Rotating target detection method, device and equipment based on single-view complex SAR (Synthetic Aperture Radar) image

Abstract

The invention relates to a rotating target detection method, device and equipment based on a single-view complex SAR (Synthetic Aperture Radar) image. The method comprises the following steps: decoupling a single-view complex SAR image to obtain amplitude and phase components, constructing a corresponding input tensor after preprocessing, extracting features through an asymmetric double-current trunk, extracting amplitude multi-scale global semantic features through a hierarchical Transform architecture, extracting phase structure texture features under semantic constraint through a phase auxiliary branch formed by a lightweight convolution residual network, and obtaining a single-view complex SAR image. Based on a fusion thought of amplitude-oriented phase enhancement, an enhanced amplitude feature sequence is obtained through feature alignment and semantic hierarchy differential fusion, the enhanced amplitude feature sequence is recombined into a multi-scale feature pyramid, and a target category and a rotation bounding box are output through a rotation target detection branch. According to the method, double information is fully utilized, phase noise interference is avoided, detection precision and positioning accuracy of targets at different scale angles are improved, global semantic capture and model lightweight are considered, and engineering application feasibility is guaranteed.

Description

Technical Field

[0001] This application relates to the fields of radar signal processing and deep learning technology, and in particular to a method, apparatus and device for rotating target detection based on single-view complex SAR images. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging sensor capable of acquiring high-resolution surface observation data under all-weather, all-day conditions, and has been widely used in fields such as marine surveillance and disaster assessment. SAR target detection refers to the automatic detection and location of targets of interest such as vehicles, ships, and aircraft in a strong cluttered background, and is a key foundational technology in SAR image interpretation and intelligent processing.

[0003] With the development of deep learning, convolutional neural networks (CNNs) have replaced traditional statistical modeling methods such as constant false alarm rate (CFAR) and become the mainstream technology for SAR target detection. Most existing deep learning methods employ an "amplitude-only" processing paradigm, converting single-view complex (SLC) SAR data into amplitude / intensity images and discarding phase information. While this allows for the reuse of mature real-valued network structures and the extraction of macroscopic geometric and scattering intensity features of the target, it neglects the coherence characteristics of SAR imaging. SAR echoes are essentially complex data; amplitude reflects the macroscopic structure and scattering intensity of the target, while phase encodes subtle physical information such as target surface roughness and scattering center topology. Relying solely on amplitude information makes it difficult to distinguish geometrically similar targets with different scattering mechanisms, limiting detection accuracy and generalization ability in complex backgrounds.

[0004] To overcome the aforementioned shortcomings, existing technologies attempt to introduce phase information, but they have significant deficiencies in feature modeling and fusion mechanisms: First, simply cascading the phase and amplitude maps into a real-valued network severs their inherent physical coupling, making it easy for high-frequency noise in the phase to interfere with the extraction of amplitude backbone features; Second, while using complex convolutional networks to model SLC data can maintain the complex algebraic structure, it suffers from high computational complexity and training difficulty, and its integration with mainstream deep architectures such as Transformer is still immature; Third, the amplitude-phase dual-stream fusion architecture lacks an effective mutual guidance mechanism and fails to fully reflect the asymmetry of amplitude and phase information contributions in SAR complex imaging data (amplitude provides stable macroscopic scattering intensity / contour characterization, while phase provides fine-grained structural clues but is more susceptible to noise).

[0005] Furthermore, the dynamic range and distribution of SAR amplitude components vary greatly under different bands and polarization modes, and the statistical characteristics of phase components are completely different. Existing technologies lack adaptive normalization and alignment mechanisms to address the statistical heterogeneity of these two components, making it difficult to achieve collaborative modeling and stable fusion within a unified feature space. This problem is particularly prominent in complex scenarios and under multi-source data conditions. Summary of the Invention

[0006] Therefore, it is necessary to provide a method, apparatus, and device for rotating target detection based on single-view complex SAR images that can effectively utilize the amplitude and phase information of single-view complex SAR images, avoid phase noise interference, and enhance the dominant role of amplitude features, in order to address the above-mentioned technical problems.

[0007] A rotating target detection method based on single-look complex SAR images, the method comprising: Acquire single-look complex SAR images; The single-view complex SAR image is decoupled using a data preprocessing unit to obtain amplitude and phase components. The amplitude and phase components are then preprocessed to construct amplitude input tensors and phase input tensors. Using an asymmetric dual-stream backbone feature extraction unit, multi-scale feature extraction is performed on the amplitude input tensor and the phase input tensor respectively. Specifically, the amplitude feature backbone extraction unit with a hierarchical Transformer network architecture extracts the multi-scale global semantic features of the amplitude input tensor, and the phase auxiliary branch with a lightweight convolutional residual network architecture extracts the multi-scale structural texture features of the phase input tensor under the semantic constraints of the multi-scale global semantic features. Using structured mutual guidance units, based on the fusion approach of amplitude as the main factor and phase enhancement, the same-level features of the multi-scale global semantic features and multi-scale structural texture features are spatially and channel aligned, and then a differentiated fusion strategy is adopted according to the semantic level to achieve amplitude-phase collaborative fusion, resulting in an enhanced amplitude feature sequence. The enhanced amplitude feature sequence is reorganized into a multi-scale feature pyramid. The target detection output unit is then used to obtain the target category and the rotated bounding box based on the multi-scale feature pyramid by rotating the target detection branch.

[0008] In one embodiment, preprocessing the amplitude components to construct the corresponding amplitude input tensor includes: The amplitude component is subjected to dynamic range compression processing, and then the compressed amplitude component is normalized to obtain a normalized amplitude image. The normalized amplitude image is copied along the channel dimension to construct a three-channel amplitude input tensor.

[0009] In one embodiment, logarithmic compression is used when performing dynamic range compression on the amplitude component; The normalization process involves normalizing the logarithmically compressed amplitude components according to their value range to obtain a normalized amplitude image.

[0010] In one embodiment, preprocessing the phase components to construct the corresponding phase input tensor includes: The phase components are subjected to continuous embedding processing to map the phase information to Euclidean space, thereby obtaining the phase embedding features of the two channels. The phase embedding features of the two channels are then used as the phase input tensor.

[0011] In one embodiment, the amplitude feature backbone extraction unit includes an image block layer, an input embedding layer, multiple cascaded Transformer layer blocks, and a multi-scale feature output layer connected in sequence. The image segmentation layer and the input embedding layer first process the three-channel amplitude input tensor into blocks, and then complete the dimension adjustment and feature embedding through the embedding layer to obtain the initial embedded features. The multiple cascaded Transformer layer blocks perform hierarchical progressive processing on the initial embedded features to obtain semantic features at different scales. Each Transformer layer block adopts a multi-head self-attention mechanism and a feedforward neural network structure, and adjacent Transformer layer blocks are connected through residual connections and layer normalization. The multi-scale feature output layer directly outputs each level block to generate semantic features at different scales, thus obtaining the multi-scale global semantic features.

[0012] In one embodiment, the phase-assisted branch includes a start block, multiple cascaded convolutional residual layers, and a feature adaptation layer connected in sequence; The initial block performs convolution, normalization, activation, and pooling on the phase input tensor to obtain initial phase features; The multiple cascaded convolutional residual layers perform hierarchical downsampling and feature extraction on the initial phase features. Each convolutional residual layer is composed of several stacked basic residual blocks. The multi-scale global semantic features at the same level output by the amplitude feature backbone extraction unit are synchronously input to the structured mutual guidance unit and interact with the phase features output by the corresponding convolutional residual layer to guide the phase auxiliary branch to focus on the target-related region through feature interaction. The feature adaptation layer adjusts the dimension of the phase features output by the last cascaded convolutional residual layer to obtain the multi-scale structural texture features that match the amplitude semantics.

[0013] In one embodiment, the structured mutual guidance unit employs a differentiated fusion strategy according to semantic hierarchy, including: The global semantic features and structural texture features corresponding to the shallow semantic layer are fused using a linear residual coupling method; The global semantic features and structural texture features corresponding to the semantic middle layer are fused using a unidirectional cross-attention approach; The global semantic features and structural texture features corresponding to the deep semantic layer are fused using a mutually guided closed-loop approach.

[0014] This application also provides a rotating target detection device based on single-look complex SAR images, the device comprising: SAR image acquisition module, used to acquire single-look complex SAR images; The decoupling preprocessing module is used to decouple the single-view complex SAR image using the data preprocessing unit to obtain amplitude components and phase components, and to preprocess the amplitude components and phase components respectively to construct amplitude input tensors and phase input tensors. The feature extraction module is used to extract multi-scale features from the amplitude input tensor and the phase input tensor using an asymmetric dual-stream backbone feature extraction unit. Specifically, the amplitude feature backbone extraction unit with a hierarchical Transformer network architecture extracts the multi-scale global semantic features of the amplitude input tensor, and the phase auxiliary branch with a lightweight convolutional residual network architecture extracts the multi-scale structural texture features of the phase input tensor under the semantic constraints of the multi-scale global semantic features. The amplitude feature enhancement module is used to utilize structured mutual guidance units, based on the fusion idea of ​​amplitude as the main factor and phase enhancement, to spatially and channel align the same-level features of the multi-scale global semantic features and multi-scale structural texture features, and then adopt a differentiated fusion strategy according to the semantic level to achieve amplitude-phase collaborative fusion, thereby obtaining an enhanced amplitude feature sequence. The rotating target detection module is used to reorganize the enhanced amplitude feature sequence into a multi-scale feature pyramid, and use the target detection output unit to obtain the target category and rotating bounding box based on the multi-scale feature pyramid through the rotating target detection branch.

[0015] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the above-described rotating target detection method based on single-view complex SAR images.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described rotating target detection method based on single-view complex SAR images.

[0017] The aforementioned method, apparatus, and device for rotating target detection based on single-view complex SAR images decouple the single-view complex SAR image using a data preprocessing unit to obtain amplitude and phase components. These components are then preprocessed to construct amplitude and phase input tensors. An asymmetric dual-stream backbone feature extraction unit is used to extract multi-scale features from both the amplitude and phase input tensors. Specifically, the amplitude feature backbone extraction unit, employing a hierarchical Transformer network architecture, extracts the multi-scale global semantic features of the amplitude input tensor, while the phase input tensor, employing a lightweight convolutional residual network architecture, extracts the multi-scale global semantic features of the phase input tensor. The auxiliary branch extracts multi-scale structural texture features from the phase input tensor under the semantic constraints of multi-scale global semantic features. Using a structured mutual guidance unit, based on the fusion approach of amplitude-centric and phase-enhanced, it performs spatial and channel alignment on the same-level features of multi-scale global semantic features and multi-scale structural texture features. Then, it adopts a differentiated fusion strategy according to semantic level to achieve amplitude-phase collaborative fusion, resulting in an enhanced amplitude feature sequence. The enhanced amplitude feature sequence is reorganized into a multi-scale feature pyramid. Using the target detection output unit, the target category and rotated bounding box are obtained from the multi-scale feature pyramid through the rotation target detection branch.

[0018] This method can fully utilize the amplitude and phase information of single-view complex SAR images. Through an asymmetric dual-stream architecture, it achieves amplitude dominance and phase assistance, which not only makes up for the lack of single amplitude feature information, but also avoids phase noise interference. Combined with semantic hierarchical differential fusion and multi-scale feature pyramid, it effectively improves the detection accuracy and positioning accuracy of targets with different scales and angles of rotation. At the same time, it takes into account the global semantic capture capability and model lightweighting, ensuring the feasibility of engineering applications. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a rotating target detection method based on a single-look complex SAR image in one embodiment. Figure 2 This is a schematic diagram of the overall structure of a structured, mutually guided attention network with amplitude-phase fusion in one embodiment, wherein... Figure 2 (a) A schematic diagram showing the overall structure of the network. Figure 2 (b) shows a schematic diagram of the amplitude feature extraction unit, i.e., the amplitude backbone. Figure 2 (c) shows a schematic diagram of the phase auxiliary branch, i.e., the phase trunk; Figure 3 This is a schematic diagram illustrating a differentiated fusion strategy employed at the semantic level in one embodiment; Figure 4 This is a schematic diagram illustrating the result of feature fusion using a differentiated fusion strategy at the semantic level in one embodiment. Figure 5 This is a structural block diagram of a rotating target detection device based on a single-look complex SAR image in one embodiment; Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] Many existing methods employ an "amplitude-only" paradigm, converting SLC (Single Look Complex) data into amplitude or intensity images while discarding phase information. This fails to utilize the microscopic scattering structure and topological information inherent in the phase, limiting their ability to distinguish between complex clutter backgrounds, weak targets, and targets with similar scattering mechanisms. Some methods simply cascade the phase map as an additional channel with the amplitude map into a real-valued network, without explicitly modeling the physical coupling between amplitude and phase in complex data. Since phase is severely affected by speckle noise, in the absence of structural constraints, phase features can easily contaminate the amplitude backbone, weakening the overall feature representation capability. Some methods employ fully complex networks to maintain the algebraic form of complex data, resulting in high parameter and computational overhead, high optimization difficulty, and immature integration methods with current mainstream large-scale deep networks. Furthermore, existing amplitude-phase dual-stream architectures often use symmetrical fusion strategies, failing to fully reflect the asymmetry of amplitude and phase information contributions under SAR coherent imaging conditions, and lacking a mutual guidance mechanism for selectively and controllably importing phase information at different semantic levels. Furthermore, given the significant differences between amplitude and phase in terms of value range, statistical distribution, and noise characteristics, existing methods often employ uniform feature processing and normalization strategies, making it difficult to achieve stable and effective collaborative representation within a unified feature space.

[0022] Therefore, this application provides a method for detecting rotating targets based on single-look complex SAR images, comprising the following steps: Step S100: Acquire a single-look complex SAR image.

[0023] Step S110: Using the data preprocessing unit, the single-view complex SAR image is decoupled to obtain the amplitude component and the phase component. The amplitude component and the phase component are preprocessed respectively to construct the amplitude input tensor and the phase input tensor.

[0024] Step S120: Using the asymmetric dual-stream backbone feature extraction unit, multi-scale feature extraction is performed on the amplitude input tensor and the phase input tensor respectively. Specifically, the amplitude feature backbone extraction unit with a hierarchical Transformer network architecture extracts the multi-scale global semantic features of the amplitude input tensor, and the phase auxiliary branch with a lightweight convolutional residual network architecture extracts the multi-scale structural texture features of the phase input tensor under the semantic constraints of the multi-scale global semantic features.

[0025] Step S130: Using structured mutual guidance units, based on the fusion approach of amplitude as the main factor and phase enhancement, spatial and channel alignment is performed on the same-level features of multi-scale global semantic features and multi-scale structural texture features. Then, a differentiated fusion strategy is adopted according to the semantic level to achieve amplitude-phase collaborative fusion, resulting in an enhanced amplitude feature sequence.

[0026] Step S140: The enhanced amplitude feature sequence is reorganized into a multi-scale feature pyramid. The target category and the rotated bounding box are obtained by using the target detection output unit through the rotated target detection branch based on the multi-scale feature pyramid.

[0027] In this method, for single-look complex SAR data, an asymmetric dual-stream network structure with amplitude information as the main component and phase information as the auxiliary component, and a corresponding mutual guidance feature fusion mechanism are provided. This enables amplitude features and phase features to be modeled collaboratively within a real-valued network framework, reducing the interference of phase noise on feature representation, thereby improving the accuracy and stability of SAR target detection in complex backgrounds.

[0028] like Figure 2 The diagram shows the overall structure of the structured mutual-guided attention network for amplitude-phase fusion proposed in this method, including the aforementioned data preprocessing unit, asymmetric dual-stream backbone feature extraction unit, structured mutual-guided unit, and target detection output unit. Specifically, the network takes a single-view complex (SLC) SAR image as input, completes amplitude and phase decoupling and corresponding preprocessing and encoding, extracts multi-scale features from the amplitude main semantic stream and the phase structure auxiliary stream, and performs hierarchical fusion through the structured mutual-guided module (i.e., the structured mutual-guided unit). These features are then input to the feature pyramid and the rotated target detection head (i.e., the target detection output unit), and output the target category and rotated bounding box detection results.

[0029] In step S100, the imaging modes of the single-view complex SAR image include strip and spotlight, covering typical scenes such as land and sea. The image resolution can be selected according to the detection requirements to adapt to the detection scenes of rotating targets at different scales. The acquired single-view complex SAR image needs to undergo preliminary quality verification to remove invalid data with severe noise pollution and blurred imaging, so as to ensure the effect of subsequent processing.

[0030] In step S110, the single-look complex SAR image is... Decoupled to real part With the imaginary part And calculate the corresponding amplitude component M and phase component. The following formula is used: (1) (2) In this embodiment, the amplitude component is used to characterize the macroscopic scattering intensity and geometric contour information, while the phase component is used to characterize the structural texture and scattering topological clues under coherent imaging.

[0031] In this embodiment, when preprocessing the amplitude components to construct the corresponding amplitude input tensor, the process includes: performing dynamic range compression on the amplitude components, then normalizing the compressed amplitude components to obtain a normalized amplitude image, and copying the normalized amplitude image along the channel dimension to construct a three-channel amplitude input tensor.

[0032] Specifically, given the large dynamic range of the amplitude component and its significant distribution differences under different imaging conditions, logarithmic compression and normalization are performed on the amplitude component to improve numerical stability and cross-scene comparability. Logarithmic compression is used for dynamic range compression of the amplitude component, and then the logarithmically compressed amplitude component is normalized according to its value range to obtain a normalized amplitude image. The process is represented as follows: (3) In formula (3), This is the normalized amplitude image. To prevent the use of tiny positive constants with a denominator of zero, preferably, Set as .

[0033] Furthermore, the normalized single-channel amplitude Copy along the channel dimension to construct a three-channel amplitude tensor This makes it compatible with the input format of backbone networks that use natural image pre-trained parameters.

[0034] In this embodiment, when preprocessing the phase components to construct the corresponding phase input tensor, the process includes: performing continuous embedding processing on the phase components, mapping the phase information to Euclidean space, obtaining the phase embedding features of the two channels, and using the phase embedding features of the two channels as the phase input tensor.

[0035] Specifically, since the domain of the phase channel is Direct input will Periodic jumps occur at certain points, affecting the stability of network learning. To avoid this problem, a two-channel sine and cosine encoding method is used to map the phase to Euclidean space, obtaining the phase embedding. , represented as: (4) Furthermore, the two-channel phase is embedded As input to the phase-assisted branch, construct the phase tensor.

[0036] In this embodiment, before inputting the amplitude input tensor and phase input tensor into the asymmetric dual-stream backbone feature extraction unit, a consistent scaling transformation (e.g., interpolation-based resizing) is performed on the amplitude input tensor and phase input tensor to unify the input size. The unified input size can be set to any fixed value according to actual application requirements. And preferably satisfy the network step size alignment constraint, for example, it is an integer multiple of 32.

[0037] In one embodiment, the input size can be normalized to The corresponding magnitude input tensor size is The phase input tensor size is .

[0038] In step S120, the amplitude feature backbone extraction unit and the phase auxiliary branch in the asymmetric dual-stream backbone feature extraction unit are used to model global semantic features and extract coherent texture and fine-grained structural information, respectively. Preferably, the phase auxiliary branch and the subsequent structured mutual guidance unit can use group normalization to enhance the robustness of mini-batch statistics, while the amplitude feature backbone extraction unit can use the default normalization form of its pre-trained structure (e.g., layer normalization) to ensure the consistency and stability of pre-trained weight transfer.

[0039] In this embodiment, the amplitude feature backbone extraction unit includes an image segmentation layer, an input embedding layer, multiple cascaded Transformer layer blocks, and a multi-scale feature output layer connected in sequence. The image segmentation layer and the input embedding layer first segment the three-channel amplitude input tensor into blocks, and then the embedding layer performs dimensionality adjustment and feature embedding to obtain initial embedded features. The multiple cascaded Transformer layer blocks then perform hierarchical progressive processing on the initial embedded features to obtain semantic features at different scales. Each Transformer layer block employs a multi-head self-attention mechanism and a feedforward neural network structure, and adjacent Transformer layer blocks are connected through residual connections and layer normalization. Finally, the multi-scale feature output layer directly outputs the semantic features generated by each layer block at different scales to obtain multi-scale global semantic features.

[0040] Specifically, the amplitude feature extraction unit adopts a hierarchical Transformer backbone network to model long-range dependencies and multi-scale contexts, which is suitable for extracting target structural semantics in complex backgrounds and supports the transfer of pre-trained parameters from natural images to accelerate convergence and improve generalization ability.

[0041] like Figure 2 As shown, each Transformer layer block adopts a structure combining a window multi-head self-attention module (W-MSABlock), a shift window multi-head self-attention module (SW-MSA Block), and a feedforward neural network (MLP). The modules are connected internally through layer normalization (LN) and residual connections, and adjacent layer blocks output features at different scales through feature transfer.

[0042] In this embodiment, the phase-assisted branch uses two-channel input after phase sine and cosine encoding. As input, a lightweight convolutional residual backbone network, PhaseCNN, is used to extract multi-scale phase structure features. The "lightweight / auxiliary" aspect refers to the fact that this branch does not employ global modeling structures such as the Transformer, but instead uses a small number of convolutional residual blocks for local structure representation. Furthermore, its output is not used as an independent input to the detection head, but only as a source of structural information for subsequent structured mutual guidance units to perform residual supplementation and controlled modulation of the amplitude backbone features (see [link to relevant documentation]). Figure 3 The shallow layer adopts a residual injection method based primarily on amplitude. The deep layer achieves mutual guidance through cross-attention, and the final detection is based solely on the amplitude features after phase enhancement.

[0043] Specifically, the phase-assisted branch comprises a sequentially connected starting block, multiple cascaded convolutional residual layers, and a feature adaptation layer. The starting block performs convolution, normalization, activation, and pooling on the phase input tensor to obtain initial phase features. Subsequently, multiple cascaded convolutional residual layers perform hierarchical downsampling and feature extraction on the initial phase features. Each convolutional residual layer consists of stacked basic residual blocks. The multi-scale global semantic features output from the amplitude feature backbone extraction unit are synchronously input to the structured mutual guidance unit and interact with the phase features output from the corresponding convolutional residual layers to guide the phase-assisted branch to focus on the target-related region. Finally, the feature adaptation layer adjusts the dimensions of the phase features output from the last cascaded convolutional residual layer to obtain multi-scale structural texture features that match the amplitude semantics.

[0044] In one embodiment, the phase-assisted branch includes three cascaded convolutional residual layers.

[0045] In one embodiment, the starting block is composed of Convolution (stride 2), group normalization, GELU activation, and Max pooling (step 2) is constructed by sequentially connecting the elements.

[0046] In one embodiment, each convolutional residual layer is composed of several basic residual blocks stacked together. The first residual block in each convolutional residual layer preferably uses a stride of 2 for downsampling. The basic residual blocks preferably employ a pre-activated residual connection of group normalization, GELU, convolution, group normalization, GELU, and convolution, and the block output is obtained by adding the main branch output to the shortcut branch output. When the scale / channel changes, the shortcut branch preferably... Convolution achieves alignment.

[0047] Preferably, to facilitate alignment and mutual guidance fusion with the multi-scale semantic features of the amplitude backbone, the phase-assisted branch is configured as follows: the number of output channels of the starting block is 96, and the number of output channels of the three convolutional residual layers are 192, 384, and 768, respectively.

[0048] In step S130, for the same level and After spatial and channel alignment, the input structured mutual guidance unit performs three-stage semantic hierarchical processing using a differentiated fusion strategy according to semantic level. This includes: fusing global semantic features and structural texture features corresponding to the shallow semantic level using linear residual coupling; fusing global semantic features and structural texture features corresponding to the middle semantic level using unidirectional cross-attention; and fusing global semantic features and structural texture features corresponding to the deep semantic level using a mutual guidance closed-loop approach. The specific process is as follows: Figure 3 As shown.

[0049] In this embodiment, the structured mutual guidance unit adopts a three-stage differentiated fusion strategy according to semantic levels. Specifically, for the shallow semantic layer, a linear residual coupling method is used. After concatenating the amplitude feature and phase feature channels, the amplitude backbone residual signal is learned through 1×1 convolution, normalization, and activation function. The residual term is added to the amplitude feature and normalized to obtain the fused feature. For the middle semantic layer, a unidirectional cross-attention method is used. The amplitude feature is used as the query, and the downsampled phase feature is used as the key and value. After feature rearrangement and cross-modal attention calculation, phase information is injected in the residual manner and normalized to obtain the fused feature. For the deep semantic layer, a mutual guidance closed-loop method is used, including two stages. In the first stage, attention is calculated with the phase feature as the query and the amplitude feature as the key and value to achieve phase feature purification and update. In the second stage, attention is calculated with the amplitude feature as the query and the purified phase feature as the key and value to achieve selective enhancement of the amplitude feature and obtain the deep fused feature.

[0050] Specifically, in the shallow stage, the feature map has high resolution but weak semantics. To avoid excessive perturbation of the amplitude-based backbone statistics while introducing phase structure cues, a linear residual coupling method with "amplitude as the backbone and phase as the residual correction" is adopted.

[0051] First, the splicing is performed in the channel dimension: (5) In formula (5), the symbol | represents splicing along the channel dimension.

[0052] Subsequently, through convolution Intermediate normalization (e.g., group normalization) and activation functions (For example, GELU) learns the residual signal of the amplitude backbone, and the process is represented as follows: (6) Finally, the residual terms are added to the amplitude characteristics and then normalized at the end. The process of obtaining the fusion features is represented as follows: (7) Under this mechanism, It is explicitly preserved as a core semantic representation. Phase branches are only passed through... Apply local residual correction. This applies when the phase information is consistent with and complementary to the amplitude semantics. The network generates a non-zero response at the corresponding position. When the phase information is redundant or unreliable, the network can learn to suppress it to near zero, thereby avoiding the direct injection of high-noise phase features into the main semantic stream at shallow layers.

[0053] Furthermore, spatial resolution is reduced and semantic information is enhanced in the middle stage. Considering the computational complexity and feature alignment difficulty, a one-way cross-attention strategy is adopted to inject phase texture into amplitude semantics in one go, using amplitude features as query Q and downsampled phase features as key K and value V.

[0054] Specifically, the two-dimensional features are rearranged into a sequence and spatial downsampling is introduced to reduce computational overhead, as follows: (8) In formula (8), and These represent the flattening and inverse flattening operators, respectively, used to flatten spatial dimensions. Rearranged with channel dimension to a length of or The sequence is then processed and the inverse operation is performed. It is a linear transformation. Indicates that the core is An average pooling operator with a step size of 2 is used to reduce the size of the attention matrix and smooth local noise.

[0055] Subsequently, cross-modal dot product attention is calculated and phase information is injected in a residual manner. After terminal normalization, the fused features are obtained, represented as: (9) In formula (9), The feature dimension representing the attention space; This is a cross-modal attention graph, where each row, after normalization, represents the contribution of different phase positions to the current query position within a given amplitude semantic context. This mechanism enables attention weights... It is fully modulated by amplitude context, thereby enabling fine-grained adaptive weighting of downsampled phase channels; phase information only participates in the reconstruction of amplitude representation when amplitude semantics considers a certain spatial location / channel to be important and the corresponding phase response to be reliable.

[0056] Furthermore, at the deepest scale, amplitude features already possess strong class discrimination capabilities. To avoid indiscriminate injection of phase information after multi-layer propagation, which weakens the separability of the semantic space, the structured mutual guidance module constructs a mutual guidance closed-loop structure at this scale, and adopts a two-stage cross-attention strategy to achieve a deep fusion strategy of "cleaning first, then enhancing".

[0057] Specifically, the first stage aims to utilize the strong semantic information of amplitude features as guidance to reconstruct phase features, thereby suppressing the interference of background noise on phase features, i.e., achieving "purification" of phase features. Using phase feature P as the query and amplitude feature M as the key and value, the key and value are preferably downsampled using average pooling before being used in attention calculation to obtain the first cross-attention weight matrix. Then, the amplitude value vector is weighted and summed using these weights, and added to the residual of the original phase feature to achieve a reweighted update of the phase features, generating the updated phase features. (10) (11) In formulas (10) and (11), This represents the attention weight matrix calculated using phase as the query and amplitude as the key / value. This is the value vector obtained by linearly projecting the pooled magnitude features; The phase features are "purified" under the guidance of amplitude semantic prior.

[0058] Next, the second stage uses the "cleaned" phase features as a reliable supplement to structural information to selectively enhance the amplitude representation.

[0059] Specifically, the amplitude feature M is used as the query, and the updated phase feature is used as the query. As keys and values, and preferably after performing average pooling downsampling on the keys and values, they are used in attention calculation to obtain the second cross-attention weight matrix; the phase value vector is weighted and summed using the weights and added to the original amplitude feature residual to achieve selective enhancement of the amplitude representation, resulting in the final deep fusion representation F: (12) (13) In formulas (12) and (13), Indicated by amplitude characteristics For Query, to purify phase features The attention weights calculated for the Key; To The value vector obtained by linear projection.

[0060] like Figure 4 As shown, this is a schematic diagram of the results after performing a hierarchical fusion strategy of residual linear injection, cross-attention fusion, and mutual guidance loop sequentially at different semantic levels. The shallow layer adopts linear coupling with amplitude as the main body and phase as the residual correction. The middle layer adopts unidirectional cross-attention with amplitude conditionalization and performs spatial downsampling on the phase key value branch to reduce the amount of computation and smooth the noise. The deep layer constructs a two-stage mutual guidance loop of "first purifying the phase and then enhancing the amplitude".

[0061] In step S140, the enhanced feature sequence is obtained after multi-stage fusion in step S130. In the target detection output unit, the feature sequence is reorganized into a multi-scale feature pyramid and input into the subsequent rotated target detection branch, outputting the target detection result.

[0062] Specifically, the independent output path of the phase branch is not retained in the structure. Only the "amplitude representation after phase enhancement" is retained as the detection basis. Thus, the design principle of "amplitude as the main factor and phase enhancement" is implemented at the network structure level. The position and intensity of phase injection are explicitly controlled through a deep "purification-enhancement" closed loop, which improves the interpretability and robustness of the cross-modal fusion process.

[0063] In this embodiment, the target detection output unit adopts a hierarchical cascaded structure, containing three sequentially connected functional components. Within this unit, a rotational candidate region is initially generated by rotating the RPN head. Next, the rotational RoI feature extraction component extracts and refines the features corresponding to the candidate regions. Finally, the 2FC bounding box head is rotated, and the refined RoI features are processed by a fully connected layer to output the bounding box parameters of the rotated target, thus completing the rotating target detection task.

[0064] Furthermore, this method can achieve stable target detection under single-view complex SAR image conditions and obtain good robustness and discriminability in complex backgrounds. Its feasibility and beneficial effects are mainly based on the following technical mechanisms and structural constraints.

[0065] First, the complementary nature and stable representation of amplitude and phase information form the basis for input-side realization. SLC SAR echoes simultaneously contain amplitude and phase information. The amplitude component primarily reflects macroscopic scattering intensity and geometric contours, while the phase component contains fine-grained texture structures and scattering topological cues under coherent imaging conditions. To ensure stable learning of both types of information in a real-valued network, this invention applies logarithmic compression and normalization to the amplitude component to mitigate numerical instability caused by excessive dynamic range and cross-scene distribution drift; and employs sine and cosine Euclidean embedding for the phase component to eliminate... The discontinuous nature of the periodicity enhances the continuity and optimizability of the training process. This provides a stable and comparable input representation foundation for subsequent feature extraction, cross-modal fusion, and end-to-end optimization.

[0066] Secondly, an asymmetric dual-stream division of labor is adopted to implement the fusion approach of "amplitude-centric, phase-enhanced," ensuring that the feature extraction and fusion paths are consistent with the imaging meaning and statistical characteristics of amplitude / phase components under coherent imaging conditions. Considering that target category discrimination relies more on stable semantic and contour representations, this method sets the amplitude branch as the main semantic stream, using a hierarchical Transformer backbone to extract multi-scale contextual and long-range dependent semantic features, and sets the phase branch (PhaseCNN) as the structure-assisted stream, using a lightweight convolutional network to extract coherent texture and fine-grained structural cues, limiting it to a "supplementary / modulated" information source for amplitude semantics rather than an independent decision path. Thus, this asymmetric division of labor utilizes the relatively stable semantic representation capabilities of the amplitude component and limits the phase component to a controlled supplementary / modulated information source under amplitude semantic constraints. This avoids directly injecting phase features, which are significantly affected by speckle noise, into the backbone in a symmetrical manner, causing feature contamination or a decrease in semantic representation capabilities, thereby improving the overall modeling stability.

[0067] Finally, the hierarchical mutual guidance mechanism of the structured mutual guidance module (structured mutual guidance unit) is used to achieve controlled fusion of controlled injection, noise suppression, and discrimination enhancement.

[0068] In this paper, considering that while phase information contains structural details, it is significantly affected by speckle noise, and its unconstrained introduction across all levels can easily lead to feature contamination and disrupt shallow statistical stability, a differentiated fusion strategy is implemented according to semantic levels. For shallow features, a structure-preserving linear residual coupling is adopted, with weak injection achieved through amplitude backbone and phase residual correction. This ensures that the phase only participates in updates as learnable residuals, structurally reducing the risk of noise spreading to the main semantic flow. For the mid-level stage, amplitude-conditional unidirectional cross-attention fusion is employed. Amplitude semantics serves as the query, determining the introduction position and intensity of phase information. Spatial downsampling is performed on the phase key / value side to reduce the attention scale and smooth local noise, achieving selective reweighting of phase structural information. For the deep level, a mutually guided closed loop is constructed (phase is purified first, amplitude is enhanced later). First, amplitude semantics guides the reweighting of phase features to complete purification, and then the purified phase selectively enhances the amplitude representation. This avoids the decrease in semantic space separability caused by indiscriminate phase superposition and ensures that phase information mainly acts on key regions related to target discrimination.

[0069] Meanwhile, the fully real-valued implementation path of this method facilitates engineering reproduction and framework integration. Through input representation, dual-stream division of labor, and structured mutual guidance mechanisms, it achieves effective utilization of complex information within the real-valued network framework without introducing special operators such as complex convolution. Therefore, it can be smoothly integrated with commonly used rotating target detection workflows (feature pyramids, multi-scale detection heads, and rotating NMS, etc.), facilitating reproduction, comparison, and transfer within general deep learning training and evaluation systems.

[0070] This paper also demonstrates the effectiveness of the proposed method through experiments. In the experiments, FAIR-CSAR was used. Training and testing were conducted using the SLC subset of the Composite Image Target Dataset V1.0. This dataset contains 22 typical aerial, maritime, and ground targets, labeled with rotated bounding boxes. The experimental data was divided into a predefined training set, validation set, and test set for model training, parameter selection, and final performance evaluation. During data processing, all SLC SAR images were uniformly resampled to... (Preferred H=W=1024). Amplitude component M adopts... Dynamic range compression and min–max normalization are performed, followed by copying the single-channel amplitude to three channels to match the input format of the ImageNet pre-trained Swin-T backbone; phase components Using sine and cosine two-channel encoding This is to ensure the continuity of the phase representation in the numerical space.

[0071] In the experiment, the training strategies employed included amplitude backbone and correlation methods used for comparison. The weights were initialized using Swin-T weights pre-trained on ImageNet-1K; data augmentation was performed using random horizontal and vertical flipping; the MMDetection / MMRotate framework was used, with AdamW as the optimizer, an initial learning rate of 1×10^(-4), weight decay of 0.05, and a batch size of 4; training was conducted for 24 epochs, with learning rate decay performed at epochs 16 and 22.

[0072] Furthermore, during the training of the entire network framework, the total network loss L is composed of the Region Candidate Generation (RPN) loss and the Rotating Detector Head (RoI) loss, which can be expressed as: (14) In formula (14), For foreground / background binary classification loss, cross-entropy loss with Sigmoid is preferred; For multi-class classification loss, Softmax cross-entropy loss is preferred; The rotation bounding box regression loss is used to constrain the position, scale, and angle errors between the predicted rotation box and the labeled rotation box. The weighting coefficients used to balance the regression loss contributions of the candidate region generation stage and the rotating detection head stage are positive. In a preferred embodiment of the invention, the following values ​​are used: , To match the default weight configuration of the classification / regression loss; in other implementations, , It can also be adjusted within a preset range based on the degree of imbalance in the dataset categories, the target scale distribution, and the convergence stability. Furthermore, in the preferred embodiment, the regression loss uses the Smooth L1 loss, applied to data processed by a rotated frame encoder (e.g., ...). The regression objective is calculated after (or equivalent parameterization); in addition, a distribution / geometric distance loss suitable for rotating frames (such as KLD or GWD loss) can be selected to further improve the stability of angle regression.

[0073] In the experiment, the average precision (mAP) of rotating target detection with an IoU threshold of 0.5 was used as the main performance indicator. The AP for each category was calculated by the area under the precision-recall (P-R) curve, and the final average of the APs for the 22 categories was obtained as mAP(50). The number of parameters and floating-point operations (FLOPs) were recorded as complexity reference indicators.

[0074] Under the same training and evaluation settings, the proposed method is compared with several representative rotating object detection methods, including the two-stage ROI Transformer, the anchor-box-based Rotated RetinaNet, and the anchor-box-free Rotated FCOS. As shown in Table 1, the proposed method achieves mAP(50)=40.6% on the test set; it achieves higher average accuracy compared to ROI Transformer (34.6%), Rotated RetinaNet (34.2%), and Rotated FCOS (35.4%), with an improvement of 5.2 percentage points compared to Rotated FCOS. Meanwhile, the proposed method has 95.01M parameters and 343.58G FLOPs under these settings, representing the computational scale corresponding to the two-stream fusion and attention interaction structure.

[0075] Further ablation experiments were conducted to verify the effectiveness of the "amplitude-centric, phase-enhanced" structural constraint and the structured mutual guidance module. Stepwise ablation was performed under the same dataset and evaluation settings, including: Amplitude-only baseline: Only the amplitude channel was input, and the model degenerated into a conventional rotating detector based on Swin-T, with mAP(50)=39.5%. This configuration tends to miss some weak targets or scenes with complex scattering mechanisms; for example, the AP of the Warship class is 9.7%. Dual-stream and shallow residual fusion: A phase branch was introduced on the basis of the amplitude-only baseline, and a shallow linear residual method was used for fusion, with mAP(50)=39.8%, which is 0.3 percentage points higher than the amplitude-only baseline; at the same time, the AP of the Warship class increased from 9.7% to 18.5%, indicating that phase structure cues can provide supplementary information for some targets. However, since this fusion method lacks selective suppression of phase noise and effective structural information, the overall gain is still limited.

[0076] In this method, a hierarchical fusion pipeline (mid-level cross-attention and deep-level mutual guidance closed loop) with a structured mutual guidance module (structured mutual guidance unit) is enabled on the basis of dual-stream and shallow residual fusion. The mAP(50) = 40.6%, which is 1.1 and 0.8 percentage points higher than that of amplitude baseline fusion and dual-stream and shallow residual fusion, respectively. At the category level, the improvement is more stable in complex scenarios and targets with significant morphological changes. For example, the AP of the Bridge category increased from 27.9% to 30.8%, and the AP of the Fokker-50 category increased from 73.4% to 84.8%.

[0077] Table 1 Comparison of detection performance of each model and ablation experimental data

[0078] Furthermore, in the visualization comparison of typical complex scenarios, only the amplitude baseline is more prone to missed detection or broken target contours in situations such as dense target distribution, weak targets surrounded by clutter, and slender target structures. This method introduces phase topological cues through a structured mutual guidance module (structured mutual guidance unit) and performs controlled fusion, which can enhance the response of weak targets and maintain the structural continuity of slender targets, thereby obtaining more stable detection performance in complex backgrounds.

[0079] The aforementioned rotating target detection method based on single-look complex SAR images takes single-look complex SAR data as the object and incorporates both amplitude and phase information into a unified modeling framework, making fuller use of the information in complex data compared to traditional "amplitude-only" methods. Through an asymmetric dual-stream structure that prioritizes amplitude semantics and uses phase structure as an auxiliary mechanism, this method demonstrates better discriminative ability for targets with similar geometric contours but different scattering mechanisms in complex sea surface and port / land surface clutter backgrounds. Experimental results based on a typical SLC SAR target detection dataset show that the proposed method improves detection accuracy and robustness compared to existing representative rotating target detection methods.

[0080] Furthermore, this method proposes a structured mutual guidance module with differentiated designs at different semantic levels (shallow, mid-layer, and deep layers). The fusion approach of amplitude-cleaned phase and phase-assisted amplitude is explicitly encoded into the network structure: lightweight residual fusion is used in the shallow layers, while the mid-to-high layers utilize a cross-attention module with spatial downsampling to achieve information exchange between the amplitude and phase branches. Under the constraint of amplitude features, phase features are selectively guided and reweighted. Simultaneously, phase features are used to supplement and optimize amplitude features in the deep layers, which helps suppress the propagation of speckle noise in the network and mitigate feature contamination that may result from simple channel cascading and symmetrical fusion. Thus, the network can maintain the stability of amplitude feature representation while fully utilizing fine-grained phase texture information to improve target boundary and pose characterization.

[0081] This method addresses the differences in amplitude and phase values, statistical distributions, and noise characteristics by designing corresponding preprocessing and representation methods: the amplitude branch employs logarithmic compression and dynamic range normalization, while the phase branch uses sine and cosine encoding to avoid phase jumps, supplemented by independent lightweight convolutional branches for feature extraction. Combined with a hierarchical, mutually guided feature fusion strategy, a collaborative modeling framework adapted to amplitude-phase statistical heterogeneity is constructed. In multi-scenario, multi-class experiments on publicly available SLC SAR target detection datasets, this framework demonstrates good detection accuracy and robustness, providing a technical foundation for further joint amplitude and phase modeling in multi-polarization, multi-source SAR tasks.

[0082] Finally, this method enables joint modeling of complex SAR data on a fully real-valued deep network architecture without the need for special operators such as complex convolution. It can be smoothly integrated with existing rotating target detection network structures, training processes, and evaluation metrics. This method facilitates transfer, reproduction, and comparative studies within general target detection frameworks, and is also beneficial for subsequent expansion and functional development in other SAR application scenarios and tasks (such as scene classification and fine segmentation).

[0083] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0084] In one embodiment, such as Figure 5 As shown, a rotating target detection device based on single-look complex SAR images is provided, comprising: a SAR image acquisition module 200, a decoupling preprocessing module 210, a feature extraction module 220, an amplitude feature enhancement module 230, and a rotating target detection module 240, wherein: The SAR image acquisition module 200 is used to acquire single-look complex SAR images.

[0085] The decoupling preprocessing module 210 is used to decouple the single-view complex SAR image using the data preprocessing unit to obtain amplitude components and phase components, and to preprocess the amplitude components and phase components respectively to construct amplitude input tensors and phase input tensors.

[0086] The feature extraction module 220 is used to extract multi-scale features from the amplitude input tensor and the phase input tensor using an asymmetric dual-stream backbone feature extraction unit. Specifically, the amplitude input tensor's multi-scale global semantic features are extracted by the amplitude feature backbone extraction unit, which employs a hierarchical Transformer network architecture. Under the semantic constraints of the multi-scale global semantic features, the phase input tensor's multi-scale structural texture features are extracted by the phase auxiliary branch, which employs a lightweight convolutional residual network architecture.

[0087] The amplitude feature enhancement module 230 is used to utilize the structured mutual guidance unit, based on the fusion idea of ​​amplitude as the main factor and phase enhancement, to spatially and channel align the same-level features of the multi-scale global semantic features and multi-scale structural texture features, and then adopt a differentiated fusion strategy according to the semantic level to achieve amplitude-phase collaborative fusion, thereby obtaining the enhanced amplitude feature sequence.

[0088] The rotating target detection module 240 is used to reorganize the enhanced amplitude feature sequence into a multi-scale feature pyramid, and use the target detection output unit to obtain the target category and rotating bounding box according to the multi-scale feature pyramid through the rotating target detection branch.

[0089] Specific limitations regarding the rotating target detection device based on single-look complex SAR images can be found in the limitations of the rotating target detection method based on single-look complex SAR images mentioned above, and will not be repeated here. Each module in the aforementioned rotating target detection device based on single-look complex SAR images can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0090] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a rotating target detection method based on single-look complex SAR images. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0091] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0092] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: Acquire single-look complex SAR images; The single-view complex SAR image is decoupled using a data preprocessing unit to obtain amplitude and phase components. The amplitude and phase components are then preprocessed to construct amplitude input tensors and phase input tensors. Using an asymmetric dual-stream backbone feature extraction unit, multi-scale feature extraction is performed on the amplitude input tensor and the phase input tensor respectively. Specifically, the amplitude feature backbone extraction unit with a hierarchical Transformer network architecture extracts the multi-scale global semantic features of the amplitude input tensor, and the phase auxiliary branch with a lightweight convolutional residual network architecture extracts the multi-scale structural texture features of the phase input tensor under the semantic constraints of the multi-scale global semantic features. Using structured mutual guidance units, based on the fusion approach of amplitude as the main factor and phase enhancement, the same-level features of the multi-scale global semantic features and multi-scale structural texture features are spatially and channel aligned, and then a differentiated fusion strategy is adopted according to the semantic level to achieve amplitude-phase collaborative fusion, resulting in an enhanced amplitude feature sequence. The enhanced amplitude feature sequence is reorganized into a multi-scale feature pyramid. The target detection output unit is then used to obtain the target category and the rotated bounding box based on the multi-scale feature pyramid by rotating the target detection branch.

[0093] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire single-look complex SAR images; The single-view complex SAR image is decoupled using a data preprocessing unit to obtain amplitude and phase components. The amplitude and phase components are then preprocessed to construct amplitude input tensors and phase input tensors. Using an asymmetric dual-stream backbone feature extraction unit, multi-scale feature extraction is performed on the amplitude input tensor and the phase input tensor respectively. Specifically, the amplitude feature backbone extraction unit with a hierarchical Transformer network architecture extracts the multi-scale global semantic features of the amplitude input tensor, and the phase auxiliary branch with a lightweight convolutional residual network architecture extracts the multi-scale structural texture features of the phase input tensor under the semantic constraints of the multi-scale global semantic features. Using structured mutual guidance units, based on the fusion approach of amplitude as the main factor and phase enhancement, the same-level features of the multi-scale global semantic features and multi-scale structural texture features are spatially and channel aligned, and then a differentiated fusion strategy is adopted according to the semantic level to achieve amplitude-phase collaborative fusion, resulting in an enhanced amplitude feature sequence. The enhanced amplitude feature sequence is reorganized into a multi-scale feature pyramid. The target detection output unit is then used to obtain the target category and the rotated bounding box based on the multi-scale feature pyramid by rotating the target detection branch.

[0094] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for detecting rotating targets based on single-look complex SAR images, characterized in that, The method includes: Acquire single-look complex SAR images; The single-view complex SAR image is decoupled using a data preprocessing unit to obtain amplitude and phase components. The amplitude and phase components are then preprocessed to construct amplitude input tensors and phase input tensors. Using an asymmetric dual-stream backbone feature extraction unit, multi-scale feature extraction is performed on the amplitude input tensor and the phase input tensor respectively. Specifically, the amplitude feature backbone extraction unit with a hierarchical Transformer network architecture extracts the multi-scale global semantic features of the amplitude input tensor, and the phase auxiliary branch with a lightweight convolutional residual network architecture extracts the multi-scale structural texture features of the phase input tensor under the semantic constraints of the multi-scale global semantic features. Using structured mutual guidance units, based on the fusion approach of amplitude as the main factor and phase enhancement, the same-level features of the multi-scale global semantic features and multi-scale structural texture features are spatially and channel aligned, and then a differentiated fusion strategy is adopted according to the semantic level to achieve amplitude-phase collaborative fusion, resulting in an enhanced amplitude feature sequence. The enhanced amplitude feature sequence is reorganized into a multi-scale feature pyramid. The target detection output unit is then used to obtain the target category and the rotated bounding box based on the multi-scale feature pyramid by rotating the target detection branch.

2. The rotating target detection method based on single-look complex SAR images according to claim 1, characterized in that, When preprocessing the amplitude components to construct the corresponding amplitude input tensor, the process includes: The amplitude component is subjected to dynamic range compression processing, and then the compressed amplitude component is normalized to obtain a normalized amplitude image. The normalized amplitude image is copied along the channel dimension to construct a three-channel amplitude input tensor.

3. The rotating target detection method based on single-look complex SAR images according to claim 2, characterized in that, Logarithmic compression is used when performing dynamic range compression on the amplitude components. The normalization process involves normalizing the logarithmically compressed amplitude components according to their value range to obtain a normalized amplitude image.

4. The rotating target detection method based on single-look complex SAR images according to claim 1, characterized in that, When preprocessing the phase components to construct the corresponding phase input tensor, the process includes: The phase components are subjected to continuous embedding processing to map the phase information to Euclidean space, thereby obtaining the phase embedding features of the two channels. The phase embedding features of the two channels are then used as the phase input tensor.

5. The rotating target detection method based on single-look complex SAR images according to claim 1, characterized in that, The amplitude feature backbone extraction unit includes an image block layer, an input embedding layer, multiple cascaded Transformer layer blocks, and a multi-scale feature output layer connected in sequence. The image segmentation layer and the input embedding layer first process the three-channel amplitude input tensor into blocks, and then complete the dimension adjustment and feature embedding through the embedding layer to obtain the initial embedded features. The multiple cascaded Transformer layer blocks perform hierarchical progressive processing on the initial embedded features to obtain semantic features at different scales. Each Transformer layer block adopts a multi-head self-attention mechanism and a feedforward neural network structure, and adjacent Transformer layer blocks are connected through residual connections and layer normalization. The multi-scale feature output layer directly outputs each level block to generate semantic features at different scales, thus obtaining the multi-scale global semantic features.

6. The rotating target detection method based on single-look complex SAR images according to claim 5, characterized in that, The phase-assisted branch includes a start block connected in sequence, multiple cascaded convolutional residual layers, and a feature adaptation layer; The initial block performs convolution, normalization, activation, and pooling on the phase input tensor to obtain initial phase features; The multiple cascaded convolutional residual layers perform hierarchical downsampling and feature extraction on the initial phase features. Each convolutional residual layer is composed of several stacked basic residual blocks. The multi-scale global semantic features at the same level output by the amplitude feature backbone extraction unit are synchronously input to the structured mutual guidance unit and interact with the phase features output by the corresponding convolutional residual layer to guide the phase auxiliary branch to focus on the target-related region through feature interaction. The feature adaptation layer adjusts the dimension of the phase features output by the last cascaded convolutional residual layer to obtain the multi-scale structural texture features that match the amplitude semantics.

7. The rotating target detection method based on single-look complex SAR images according to claim 1, characterized in that, The structured mutual guidance unit adopts a differentiated fusion strategy according to semantic hierarchy, including: The global semantic features and structural texture features corresponding to the shallow semantic layer are fused using a linear residual coupling method; The global semantic features and structural texture features corresponding to the semantic middle layer are fused using a unidirectional cross-attention approach; The global semantic features and structural texture features corresponding to the deep semantic layer are fused using a mutually guided closed-loop approach.

8. A rotating target detection device based on single-look complex SAR images, characterized in that, The device includes: SAR image acquisition module, used to acquire single-look complex SAR images; The decoupling preprocessing module is used to decouple the single-view complex SAR image using the data preprocessing unit to obtain amplitude components and phase components, and to preprocess the amplitude components and phase components respectively to construct amplitude input tensors and phase input tensors. The feature extraction module is used to extract multi-scale features from the amplitude input tensor and the phase input tensor using an asymmetric dual-stream backbone feature extraction unit. Specifically, the amplitude feature backbone extraction unit with a hierarchical Transformer network architecture extracts the multi-scale global semantic features of the amplitude input tensor, and the phase auxiliary branch with a lightweight convolutional residual network architecture extracts the multi-scale structural texture features of the phase input tensor under the semantic constraints of the multi-scale global semantic features. The amplitude feature enhancement module is used to utilize structured mutual guidance units, based on the fusion idea of ​​amplitude as the main factor and phase enhancement, to spatially and channel align the same-level features of the multi-scale global semantic features and multi-scale structural texture features, and then adopt a differentiated fusion strategy according to the semantic level to achieve amplitude-phase collaborative fusion, thereby obtaining an enhanced amplitude feature sequence. The rotating target detection module is used to reorganize the enhanced amplitude feature sequence into a multi-scale feature pyramid, and use the target detection output unit to obtain the target category and rotating bounding box based on the multi-scale feature pyramid through the rotating target detection branch.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images