure modeling backbone network module adopts a phased hierarchical structure, including a lightweight spatial-channel co-coding unit, a direction-aware structure modeling unit, a nearshore scene-aware background suppression unit, and a globally consistent guided feature recalibration unit. The lightweight spatial-channel co-coding unit processes the original input features. The output features of the encoding unit are obtained by combining partial channel spatial convolution with channel dimension; specifically as follows: First, its channel dimension is divided into a subset of spatial convolution-enhanced channels and a subset of preserved channels, calculated using the following formula:

[0012] In the formula, This indicates the proportion of channels participating in spatial convolution. This represents the subset of spatially enhanced convolution channels that participate in spatial convolution. This represents the subset of channels that are not involved in spatial convolution; Next, regarding Applying a 3×3 spatial convolution yields locally enhanced features, while Keeping the same, splicing and blending yields the fused features. The calculation formula is as follows: ,

[0013] In the formula, This represents a 3×3 spatial convolution operation; Subsequently, the fused features are non-linearly mapped along the channel dimension through pointwise convolution to the local enhanced features. The calculation formula is as follows: , In the formula, Each represents a learnable weight parameter of a 1×1 convolution. This represents the Sigmoid activation function. This represents the mapping characteristics after channel nonlinear mapping; Finally, a channel recalibration mechanism is introduced to adaptively adjust the response of each channel. The calculation formula is as follows:

[0014] In the formula, Indicates global average pooling. and Both represent fully connected layers. Represents the ReLU activation function. This represents the channel recalibration weight vector. This indicates channel-by-channel multiplication. This represents the output characteristics of the coding unit; The orientation-aware structural modeling unit is based on the output features of the encoding unit. Through multi-directional convolution extraction, channel splicing and fusion, and residual adaptive fusion, the output features of the modeling unit are obtained, as follows: Horizontal and vertical structural response features are extracted using anisotropic convolution kernels, and local structural response features are obtained by extracting local enhancement features, expressed as follows: ,

[0015] In the formula, For the length of the directional receptive field, The horizontal structural response characteristics, This refers to the structural response characteristics in the vertical direction. For local structural response characteristics, This is a horizontal convolution operation with a kernel size of [size missing]. , This is a convolution operation in the straight direction, with a kernel size of [size missing]. ; Will After concatenating the three data points along the channel dimension, channel fusion and dimensionality reduction are performed using 1×1 pointwise convolution to obtain the orientation enhancement feature. The calculation formula is:

[0016] In the formula, for Pointwise convolution operation; Learnable residual coefficients are introduced, and the orientation enhancement features are connected to the output features of the coding unit through residual connections. The output features of the modeling unit are obtained through fusion, and the calculation formula is as follows: ,

[0017] In the formula, These are learnable parameters, and their initial values ​​are small. The output features of the modeling unit; The nearshore scene perception background suppression unit is based on the output features of the modeling unit. The similarity between features and background patterns is estimated using global contextual information, and highly similar regions are suppressed to obtain the output features of the suppression unit. Specifically, spatial dimension compression is first performed through global pooling to generate scene description vectors. The calculation formula is: ,

[0018] Subsequently, the background similarity weights are obtained through two layers of nonlinear mapping. The calculation formula is: ,

[0019] The background similarity weights are extended to the same spatial dimension as the original input features to form a similarity map. Then, the original input features are processed element-wise. Background suppression is performed to obtain background suppression features. The calculation formula is: ,

[0020] Learnable residual coefficients are introduced, and the background suppression features are fused with the original input features through residual connections to obtain the output features of the suppression unit. The calculation formula is: , In the formula, These are learnable parameters;

[0021] The globally consistent guided feature recalibration unit is based on the output features of the suppression unit. By adaptively gating and dynamically enhancing the feature channel responses using global contextual statistics, the output features of the recalibration unit are obtained, as follows: The original input features are then subjected to global average pooling. Spatial dimensionality compression is performed to obtain a global statistical vector; subsequently, a global semantic weight vector is generated through nonlinear mapping of a two-layer fully connected network. Its calculation formula is expressed as: , In the formula, and All indicate Learnable weight parameters for convolution;

[0022] The global semantic weight vector is expanded to the same spatial dimension as the original input features, and the original input features are then adaptively enhanced at the channel level through element-wise multiplication to obtain the recalibrated features. The calculation formula is:

[0023] Learnable residual coefficients are introduced, and the recalibration features are connected to the output features of the suppression unit through residual connections. The output characteristics of the recalibration unit are obtained by fusion. The calculation formula is: , In the formula, These are learnable parameters;

[0024] Will Dimensional normalization, channel unification, and scale filtering are performed to generate four sets of multi-scale features with different resolutions. As the final output data of the hierarchical global structure modeling backbone network module, it is input into the salient structured gated feature pyramid network module.

[0025] Preferably, in step S2, the saliency structured gated feature pyramid network module outputs multi-scale fusion detection features, as follows: Sa1, based on multi-scale features ,in In the formula For batch size, For the first Number of layer feature channels For the first Number of native channels in the scale feature layer The first Height and width of layer feature maps, scale index Spatial resolution decreases with increasing scale; firstly, for each layer Perform a horizontal channel alignment operation to obtain horizontal features. ,in ,and To unify the number of output channels, the lateral features are then processed using the Denoise operator for saliency decomposition and background noise reduction to obtain the saliency features of each layer. With denoising lateral features ; Sa2, for any adjacent scale and First of all Perform upsampling to make its resolution the same as the first By maintaining consistency across layers, upsampled lateral features are obtained. Simultaneously, saliency features Upsampling yields upsampling significance features Then, saliency propagation is enabled to upsample saliency features. With the Layer saliency features Weighted fusion, after activation function normalization, forms a consistent significance prior across scales. ; Sa3, based on the first Layer denoising lateral features Spatial gating features are generated through lightweight convolution and Sigmoid normalization. ; Use significant priors By applying weighted constraints to the spatial gating features, the optimized gating features are obtained. Subsequently, high-level semantic features are selectively injected using a gated residual fusion method to complete the gated residual fusion, resulting in the fused product. After fusion is complete, the Denoise operator is called again to process the fused product. Perform further denoising and recombination, and reuse the saliency prior after propagation. As a structured prior, saliency decomposition and background noise suppression are performed to obtain the final fusion features after secondary denoising. This allows for timely suppression of potential background interference after each cross-scale injection, enabling the fusion link to have a closed-loop steady-state characteristic of "injection-restructuring-reinjection". Sa4. Apply local convolutional smoothing to each layer of denoised and fused features to obtain multi-scale fused detection features. .

[0026] Preferably, in Sa1, the Denoise operator obtains saliency features. With denoising lateral features The calculation process is as follows: Significance features Acquisition: Based on lateral features Significant features are obtained through lightweight convolution and Sigmoid normalization. This is used to indicate the confidence level that a spatial location belongs to a candidate target region at this scale; Denoising Lateral Features Acquisition: Based on salient features The lateral features are structurally decomposed into foreground and background components. The foreground component is obtained by element-wise multiplication of the salient features and the lateral features. The background component is obtained by element-wise multiplication of the inverse saliency with the lateral features to emphasize the response of salient regions. This is used to aggregate textures and scattering patterns in insignificant regions; then a light noise-suppressing transformation is applied to the background component. and the foreground components The noise-suppressed background features are weighted and fused with the noise-suppressed background features according to a set controllable intensity to form denoised lateral features. In the formula This is the hyperparameter for background suppression intensity.

[0027] Preferably, in step S2, the specific steps of the decoupled rotating frame structured regression loss module are as follows: Sb1, based on multi-scale fusion detection features The bounding box is passed through a weighted rotation detector head, which consists of classification and regression branches. Each branch contains four 3x3 convolutional layers and a ReLU activation function. Finally, a 1x1 convolution is used to output the predicted bounding box parameters, which are represented in five-dimensional parameter form. ,in Represents the translation coordinates of the target center. Indicates the width and height of the target dimensions. The main direction of the target angle is represented; the five-dimensional parameters are completely decoupled into three independent regression subtasks, and structured regression subtasks for the three types of attributes of center, size and angle are constructed respectively; Sb2, Introduce an angle wrapper operator to normalize the angle error to... Within the range; Sb3, Smooth L1 form with adaptive design scale; based on geometric scale As a scale metric, and to construct scale ratios ,in For reference scale, This is a truncation function. The scale ratio is the upper and lower limit; an adaptive smoothing interval is dynamically generated based on the scale ratio. In the formula The preset basic smoothing interval; Sb4, employs adaptive smoothing intervals Calculating the five-dimensional loss vector using the scale-adaptive smoothing L1 loss function. Its expression is:

[0028] in, Represents the regression residuals, including the central residuals. Dimensional residuals and angular residual ; when When the value is small, the quadratic interval shrinks, and the loss function is closer to the L1 form near the zero point, with a steep gradient at the zero point, thus imposing a stronger and more precise constraint on small targets; when... When the value is large, the second interval expansion keeps the loss smooth within a large error range, which is beneficial for stable optimization of large targets or the initial regression stage. Finally, based on the five-dimensional loss vector, backpropagation optimization is used to gradually bring the predicted rotation box parameters closer to the actual ship target rotation box parameters, and the detection box position of the ship target is finally output.

[0029] Compared with existing technologies, this invention proposes a SAR image ship detection method based on multi-scale feature collaborative modeling. The advantages of this method are: (1) The present invention adopts a hierarchical global structure to model the backbone network. In the feature extraction stage, differentiated functional units and cross-scale semantic interaction mechanisms are introduced to enhance the consistency of the expression of features at different resolutions. It can stably represent the features of small-scale, weak-contrast ship targets and improve the robustness of multi-scale target feature extraction.

[0030] (2) This invention is equipped with a saliency structured gating feature pyramid network, which completes foreground and background feature decomposition and background noise suppression based on saliency prior. It combines spatial gating and cross-scale saliency propagation mechanism to achieve adaptive semantic injection, which can accurately block the cumulative propagation of sea clutter and shoreline interference in the fusion link and optimize the multi-scale feature fusion effect.

[0031] (3) The present invention designs a decoupled rotating frame structured regression loss module, which decouples the rotation frame center, scale and angle parameters into a structured and decoupled calculation loss. Combined with angle interval correction and scale adaptive smoothing constraint, it effectively eliminates gradient oscillation caused by angle periodicity and target aspect ratio difference, and greatly improves the positioning accuracy and geometric consistency of the rotating frame.

[0032] (4) The modules of this invention have clear division of labor and coordinated functions. The overall model structure is highly interpretable and has excellent comprehensive detection performance in complex marine SAR scenarios. It not only achieves high-precision detection of ship targets, but also provides feasible technical reference for remote sensing target detection research in extreme sea conditions and dense ship scenarios. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the SAR image ship detection framework of the present invention; Figure 3 This is a diagram of the salient structured gated feature pyramid network module architecture of the present invention; Figure 4 This is a diagram of the architecture of the decoupled rotating frame structured regression loss module of the present invention; Figure 5 This is a structural diagram of the orientation-sensing structure modeling unit of the present invention; Figure 6 The adaptive smoothing interval of the present invention Schematic diagram of the adjustment of changes; Figure 7 The adaptive smoothing interval in the Smooth L1 function of this invention A schematic diagram showing the changes and adjustments. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be further described clearly and completely below with reference to the accompanying drawings. It should be noted that the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0035] To make the inventive objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be further described in detail below with reference to the accompanying drawings: In order to better understand the above-mentioned objectives, features, and advantages of this invention, the advantages of this invention will be further illustrated below by comparing the embodiments with the accompanying drawings and specific implementation methods.

[0036] This invention proposes a SAR image ship detection method based on multi-scale feature collaborative modeling, such as... Figure 1 The process of this method is described in detail below: S1. Collect SAR images and construct a SAR image dataset; extract features from the SAR images to obtain the original input features; the feature extraction involves performing radiometric correction, logarithmic transformation and normalization on the SAR images, and copying the single-channel grayscale image into a three-channel image to obtain the original input features; S2. Construct a SAR image ship detection model based on multi-scale feature collaborative modeling. The model includes a hierarchical global structure modeling backbone network module, a salient structured gated feature pyramid network module, and a decoupled rotating box structured regression loss module. The hierarchical global structure modeling backbone network module, based on the original input features, includes four types of phased hierarchical units: a lightweight spatial-channel co-coding unit, a direction-aware structure modeling unit, a nearshore scene-aware background suppression unit, and a globally consistent guided feature recalibration unit. It completes high-dimensional feature extraction and structure modeling in stages. Specifically, the lightweight spatial-channel co-coding unit performs channel convolution operations on the original input features to achieve local texture layering encoding; the direction-aware structure modeling unit extracts horizontal and vertical structural response features based on anisotropic convolution kernels to achieve directional structure layering encoding; and the nearshore scene-aware background suppression unit obtains background similarity weights through global context information and weakens the interference of strong scattering background in the nearshore region on feature expression, achieving nearshore background layering. Encoding; the global consistency-guided feature recalibration unit generates a global semantic weight self-vector and performs adaptive enhancement to achieve global semantic hierarchical encoding; the output features of the four types of units are obtained through staged hierarchical serial processing and then uniformly processed to generate four sets of multi-scale features with different resolutions and output them; the saliency structured gating feature pyramid network module takes the multi-scale features as input, and through saliency decomposition, background noise suppression, spatial gating control and cross-scale saliency propagation, it achieves foreground feature enhancement and background interference suppression, and outputs multi-scale fused detection features; the decoupled rotating box structured regression loss module takes the multi-scale fused detection features as input, performs structured decoupling modeling and scale adaptive constraints on the center, size and angle terms of the rotating box, and outputs the position of the ship's rotating box, such as Figure 2-4 As shown; Specifically, in step S2, the hierarchical global structure modeling backbone network module adopts a phased hierarchical structure, including a lightweight spatial-channel co-coding unit (LSCE Block), a direction-aware structure modeling unit (DASM Block), a nearshore scene-aware background suppression unit (NSAS Block), and a globally consistent guided feature recalibration unit (GCGR Block). The Lightweight Spatial-Channel Encoding Block (LSCE Block) serves as the basic building block; it processes the original input features... By combining partial channel spatial convolution with channel dimension, the output features of the encoding unit are obtained, achieving collaborative modeling of local structural information and channel dependencies. Specifically, the channel dimension is first divided into a spatial convolution-enhanced channel subset and a preserved channel subset, calculated using the following formula:

[0037] In the formula, This indicates the proportion of channels participating in spatial convolution. This represents the subset of spatially enhanced convolution channels that participate in spatial convolution. This represents the subset of channels that are not involved in spatial convolution; Next, regarding Applying a 3×3 spatial convolution yields locally enhanced features, while Keeping the same, splicing and blending yields the fused features. The calculation formula is as follows:

[0038] In the formula, This represents a 3×3 spatial convolution operation; Subsequently, the fused features are non-linearly mapped along the channel dimension through pointwise convolution to the local enhanced features. The calculation formula is as follows:

[0039] In the formula, Each represents a learnable weight parameter of a 1×1 convolution. This represents the Sigmoid activation function. This represents the mapping characteristics after channel nonlinear mapping; Finally, a channel recalibration mechanism is introduced to adaptively adjust the response of each channel. The calculation formula is as follows:

[0040] In the formula, Indicates global average pooling. Both represent fully connected layers (or pointwise convolutional layers). Represents the ReLU activation function. This represents the channel recalibration weight vector. This indicates channel-by-channel multiplication. This represents the output characteristics of the coding unit; The lightweight spatial-channel co-coding unit effectively retains sensitivity to local scattering structures and edge contours while controlling computational complexity, providing stable underlying feature support for subsequent global modeling. The Directional-Aware Structural Modeling Block (DASM Block), such as Figure 5 As shown, based on the output features of the coding unit Through multi-directional convolution extraction, channel splicing and fusion, and residual adaptive fusion, the output features of the modeling unit are obtained, as follows: Horizontal and vertical structural response features are extracted using anisotropic convolution kernels, and local structural response features are obtained by extracting local enhancement features, expressed as follows: ,

[0041] In the formula, For the length of the directional receptive field, The horizontal structural response characteristics, This refers to the structural response characteristics in the vertical direction. For local structural response characteristics, This is a horizontal convolution operation with a kernel size of [size missing]. , This is a convolution operation in the straight direction, with a kernel size of [size missing]. ; Will After concatenating the three data points along the channel dimension, channel fusion and dimensionality reduction are performed using 1×1 pointwise convolution to obtain the orientation enhancement feature. The calculation formula is:

[0042] In the formula, for Pointwise convolution operation; Considering that excessively strong directional constraints may affect the model's generalization ability, learnable residual coefficients are introduced. These residual connections link the directional enhancement features with the output features of the encoding unit. The output features of the modeling unit are obtained through fusion, and the calculation formula is as follows: ,

[0043] In the formula, These are learnable parameters, and their initial values ​​are small. The output features of the modeling unit; This allows the model to maintain a stable mapping in the early stages of training, while gradually enhancing its directional modeling capabilities as training progresses. DASM Block explicitly decomposes the spatial structure through horizontal and vertical convolutions, enabling the network to better distinguish between ship targets and background interference with similar local textures but different directional distributions. The Nearshore Scene-Aware Suppression Block (NSAS Block) is based on the output features of the modeling unit. The similarity between features and background patterns is estimated using global contextual information, and highly similar regions are suppressed to obtain the output features of the suppression unit. Specifically, spatial dimension compression is first performed through global pooling to generate scene description vectors. The calculation formula is: ,

[0044] Subsequently, the background similarity weights are obtained through two layers of nonlinear mapping. The calculation formula is:

[0045] In the formula, and All indicate Learnable weight parameters for convolution;

[0046] The background similarity weights are extended to the same spatial dimension as the original input features to form a similarity map. Then, the original input features are processed element-wise. Background suppression is performed to obtain background suppression features. The calculation formula is: ,

[0047] Learnable residual coefficients are introduced, and the background suppression features are fused with the original input features through residual connections to obtain the output features of the suppression unit. The calculation formula is: , In the formula, These are learnable parameters;

[0048] This design enables the network to effectively reduce the interference of strong scattering background in the nearshore region on feature representation while maintaining the target response; The Global Consistency-Guided Recalibration Block (GCGR Block) is based on the output features of the suppression unit. By adaptively gating and dynamically enhancing the feature channel responses using global contextual statistics, the output features of the recalibration unit are obtained, thereby strengthening discriminative features and suppressing redundant responses. Specifically, global average pooling is used to enhance the original input features. Spatial dimensionality compression is performed to obtain a global statistical vector; subsequently, a global semantic weight vector is generated through nonlinear mapping of a two-layer fully connected network. Its calculation formula is expressed as: , In the formula, and Each represents a learnable weight parameter of a 1×1 convolution;

[0049] The global semantic weight vector is expanded to the same spatial dimension (H×W×C) as the original input features, and the original input features are adaptively enhanced at the channel level through element-wise multiplication to obtain the recalibrated features. The calculation formula is:

[0050] Learnable residual coefficients are introduced, and the recalibration features are connected to the output features of the suppression unit through residual connections. The output characteristics of the recalibration unit are obtained by fusion. The calculation formula is: , In the formula, These are learnable parameters;

[0051] This allows the model to maintain a near-identical mapping in the early stages of training, thus avoiding feature degradation due to excessive gating. As training progresses, the model gradually learns more reasonable global semantic weights, thereby enhancing discriminative features and suppressing redundant responses. Will Dimensional normalization, channel unification, and scale filtering are performed to generate four sets of multi-scale features with different resolutions. As the final output data of the hierarchical global structure modeling backbone network module, it is input into the salient structured gated feature pyramid network module, providing all the original multi-scale feature support for cross-scale feature fusion, and realizing seamless data connection between the backbone network and the feature pyramid module. HGSM-Net achieves a gradual transition from local texture modeling to global structural representation by selectively introducing LSCE Block, DASM Block, NSAS Block, and GCGR Block at different levels. The shallow layer mainly uses LSCE Block to ensure the stability of low-level features; the middle layer introduces DASM Block to strengthen directional structural modeling; and the deep layer combines NSAS Block and GCGR Block to suppress complex backgrounds and improve semantic consistency. Through this hierarchical structural design, the network can simultaneously model local details and global contextual information in multi-scale feature spaces, thereby significantly improving the robustness and accuracy of SAR ship target detection in complex marine scenarios.

[0052] Specifically, in step S2, the salient structured gated feature pyramid network module outputs multi-scale fused detection features, as follows: The saliency-structured gated feature pyramid network module explicitly predicts spatial saliency priors at each scale, decomposes features into foreground and background components, suppresses noise in the background component, and then regulates the top-down semantic injection intensity using spatial gating. Simultaneously, it propagates saliency priors across scales to form consistent candidate region constraints across scales. The specific steps are as follows: Sa1, let the multi-scale features output by the backbone network module of the hierarchical global structure modeling be... ,in In the formula For batch size, For the first Number of layer feature channels For the first Number of native channels in the scale feature layer The first Height and width of layer feature maps, scale index Spatial resolution decreases with increasing scale; firstly, for each layer Perform a horizontal channel alignment operation to obtain horizontal features. ,in ,and To standardize the number of output channels; then, unlike traditional methods, Instead of directly proceeding to cross-scale fusion, it first undergoes saliency decomposition and background denoising to construct "denoised lateral features." These lateral features are then processed through the Denoise operator for saliency decomposition and background denoising to obtain the saliency features of each layer. With denoising lateral features This achieves initial suppression of sea clutter and strong scattering interference from the shoreline; Sa2, based on saliency features With denoising lateral features It performs top-down cross-scale fusion, but the fusion method is no longer to unconditionally inject upper-layer features into lower-layer features. Instead, it introduces spatial gating features to control the injection amount positionally, and uses saliency as a propagable cross-scale prior to participate in the gating construction; for any adjacent scale... and First of all Perform upsampling to make its resolution the same as the first By maintaining consistency across layers, upsampled lateral features are obtained. Simultaneously, saliency features Upsampling yields upsampling significance features Then, saliency propagation is enabled to upsample saliency features. With the Layer saliency features Weighted fusion, after activation function normalization, forms a consistent significance prior across scales. The significance of this propagation mechanism is that the more stable and context-constrained candidate regions formed at higher scales will constrain the saliency of lower scales from top to bottom, so that the lower scale will not produce fragmented and structurally inconsistent responses when the details are rich but the noise is stronger, thereby improving the stability and consistency of small targets across scales. Sa3, based on the first Layer denoising lateral features Spatial gating features are generated through lightweight convolution and Sigmoid normalization. ; Use significant priors By applying weighted constraints to the spatial gating features, the optimized gating features are obtained. This allows semantic injection to focus on spatial regions more likely to contain the target; subsequently, a gated residual fusion method is used to selectively inject high-level semantic features, completing the gated residual fusion and obtaining the fused result. This gated fusion mechanism achieves "noise-aware semantic injection": when a location is saliency-judged as background or gated as low confidence, upper-layer semantics are suppressed from injection, thus blocking the cross-scale propagation of background semantic pollution. When a location is saliency and the gate is large, upper-layer semantics are selectively injected to enhance the semantic consistency of the target region. It is important to note that the network does not output directly after fusion, but rather calls the Denoise operator again to process the fused data. Perform further denoising and recombination, and reuse the saliency prior after propagation. As a structured prior, saliency decomposition and background denoising are performed to obtain the final fusion features after secondary denoising. This allows for timely suppression of potential background interference after each cross-scale injection, enabling the fusion link to have a closed-loop steady-state characteristic of "injection-restructuring-reinjection". This is particularly crucial for the problem of consistent background patterns across scales and easy accumulation and amplification in SAR scenarios. Sa4. Apply local convolutional smoothing to each layer of denoised and fused features to obtain multi-scale fused detection features. To further integrate neighborhood information and alleviate the local discontinuities that may be caused by upsampling and point-by-point gating; if the detection framework requires additional lower resolution layers, the network generates additional layers according to general practices: pooling is used to obtain coarser-scale features when no additional convolution is introduced, and new layers are generated step by step through stride convolution when additional convolution is enabled, and the source of the additional layers is selected from the input features, the highest layer fused features or the output features, thereby ensuring compatibility with the existing training configuration.

[0053] More specifically, in Sa1, the Denoise operator obtains saliency features. With denoising lateral features The calculation process is as follows: Significance features Acquisition: The Denoise operator is a key step in suppressing semantic pollution in the fusion link, based on lateral features. Significant features are obtained through lightweight convolution and Sigmoid normalization. This is used to indicate the confidence level that a spatial location belongs to a candidate target region at this scale; Denoising Lateral Features Acquisition: Based on salient features The lateral features are structurally decomposed into foreground and background components. The foreground component is obtained by element-wise multiplication of the salient features and the lateral features. The background component is obtained by element-wise multiplication of the inverse saliency with the lateral features to emphasize the response of salient regions. This is used to aggregate textures and scattering patterns in insignificant regions; then a light noise-suppressing transformation is applied to the background component. (Constituted by local convolutions with normalization and activation), and the foreground components The noise-suppressed background features are weighted and fused with the noise-suppressed background features according to a set controllable intensity to form denoised lateral features. In the formula The background suppression intensity hyperparameter is used to represent the feature of the salient region. The intuitive meaning of this design is that the features of the salient region are preserved as much as possible to avoid weakening the weak target, while the non-salient region is not simply discarded. Instead, it is first locally smoothed / reorganized to suppress the interference of high-frequency clutter texture and strong scattering block, and then back-injected with a lower weight. This reduces the "poisoning" effect of the background on the subsequent fusion while maintaining the continuity of the overall context. Since the process decomposes and reorganizes based on saliency as a structured prior, its effect does not rely on simple convolutional denoising. Instead, it guides the denoising effect to focus on the background component through saliency guidance, thus avoiding excessive smoothing of the target component from a mechanistic perspective.

[0054] Specifically, in step S2, the decoupled rotating box structured regression loss module treats the rotating box regression problem as a structured regression task composed of three attributes: center, scale, and orientation. Based on this, it designs a regression loss form with decoupled parameters and adaptive scale. The specific steps are as follows: Sb1, based on multi-scale fusion detection features The bounding box is passed through a weighted rotation detector head, which consists of classification and regression branches. Each branch contains four 3x3 convolutional layers (256 channels) and a ReLU activation function. Finally, a 1x1 convolution is used to output the predicted bounding box parameters, which are represented in five-dimensional parameter form. ,in Represents the translation coordinates of the target center. Indicates the width and height of the target dimensions. The main direction of the target angle is represented; structurally, the five-dimensional parameters are completely decoupled into three independent regression subtasks, constructing structured regression subtasks for the three attributes of center, size, and angle, respectively, rather than simply splicing the five-dimensional parameters together and uniformly calculating the regression loss; this decoupling is not a formal split, but based on the essential differences in the optimization behavior of different geometric components: center regression focuses more on local translation accuracy, scale regression directly determines the target coverage, while angle regression has periodicity and discontinuity, and its error has a high impact on IoU that depends on the aspect ratio of the target; if the three are treated equally, the angle term often produces disproportionate gradient feedback on specific samples, thereby interfering with the stable convergence of other parameters; Sb2, Introduce an angle wrapper operator to normalize the angle error to... Within the range, thus avoiding Numerical jumps occur at the boundaries; this process ensures the geometric continuity of the angle error, making the loss function have consistent local smoothness in the direction space, which provides a premise for subsequent smooth regression constraints. Sb3, a scale-adaptive Smooth L1 design, is based on the idea that different scale targets have different tolerances for regression accuracy. Small targets often exhibit more significant localization errors due to minor deviations in pixel space, while large targets are more robust to the same absolute deviation. Using a fixed smoothing interval would result in overly soft regression constraints for small targets and overly hard constraints for large targets, leading to scale bias during the overall training process. Therefore, a geometric scale is used instead. As a scale metric, and to construct scale ratios ,in For reference scale, This is a truncation function. The scale ratio is defined as upper and lower limits to restrict unstable amplification caused by extreme scales; the adaptive smoothing interval is dynamically obtained based on the scale ratio. In the formula The preset basic smoothing range is used to achieve an adaptive adjustment mechanism where "the smaller the scale, the harder the constraint; the larger the scale, the smoother the constraint." like Figure 6 As shown, the horizontal and vertical axes represent the target width and height, respectively, and the vertical axis represents the distance from the target to the height. The scale ratio obtained by mapping can be intuitively observed that small-scale regions correspond to lower ratio values, while the ratio rises smoothly as the scale gradually increases and is eventually limited by the upper bound. This "continuous, monotonic and constrained" mapping relationship ensures the stability of the scale adaptation mechanism in the entire scale space, so that the loss function will not produce excessive gradient changes due to extremely small or large targets. At the same time, it clearly reflects "how geometric scale controls optimization behavior" from the perspective of combining numbers and shapes. Sb4, employs adaptive smoothing intervals Calculating the five-dimensional loss vector using the scale-adaptive smoothing L1 loss function. Its expression is:

[0055] in, Represents the regression residuals, including the central residuals. Dimensional residuals and angular residual ; like Figure 7 As shown, the optimization implications of this design are explained from a geometric perspective: when When the value is small, the quadratic interval shrinks, and the loss function is closer to the L1 form near the zero point, with a steep gradient at the zero point, thus imposing a stronger and more precise constraint on small targets; when... When the value is large, the second-order interval expansion keeps the loss smooth within a larger error range, which is beneficial for stable optimization of large objectives or the initial regression stage; by... By transforming the loss function from a fixed constant to a function that varies with the target scale, adaptive "geometric fairness" is achieved across samples of different scales. Finally, based on the five-dimensional loss vector, backpropagation optimization is used to gradually bring the predicted rotation box parameters closer to the actual ship target rotation box parameters, and the detection box position of the ship target is finally output.

[0056] S3. Train the model by inputting the sample data from the dataset into the model, calculating the total loss function value, performing backpropagation, optimizing the connection weights through the selected optimizer and corresponding parameters, and obtaining the final SAR image ship detection model after multiple training rounds. S4. Based on the trained SAR image ship detection model, input the SAR image to be detected and output the ship rotation box position.

[0057] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0058] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A SAR image ship detection method based on multi-scale feature collaborative modeling, characterized in that, The specific process is as follows: S1. Collect SAR images and construct a SAR image dataset; extract features from the SAR images to obtain the original input features; S2. Construct a SAR image ship detection model based on multi-scale feature collaborative modeling. The model includes a hierarchical global structure modeling backbone network module, a salient structured gated feature pyramid network module, and a decoupled rotating box structured regression loss module. The hierarchical global structure modeling backbone network module, based on the original input features, includes four types of phased hierarchical units: a lightweight spatial-channel co-coding unit, a direction-aware structure modeling unit, a nearshore scene-aware background suppression unit, and a globally consistent guided feature recalibration unit. It completes high-dimensional feature extraction and structure modeling in stages. Specifically, the lightweight spatial-channel co-coding unit performs channel convolution operations on the original input features to achieve local texture layering encoding; the direction-aware structure modeling unit extracts horizontal and vertical structural response features based on anisotropic convolution kernels to achieve directional structure layering encoding; and the nearshore scene-aware background suppression unit obtains background similarity weights through global context information and weakens the interference of strong scattering background in the nearshore region on feature expression, achieving nearshore background layering. The encoding process involves: a global consistency-guided feature recalibration unit generating a global semantic weight self-vector and adaptively enhancing it to achieve hierarchical global semantic encoding; sequentially processing the output features of four types of units through staged hierarchical serial processing, then uniformly processing them to generate four sets of multi-scale features with different resolutions and outputting them; the saliency structured gating feature pyramid network module takes the multi-scale features as input, and through saliency decomposition, background noise reduction, spatial gating control, and cross-scale saliency propagation, achieves foreground feature enhancement and background interference suppression, outputting multi-scale fused detection features; the decoupled rotating box structured regression loss module takes the multi-scale fused detection features as input, performs structured decoupling modeling and scale adaptive constraints on the center, size, and angle terms of the rotating box, and outputs the ship's rotating box position; S3. Train the model by inputting the sample data from the dataset into the model, calculating the total loss function value, performing backpropagation, optimizing the connection weights through the selected optimizer and corresponding parameters, and obtaining the final SAR image ship detection model after multiple training rounds. S4. Based on the trained SAR image ship detection model, input the SAR image to be detected and output the ship rotation box position.

2. The SAR image ship detection method based on multi-scale feature collaborative modeling according to claim 1, characterized in that, In step S2, the hierarchical global structure modeling backbone network module adopts a phased hierarchical structure, including a lightweight spatial-channel co-coding unit, a direction-aware structure modeling unit, a nearshore scene-aware background suppression unit, and a globally consistent guided feature recalibration unit. The lightweight spatial-channel co-coding unit processes the original input features. The output features of the encoding unit are obtained by combining partial channel spatial convolution with channel dimension; specifically as follows: First, its channel dimension is divided into a subset of spatial convolution-enhanced channels and a subset of preserved channels, calculated using the following formula: ， In the formula, This indicates the proportion of channels participating in spatial convolution. This represents the subset of spatially enhanced convolution channels that participate in spatial convolution. This represents the subset of channels that are not involved in spatial convolution; Next, regarding Applying a 3×3 spatial convolution yields locally enhanced features, while Keeping the same, splicing and blending yields the fused features. The calculation formula is as follows: ， In the formula, This represents a 3×3 spatial convolution operation; Subsequently, the fused features are non-linearly mapped along the channel dimension through pointwise convolution to the local enhanced features. The calculation formula is as follows: ， In the formula, and Each represents a learnable weight parameter of a 1×1 convolution. This represents the Sigmoid activation function. This represents the mapping characteristics after channel nonlinear mapping; Finally, a channel recalibration mechanism is introduced to adaptively adjust the response of each channel. The calculation formula is as follows: ， ， In the formula, Indicates global average pooling. and Both represent fully connected layers. Represents the ReLU activation function. This represents the channel recalibration weight vector. This indicates channel-by-channel multiplication. This represents the output characteristics of the coding unit; The orientation-aware structural modeling unit is based on the output features of the encoding unit. Through multi-directional convolution extraction, channel splicing and fusion, and residual adaptive fusion, the output features of the modeling unit are obtained, as follows: Horizontal and vertical structural response features are extracted using anisotropic convolution kernels, and local structural response features are obtained by extracting local enhancement features, expressed as follows: ， In the formula, For the length of the directional receptive field, The horizontal structural response characteristics, This refers to the structural response characteristics in the vertical direction. For local structural response characteristics, This is a horizontal convolution operation with a kernel size of [size missing]. , This is a convolution operation in the straight direction, with a kernel size of [size missing]. ; Will and After concatenating the three data points along the channel dimension, channel fusion and dimensionality reduction are performed using 1×1 pointwise convolution to obtain the orientation enhancement feature. The calculation formula is: ， In the formula, for Pointwise convolution operation; Learnable residual coefficients are introduced, and the orientation enhancement features are connected to the output features of the coding unit through residual connections. By fusing, the output features of the modeling unit are obtained. The calculation formula is: ， In the formula, These are learnable parameters, and their initial values ​​are small; The nearshore scene perception background suppression unit is based on the output features of the modeling unit. The similarity between features and background patterns is estimated using global contextual information, and highly similar regions are suppressed to obtain the output features of the suppression unit. Specifically, spatial dimension compression is first performed through global pooling to generate scene description vectors. The calculation formula is: ， Subsequently, the background similarity weights are obtained through two layers of nonlinear mapping. The calculation formula is: ， The background similarity weights are extended to the same spatial dimension as the original input features to form a similarity map. Then, the original input features are processed element-wise. Background suppression is performed to obtain background suppression features. The calculation formula is: ， Learnable residual coefficients are introduced, and the background suppression features are fused with the original input features through residual connections to obtain the output features of the suppression unit. The calculation formula is: ， In the formula, These are learnable parameters; The globally consistent guided feature recalibration unit is based on the output features of the suppression unit. By adaptively gating and dynamically enhancing the feature channel responses using global contextual statistics, the output features of the recalibration unit are obtained, as follows: The original input features are then subjected to global average pooling. Spatial dimensionality compression is performed to obtain a global statistical vector; subsequently, a global semantic weight vector is generated through nonlinear mapping of a two-layer fully connected network. Its calculation formula is expressed as: ， In the formula, and All indicate Learnable weight parameters for convolution; The global semantic weight vector is expanded to the same spatial dimension as the original input features, and the original input features are then adaptively enhanced at the channel level through element-wise multiplication to obtain the recalibrated features. The calculation formula is: ， Learnable residual coefficients are introduced, and the recalibration features are connected to the output features of the suppression unit through residual connections. The output characteristics of the recalibration unit are obtained by fusion. The calculation formula is: ， In the formula, These are learnable parameters; Will Dimensional normalization, channel unification, and scale filtering are performed to generate four sets of multi-scale features with different resolutions. As the final output data of the hierarchical global structure modeling backbone network module, it is input into the salient structured gated feature pyramid network module.

3. The SAR image ship detection method based on multi-scale feature collaborative modeling according to claim 1, characterized in that, In step S2, the saliency structured gated feature pyramid network module outputs multi-scale fused detection features, as follows: Sa1, based on multi-scale features ,in In the formula For batch size, For the first Number of layer feature channels For the first Number of native channels in the scale feature layer The first Height and width of layer feature maps, scale index Spatial resolution decreases with increasing scale; firstly, for each layer Perform a horizontal channel alignment operation to obtain horizontal features. ,in ,and To unify the number of output channels, the lateral features are then processed using the Denoise operator for saliency decomposition and background noise reduction to obtain the saliency features of each layer. With denoising lateral features ; Sa2, for any adjacent scale and First of all Perform upsampling to make its resolution the same as the first By maintaining consistency across layers, upsampled lateral features are obtained. Simultaneously, saliency features Upsampling yields upsampling significance features Then, saliency propagation is enabled to upsample saliency features. With the Layer saliency features Weighted fusion, after activation function normalization, forms a consistent significance prior across scales. ; Sa3, based on the first Layer denoising lateral features Spatial gating features are generated through lightweight convolution and Sigmoid normalization. ; Use significant priors By applying weighted constraints to the spatial gating features, the optimized gating features are obtained. ; Subsequently, high-level semantic features are selectively injected using a gated residual fusion method to complete the gated residual fusion, resulting in the fused product. After fusion is complete, the Denoise operator is called again to process the fused product. Perform further denoising and recombination, and reuse the saliency prior after propagation. As a structured prior, saliency decomposition and background noise suppression are performed to obtain the final fusion features after secondary denoising. This allows for timely suppression of potential background interference after each cross-scale injection, enabling the fusion link to have a closed-loop steady-state characteristic of "injection-restructuring-reinjection". Sa4. Apply local convolutional smoothing to each layer of denoised and fused features to obtain multi-scale fused detection features. .

4. The SAR image ship detection method based on multi-scale feature collaborative modeling according to claim 3, characterized in that, In Sa1, the Denoise operator obtains saliency features. With denoising lateral features The calculation process is as follows: Significance features Acquisition: Based on lateral features Significant features are obtained through lightweight convolution and Sigmoid normalization. This is used to indicate the confidence level that a spatial location belongs to a candidate target region at this scale; Denoising Lateral Features Acquisition: Based on salient features The lateral features are structurally decomposed into foreground and background components. The foreground component is obtained by element-wise multiplication of the salient features and the lateral features. The background component is obtained by element-wise multiplication of the inverse saliency with the lateral features to emphasize the response of salient regions. This is used to aggregate textures and scattering patterns in insignificant regions; then a light noise-suppressing transformation is applied to the background component. and the foreground components The noise-suppressed background features are weighted and fused with the noise-suppressed background features according to a set controllable intensity to form denoised lateral features. In the formula This is the hyperparameter for background suppression intensity.

5. The SAR image ship detection method based on multi-scale feature collaborative modeling according to claim 1, characterized in that, In step S2, the decoupled rotating frame structured regression loss module obtains the position of the ship's rotating frame, as follows: Sb1, based on multi-scale fusion detection features The bounding box is passed through a weighted rotation detector head, which consists of classification and regression branches. Each branch contains four 3x3 convolutional layers and a ReLU activation function. Finally, a 1x1 convolution is used to output the predicted bounding box parameters, which are represented in five-dimensional parameter form. ,in Represents the translation coordinates of the target center. Indicates the width and height of the target dimensions. The main direction of the target angle is represented; the five-dimensional parameters are completely decoupled into three independent regression subtasks, and structured regression subtasks for the three types of attributes of center, size and angle are constructed respectively; Sb2, Introduce an angle wrapper operator to normalize the angle error to... Within the range; Sb3, Smooth L1 form with adaptive design scale; based on geometric scale As a scale metric, and to construct scale ratios ,in For reference scale, This is a truncation function. These are the upper and lower limits of the scale ratio; Dynamically generate adaptive smoothing intervals based on scale ratio In the formula The preset basic smoothing interval; Sb4, employs adaptive smoothing intervals Calculating the five-dimensional loss vector using the scale-adaptive smoothing L1 loss function. Its expression is: ， in, Represents the regression residuals, including the central residuals. Dimensional residuals and angular residual ; when When the value is small, the quadratic interval shrinks, and the loss function is closer to the L1 form near the zero point, with a steep gradient at the zero point, thus imposing a stronger and more precise constraint on small targets; when... When the value is large, the second interval expansion keeps the loss smooth within a large error range, which is beneficial for stable optimization of large targets or the initial regression stage. Finally, based on the five-dimensional loss vector, backpropagation optimization is used to gradually bring the predicted rotation box parameters closer to the actual ship target rotation box parameters, and the detection box position of the ship target is finally output.

Images