SAR target detection method based on phase-amplitude frequency domain cross modulation

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) target detection method based on phase-amplitude frequency domain cross modulation. The problems that due to inherent speckle noise in an SAR image, a target is difficult to distinguish, and an existing method is insufficient in frequency domain information complementarity and calculation efficiency are solved. According to the scheme, firstly, input SAR image features are converted into a frequency domain, and a phase spectrum and an amplitude spectrum of the SAR image features are obtained; the method is characterized in that through phase-amplitude frequency domain cross modulation, interaction and guidance between phases and amplitudes are promoted by using an attention mechanism, accurate adjustment, cooperative enhancement and effective noise suppression of the phases and the amplitudes are realized, and amplitude spectrums and phase spectrums after optimization modulation are generated; and finally, the optimized frequency domain features are transformed back to a spatial domain for subsequent target detection. According to the method, cross modulation is realized through a phase-amplitude feature exchange (PATE) mechanism and a frequency band division self-attention (BPSA) mechanism, so that the precision and robustness of SAR image target detection are remarkably improved, and the calculation complexity is effectively controlled.

Description

Technical Field

[0001] The present invention relates to the field of synthetic aperture radar (SAR) and SAR image processing technology, and in particular to a method for realizing high-precision SAR image target detection under strong noise interference. Background Art

[0002] SAR target detection technology has important applications in remote sensing, but its development has been constrained by the inherent speckle noise and complex scene interference inherent in SAR images. Current technology faces three key challenges: First, the inherent speckle noise in SAR images, which is highly coupled to target features. The granular distribution of speckle noise in the spatial domain closely resembles target edges and texture features. Second, the limitations of frequency domain information utilization. Existing methods often process the amplitude or phase spectrum in isolation, ignoring their complementarity in the frequency domain. Third, the conflict between computational efficiency and detection accuracy. Although global frequency domain interaction mechanisms can improve detection accuracy, their extremely high computational complexity makes it difficult to meet the real-time processing requirements of large-scale SAR imagery.

[0003] The main shortcomings of existing technologies are reflected in the following aspects: (1) Isolated processing of amplitude and phase: Existing technologies tend to process amplitude and phase information separately. Most methods focus mainly on the amplitude component and ignore the phase details that are crucial to maintaining the original spatial relationship and structural integrity of the object. (2) Increased model complexity and excessive interaction: Most methods involve the interaction of signals across the entire spectrum, which not only increases computational complexity but also may lead to excessive processing of frequency domain features. This excessive processing may manifest as excessive smoothing of object features, thereby reducing detection accuracy. Summary of the Invention

[0004] To address the above problems, this paper focuses on phase-amplitude modulation techniques and frequency domain band division techniques, aiming to systematically address the limitations of existing technologies. The purpose of the present invention is to provide a SAR target detection method DenoDet V2 based on phase-amplitude frequency domain cross-modulation, a new method for SAR image target detection using a mutual guidance mechanism through the complementary characteristics of amplitude and phase information.

[0005] The technical solution to realize the present invention is: a SAR target detection method based on phase-amplitude frequency domain cross modulation, comprising the following steps:

[0006] Step 1: preprocess the input SAR image to obtain the feature map to be processed;

[0007] Step 2: convert the feature map to be processed into the frequency domain through Fourier transform, and separate the amplitude spectrum and phase spectrum from the frequency domain representation;

[0008] Step 3: performing a frequency domain cross-modulation operation on the amplitude spectrum and the phase spectrum, wherein the information interaction between the amplitude spectrum and the phase spectrum is utilized, and modulation information is generated through an attention mechanism, and then the modulation information is used to adjust the amplitude spectrum and the phase spectrum respectively to achieve mutual enhancement and noise suppression, thereby obtaining a modulated amplitude spectrum and a modulated phase spectrum;

[0009] Step 4, reconstructing the modulated amplitude spectrum and the modulated phase spectrum into an enhanced frequency domain representation;

[0010] Step 5: Restoring the enhanced frequency domain representation to the spatial domain through inverse Fourier transform to obtain an enhanced spatial domain feature map;

[0011] Step 6: Input the enhanced spatial domain feature map into the target detection network and output the target detection result.

[0012] Furthermore, the generation of modulation information through the attention mechanism in step 3 is implemented using a phase-amplitude feature exchange mechanism PATE, which generates the modulation information based on the information interaction between the amplitude spectrum and the phase spectrum. In the PATE mechanism:

[0013] a. Implementing adjustment of the amplitude spectrum: first, using the characteristic information of the amplitude spectrum itself to establish the focus of its adjustment, then using the corresponding phase spectrum characteristic information to provide interactive guidance context information for the adjustment of the amplitude spectrum, and generating modulation information for adjusting the amplitude spectrum through attention calculation within the PATE mechanism;

[0014] b. Implement adjustment of the phase spectrum; first, utilize the characteristic information of the phase spectrum itself to establish the focus of its adjustment, then utilize the corresponding amplitude spectrum characteristic information to provide interactive guidance context information for the adjustment of the phase spectrum, and generate the modulation information for adjusting the phase spectrum through the attention calculation within the PATE mechanism.

[0015] Furthermore, the modulation information is generated through the attention mechanism described in step 3, and the frequency band partitioning self-attention mechanism BPSA is used to independently calculate the PATE attention mechanism in the divided frequency band groups, thereby generating modulation information for adjusting the amplitude spectrum and phase spectrum in each frequency band group.

[0016] Furthermore, the preprocessing described in step 1 includes normalizing the input SAR image, dividing it into image blocks of a preset size, and mirror-filling edge areas of the image blocks.

[0017] Furthermore, the object detection network in step 6 adopts a feature pyramid structure and includes bidirectional cross-scale connections to fuse multi-scale and multi-resolution features.

[0018] Compared with existing technologies, this invention offers significant advantages: DenoDet V2 is the first method to implement this phase-amplitude bidirectional reference exchange feature denoising method. DenoDet V2 integrates the frequency domain denoising module into the detection framework and achieves a fundamental difference from existing methods through the dynamic phase-amplitude feature exchange mechanism PATE. It also uses the band-partitioned self-attention mechanism (BPSA) to maintain the inherent characteristics of frequency domain signals while reducing the dimensionality of attention calculations through grouping, thereby improving efficiency.

[0019] DenoDet V2 is a significant advancement that leverages the complementary nature of amplitude and phase information, achieving mutual enhancement between the phase and amplitude spectra through a frequency-band intermodulation mechanism. DenoDet V2 demonstrates state-of-the-art performance on a variety of SAR datasets while reducing computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the overall architecture diagram of DenoDet V2.

[0021] Figure 2 This is a schematic diagram of the frequency sorting of DFT signal components.

[0022] Figure 3 This is the BPSA module flow chart.

[0023] Figure 4 It is a schematic diagram of phase-amplitude dynamic exchange.

[0024] Figure 5 This is a schematic diagram of the detection effect of DenoDet V2. DETAILED DESCRIPTION

[0025] The core of this invention is to construct a novel SAR target detection framework driven by frequency domain feature cross-modulation. The basic process begins by normalizing the input SAR image, partitioning it into blocks and preserving boundary information through edge padding. Subsequently, a block-wise Fourier transform is used to decompose the image blocks into amplitude and phase spectra. This block-wise Fourier transform significantly reduces computational complexity.

[0026] Building on this foundation, the present invention further utilizes the Band-wise Partition Self-Attention (BPSA) mechanism to first partition the amplitude and phase spectra into several frequency band groups. Crucially, within each frequency band group, the Phase and Amplitude Token Exchange (PATE) mechanism is employed to generate an adaptive frequency-domain attention map, enabling dynamic feature fusion and mutual enhancement between amplitude and phase features.

[0027] Through this PATE-based attention mechanism, the BPSA attention mechanism module is able to achieve deep, bidirectional dynamic information exchange and optimization between the amplitude and phase spectra within each frequency band. For example, it can use the relatively robust structural characteristics of the phase information to guide the amplitude spectrum to perform adaptive noise filtering and highlight target details; conversely, the optimized amplitude spectrum information can also support the calibration and accuracy improvement of the phase spectrum.

[0028] Ultimately, the amplitude and phase modulation information for each frequency band, generated using the band-partitioning self-attention mechanism (BPSA) and the phase-amplitude feature exchange mechanism (PATE), is applied to the corresponding spectral data through methods such as element-by-element multiplication, effectively suppressing noise interference and enhancing and integrating structural features that are beneficial for target detection. These finely modulated frequency domain features are then restored to the spatial domain through an inverse Fourier transform, forming a high-quality enhanced feature map. This is then input into a lightweight target detection network to produce highly accurate target detection results.

[0029] The present invention adopts: (1) Amplitude-phase mutual guidance mechanism, which utilizes the robustness of phase information to adaptively filter out noise from the amplitude spectrum, which not only improves the signal-to-noise ratio of the target feature, but also further improves the phase accuracy through the optimized amplitude spectrum, ensuring the retention of key target details. (2) Phase-amplitude feature exchange mechanism PATE, which realizes the dynamic exchange and strategic interaction of phase and amplitude roles, and can finely recalibrate and integrate information, thereby improving the fidelity of the reconstructed signal. This mutual enhancement effect between the amplitude spectrum and the phase spectrum produces a synergistic feature denoising effect, iteratively improving the target detection accuracy. The PATE mechanism is a cross-spectral feature exchange method that enables the network to adaptively choose to use phase data to guide amplitude data, or to use amplitude data to drive phase data. (3) Band-partitioned self-attention mechanism BPSA, which better adapts to the characteristics of frequency domain features, divides the spectrum into N×N local frequency band grids, so that phase and amplitude interact only within their respective frequency bands. This local frequency band-based processing method achieves fine-grained feature denoising, accurately adapts to the uniqueness of each frequency band, and fully retains key spatial and structural information.

[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0031] 1 System Configuration and Data Preparation

[0032] The environment for this example is divided into two parts: hardware and software. The hardware device uses a computing server equipped with an NVIDIA RTX4090 GPU, 24GB of video memory, and an Intel Xeon Gold 6338 processor, and the operating system is Ubuntu 20.04. The software framework is built on PyTorch 1.12 and integrates CUDA 11.6 and cuDNN 8.4 acceleration libraries. The object detection module relies on MMDetection 3.3.0, and image preprocessing is completed using OpenCV 4.5.

[0033] The experimental data uses the public datasets SARDet-100K and AIR-SARShip-1.0, which contain more than 100,000 512×512 training images, covering six types of targets such as ships and aircraft; AIR-SARShip-1.0 provides 3000×3000 high-resolution test images to simulate dense ships and complex background scenes.

[0034] 2DenoDet V2 Overall Architecture

[0035] The overall architecture of the proposed DenoDet V2 is as follows Figure 1 shown.

[0036] The architecture is based on a general object detection network and integrates the DFTDeno module for plug-and-play enhancement. Through a designed attention mechanism, DFTDeno performs dynamic soft threshold denoising in the transformed domain of feature maps. In DenoDetV2, this is incorporated into the backbone network stage to optimize the feature map extraction process.

[0037] The DFTDeno module consists of a forward two-dimensional discrete Fourier transform (DFT), a dynamic threshold denoising, and a reverse two-dimensional DFT. C×H×W , its two-dimensional DFT forward transform function can be defined as:

[0038]

[0039] where h,u∈[0,H-1] and w,v∈[0,W-1].

[0040] According to Euler's formula, the above formula can be expanded as:

[0041]

[0042] Where H and W represent the height and width of the feature map respectively, h,u∈[0,H-1] and w,v∈[0,W-1] represent the coordinate position of the statistic, m c，u，v Represents the tensor value at coordinate (u, v) in channel c after forward DFT. The signal component after DFT can be decomposed into real and imaginary parts using formula 2:

[0043]

[0044] Next, for the real part R∈R C×H×W With the imaginary part I∈R C×H×W After two nonlinear transformations, the amplitude spectrum A and phase spectrum P∈[π,-π) of the frequency domain signal are extracted. The specific calculation is as follows: Amplitude spectrum and phase spectrum calculation:

[0045]

[0046] Formula (2) shows that each signal component originates from the pixel-level values ​​in the original feature map, essentially encoding some global information. To process this highly aggregated information, an efficient attention mechanism is used for signal modulation. This operation suppresses noise while enhancing information features through the following transformation:

[0047]

[0048] in, and Represent the amplitude and phase of the modulated signal respectively, ⊙ represents the element-by-element multiplication operation, G(·)∈R C×H×W It is the attention map output by the designed module G.

[0049] Finally, the modulated amplitude and phase information are recombined, and the frequency signal is restored to the spatial domain through the inverse discrete Fourier transform IDFT. The specific steps are as follows:

[0050]

[0051]

[0052] For the sake of brevity, the normalization coefficients in both the forward and inverse transforms are omitted.

[0053] In the final implementation, the original phase characteristics are not used, but the phase characteristics are used according to formula (8) and The phase information is decoupled, effectively alleviating the angular boundary discontinuity problem. Subsequently, the two orthogonal components are reconciled using trigonometric identities, ensuring the parity consistency of the final phase representation through strict mathematical equivalence.

[0054] 3-band self-attention mechanism BPSA

[0055] In the frequency domain, adjacent features may not necessarily show local correlation, so the long-range context modeling ability of the self-attention mechanism is crucial. In addition, in the transform domain, each sampling point corresponds to a specific frequency component, which is highly compatible with the position encoding strategy. Based on this, the self-attention mechanism is adopted as the basic architecture for frequency band signal modulation. Represents the frequency domain tensor after discrete Fourier transform DFT. The output of the signal-modulated self-attention block is calculated as follows:

[0056]

[0057] in, is the frequency domain attention map, is the modulated output signal, and BPSA stands for the band-partitioned self-attention mechanism. In the basic band self-attention BSA, the input features are divided into Where d = H × W, and each coordinate pixel represents an independent frequency signal in the basic BSA. The result of each frequency band after BSA processing can be expressed as:

[0058]

[0059] in, It is used to generate the query vector Q i , key vector K i Sum value vector V i The learnable weights, E is the position encoding parameter, is a scaling factor related to the number of frequency bands.

[0060] In the feature map after DFT transformation, the frequency changes non-monotonically with the two-dimensional coordinates: it decreases first and then increases, and the frequency is is centrally symmetric. To facilitate modeling, the spectrum needs to be shifted in the center, such as Figure 2 Based on this spectral characteristic, a band-partitioned self-attention mechanism BPSA is proposed, which strategically decomposes the global attention operation into sub-band operations that can be calculated in parallel, as shown in Figure 3 As shown. Where h and w represent the vertical and horizontal step sizes of the frequency band division, respectively. In formula (13), the number of frequency band groups d should be rewritten as And Q i , K i 、V i ∈R h*w The query vector, key vector, and value vector represent the same set of frequency bands. This frequency band processing paradigm achieves dual goals: maintaining the distribution characteristics of frequency domain data and regional attention calculation after dimensionality reduction.

[0061] 4Phase-amplitude characteristic exchange PATE

[0062] Traditional frequency domain denoising methods usually process the amplitude spectrum and phase spectrum independently. This separation process limits the model's ability to adaptively interact with the two modal information. To overcome this fundamental defect, a cross-spectral feature exchange method is proposed, which enables the network to adaptively choose to use phase data to guide amplitude data, or use amplitude data to drive phase data. The feature exchange process of amplitude and phase graphs is as follows: Figure 4 shown.

[0063] 1. Band alignment and grouping

[0064] (a) Using the same step size parameter, the amplitude spectrum and phase spectrum Split into discrete frequency band groups to ensure that the frequencies of the two sets of signals are aligned within the corresponding frequency bands.

[0065] (b) Amplitude subbands obtained after grouping and phase subbands Each subband

[0066] 2. Modality-Specific Feature Mapping

[0067] Each sub-band generates a modality-specific query, key, and value representation through a multi-layer perceptron MLP:

[0068]

[0069] in, are the learnable linear mapping weights of amplitude and phase features, The query, key, and value representing each band group. The key for the amplitude Sum The corresponding subband from the phase spectrum Phase Key Sum The corresponding subband from the magnitude spectrum Therefore, when the PATE mechanism builds attention for a certain modality, it will first perform self-reference processing based on the characteristics of the modality itself, and then introduce the feature information of the other modality for cross-modal interactive guidance processing. The two work together to achieve precise feature modulation and enhancement, breaking the isolation of traditional independent processing, realizing dynamic information complementarity of amplitude and phase, and reducing attention complexity through frequency band grouping, thereby improving computational efficiency and accuracy.

[0070] 5. Performance Verification

[0071] Combine Figure 5This method was systematically validated on the public datasets SARDet-100K, SAR-AIRcraft-1.0, and AIR-SARShip-1.0. The experimental results in Tables 1 and 2 show that its detection accuracy, small target performance, and noise robustness significantly outperform existing technologies.

[0072] Table 1 Comparison results of methods on the SARDet-100K dataset

[0073]

[0074]

[0075] Table 2. Comparison results of methods on AIR-SARShip-1.0 and SAR-Aircraft datasets

[0076]

[0077] On the SARDet-100K dataset, this method achieves an average precision (mAP) of 56.71% with 32.60M parameters, which is 1.7% higher than the existing best method, and the small target detection accuracy is 51.45%.

[0078] On the AIR-SARShip-1.0 high-resolution dataset, this method achieved a mAP of 73.98%, with the detection accuracy of small-sized targets such as ships improved by 1.56%, verifying the robustness of the frequency domain interaction mechanism to complex background and noise interference.

[0079] In the SAR-AIRcraft-1.0 dataset with a large number of low signal-to-noise ratio scenes, it still maintains a mAP of 69.93%, indicating that bidirectional spectrum modulation effectively suppresses the interference of high-frequency noise on the target structure, proving the strong generalization ability of the method.

[0080] Traditional methods treat amplitude and phase spectra in isolation, ignore the complementarity of frequency domain features, and introduce excessive computational overhead in global frequency domain interactions, making it difficult to balance noise suppression and target feature preservation. To address these shortcomings, the present invention achieves a coordinated optimization of enhanced noise robustness and real-time, efficient detection through frequency domain decomposition and bidirectional modulation, combined with a lightweight architecture design.

[0081] In summary, this method, DenoDet V2, achieves mutual enhancement of the phase and amplitude spectra through dynamic exchange of amplitude-phase features, producing a synergistic feature denoising effect that gradually improves target detection accuracy. It also implements fine-grained feature denoising through the band-partitioning self-attention module (BPSA), accurately adapting to the uniqueness of each frequency band while fully preserving key spatial and structural information. This method systematically addresses the core challenges of noise coupling, small target under-detection, and scene generalization in SAR imagery. Extensive experiments on multiple SAR datasets have verified the advanced performance of DenoDet V2, demonstrating its potential for robust and accurate target detection in noisy SAR environments, providing reliable technical support for high-precision real-time detection.

Claims

1. A SAR target detection method based on phase-amplitude frequency domain cross-modulation, characterized in that: The following steps are involved: Step 1: preprocess the input SAR image to obtain the feature map to be processed; Step 2: convert the feature map to be processed into the frequency domain through Fourier transform, and separate the amplitude spectrum and phase spectrum from the frequency domain representation; Step 3: performing a frequency domain cross-modulation operation on the amplitude spectrum and the phase spectrum, wherein the information interaction between the amplitude spectrum and the phase spectrum is utilized, and modulation information is generated through an attention mechanism, and then the modulation information is used to adjust the amplitude spectrum and the phase spectrum respectively to achieve mutual enhancement and noise suppression, thereby obtaining a modulated amplitude spectrum and a modulated phase spectrum; Step 4, reconstructing the modulated amplitude spectrum and the modulated phase spectrum into an enhanced frequency domain representation; Step 5: Restoring the enhanced frequency domain representation to the spatial domain through inverse Fourier transform to obtain an enhanced spatial domain feature map; Step 6: Input the enhanced spatial domain feature map into the target detection network and output the target detection result.

2. The method according to claim 1, wherein: The generation of modulation information through the attention mechanism in step 3 is implemented using the phase-amplitude feature exchange mechanism PATE. The modulation information is generated based on the information interaction between the amplitude spectrum and the phase spectrum. In the PATE mechanism: a. Implementing adjustment of the amplitude spectrum: first, using the characteristic information of the amplitude spectrum itself to establish the focus of its adjustment, then using the corresponding phase spectrum characteristic information to provide interactive guidance context information for the adjustment of the amplitude spectrum, and generating modulation information for adjusting the amplitude spectrum through attention calculation within the PATE mechanism; b. Implement adjustment of the phase spectrum; first, utilize the characteristic information of the phase spectrum itself to establish the focus of its adjustment, then utilize the corresponding amplitude spectrum characteristic information to provide interactive guidance context information for the adjustment of the phase spectrum, and generate the modulation information for adjusting the phase spectrum through the attention calculation within the PATE mechanism.

3. The method according to claim 2, wherein: In step 3, the modulation information is generated through the attention mechanism, and the band partitioning self-attention mechanism BPSA is used to independently calculate the PATE attention mechanism in the divided frequency band groups, thereby generating modulation information for adjusting the amplitude spectrum and phase spectrum in each frequency band group.

4. The method according to claim 1, wherein: The preprocessing described in step 1 includes normalizing the input SAR image, dividing it into image blocks of a preset size, and mirror-filling the edge areas of the image blocks.

5. The method according to claim 1, wherein: The object detection network in step 6 adopts a feature pyramid structure and includes bidirectional cross-scale connections to fuse multi-scale and multi-resolution features.

Images