Sonar target detection method fusing directional echo shadow geometry and acoustic spectrum

Abstract

The application relates to the technical field of sonar target detection, and particularly discloses a sonar target detection method fusing directionality echo shadow geometry and acoustic spectrum, which comprises the following steps: acquiring a sonar image; inputting a backbone network to obtain a multi-scale feature map, and then inputting the neck structure into a detection head to output a target detection result; the backbone network comprises multiple stages, each stage adopts a stacked space-frequency module, and the space-frequency module sequentially comprises a sonar adaptive state space model, an acoustic frequency feature modulation model and a positional multilayer perception machine; the sonar adaptive state space model aligns a one-dimensional scanning sequence with an acoustic imaging axis, and then fuses each position with its neighborhood along a distance axis through an echo-shadow pairing operator, so that recursion is carried out along the direction developed by echoes and shadows; the acoustic frequency feature modulation model operates on the Fourier domain of features, decomposes the features into non-intersecting frequency subbands, and then reconstructs the features. The application realizes high-precision and low-computing-cost sonar target detection.

Description

Technical Field

[0001] This invention relates to the field of sonar target detection technology, and in particular to a sonar target detection method that integrates directional echo shadow geometry and acoustic spectrum. Background Technology

[0002] Extensive research has been conducted on sonar-based underwater target detection algorithms. In FLS detection, Valdenegro-Toro et al. pioneered the application of convolutional neural networks to target scoring in sonar images. Subsequent researchers have further introduced techniques such as multi-scale feature extraction, speckle removal, lightweight small target detection, and detector transfer based on knowledge distillation. In SSS image interpretation, early work utilized transfer learning to address the scarcity of annotations, while recent methods have superimposed task-specific modules such as feature super-resolution, deep feature selection enhancement, and wavelet-based frequency decoupling on top of a general backbone.

[0003] At the backbone network level, Convolutional Neural Networks (CNNs) model the context using local receptive fields; Transformers cover the global context with self-attention; and Visual State Space Models (SSM / Mamba) achieve long-range modeling with linear computational complexity through sequential state propagation. Furthermore, several frequency domain learning methods (such as channel frequency domain attention, global Fourier filtering, and frequency-decoupled convolution) have been developed in the general vision domain to enhance feature discrimination capabilities.

[0004] Despite the progress made by the above methods, there are still significant limitations in practical sonar target detection applications. The fundamental reason is that sonar images carry two physical cues that have no corresponding objects in optical images, and existing detectors have not modeled them within the backbone. These two cues are: (1) directional echo-shadow geometry along the imaging axis: each target will produce a bright echo and a thin shadow extending along a fixed acoustic direction (along the distance axis in FLS, and along the vertical track direction in SSS); (2) spectral texture contrast: the target echo carries energy in the mid-to-high frequency band that is not present in the background water, while the low frequency band is mainly dominated by diffuse reverberation. The two are complementary: metal cylinders and stones with similar shapes are almost identical in spatial morphology, and detectors that rely solely on spatial cues will confuse them, but their spectral characteristics are different.

[0005] The shortcomings of existing methods are specifically manifested in the following ways: (1) Existing sonar detectors still use a pure spatial backbone designed for optical input. Its operators are not organized along the acoustic axis and lack frequency domain decomposition capability. Therefore, the two physical clues mentioned above are discarded at the same time. Some methods only add sonar-specific modules at the input or output end, but the backbone is still a general space processor. The problem has not been truly solved.

[0006] (2) In terms of long-range modeling, CNN backbone can only model local context and will miss echo-shadow pairs across the scene; although Transformer backbone can cover the whole, its quadratic complexity is not practical on the embedded platform commonly used for underwater inference; although the existing visual SSM provides long-range modeling with linear complexity, its scanning direction is fixed on the image grid (rows / columns) rather than the acoustic axis, so the sequence cannot track the directional echo-shadow structure, and the target evidence is scattered into the irrelevant background.

[0007] (3) Existing methods lack a dedicated processing channel for acoustic frequency cues, which can neither suppress diffuse low-frequency reverberation in the dominant background region nor amplify the mid-to-high frequency structure of the target echo.

[0008] (4) In the multi-scale feature fusion stage, the standard feature pyramid (FPN / PANet) mixes multi-scale features by uniform element-wise addition. The spatial response is densely distributed in the scene, while the frequency response is concentrated in the sparse small target area. Uniform addition will dilute the sparse frequency signal into the dense spatial signal, causing the frequency channel to be diluted when propagating from top to bottom. Summary of the Invention

[0009] This invention aims to solve the aforementioned problems. To this end, this invention provides a sonar target detection method that integrates directional echo shadow geometry and acoustic spectrum. It designs a sonar adaptive state-space modeling (SA-SSM) method to align and pair echo-shadow along the acoustic axis, an acoustic frequency feature modulation (AFFM) method to separate the target echo from the background reverberation, and an acoustic sensing feature aggregation (AAFA) method to maintain the balance of the two channels at the neck, thereby achieving high-precision and low-computational-cost sonar target detection.

[0010] This invention provides a sonar target detection method that integrates directional echo shadow geometry and acoustic spectrum, and the technical solution adopted is as follows: including: Acquire sonar images; After sonar images are patch-embedding, they are input into the backbone network to obtain multi-scale feature maps; The aforementioned backbone network comprises multiple stages, each employing stacked space-frequency modules. Within each space-frequency module are, in sequence, a sonar adaptive state-space model, an acoustic frequency feature modulation model, and a positional multilayer perceptron. The sonar adaptive state-space model aligns the one-dimensional scanning sequence with the acoustic imaging axis, and then incorporates the neighborhood of each position along the distance axis using an echo-shadow pairing operator, allowing the recursion to unfold along the direction formed by the echo and shadow. The acoustic frequency feature modulation model operates in the Fourier domain of the features, decomposing the features into disjoint frequency subbands, and then reconstructing the features. The multi-scale feature map is aggregated with acoustic sensing features through the neck structure and then fed into the detection head to output the target detection result.

[0011] Furthermore, the acoustic frequency characteristic modulation model is only enabled in the first two stages of the backbone network.

[0012] Furthermore, the working process of the sonar adaptive state-space model is as follows: The input features of the space-frequency module are divided into a feature stream and a gated stream by linear projection; Based on the gated flow, multiplicative gating is obtained; Based on the feature flow, a flattened sequence aligned along the acoustic axis is obtained; after the sequence is enhanced by the echo-shading pairing operator, it is input into the range axis forward scan branch and the range axis reverse scan branch respectively to perform range axis forward scan and range axis reverse scan; the sequence is input into the azimuth axis scan branch to perform azimuth axis scan. The sequences output by the forward scan branch of the range axis, the reverse scan branch of the range axis, and the scan branch of the azimuth axis are respectively passed through a selective state-space model to obtain a restored grid. The three restored grids are averaged, then subjected to layer normalization, and multiplied element-wise with a multiplicative gate to obtain the output features of the sonar adaptive state-space model.

[0013] Furthermore, the range axis is the direction in which the echoes and shadows are arranged in sequence, the forward-looking sonar is along the image height, the side-scan sonar is along the image width, and the azimuth axis is orthogonal to the range axis.

[0014] Furthermore, the sequences output by the forward scan branch of the range axis, the reverse scan branch of the range axis, and the scan branch of the azimuth axis are restored by independent selective state space models according to their respective scan order to obtain the restored mesh; each of the three independent selective state space models has independent parameters.

[0015] Furthermore, the echo-shadow pairing operator encodes spatial relationships along the distance axis as a token-level one-dimensional neighborhood aggregation; the specific process is as follows: For each position, the predictor provides a Dimensional aggregation weights and a gate, The neighborhood radius on each side of the distance axis; For each channel, aggregate the token features at that location independently using aggregate weights and gating.

[0016] Furthermore, the working process of the acoustic frequency characteristic modulation model is as follows: Perform an orthogonally normalized two-dimensional real Fourier transform on the input feature map, and then perform cascaded subband separation to obtain multiple non-overlapping high-frequency subbands and a low-frequency residual. The spatial attention weights are calculated from the high-frequency subband itself, and the high-frequency subband is modulated to obtain the modulated subband; The output characteristics of the acoustic frequency characteristic modulation model are obtained by calculating the sum of the low-frequency residual and all modulated subbands.

[0017] Furthermore, the cutoff parameters used for cascaded subband separation are 2 and 4; cascaded subband separation yields two non-overlapping high-frequency subbands and a low-frequency residual.

[0018] Furthermore, each pyramid level of acoustic sensing feature aggregation employs a frequency-space cross-gating unit, enabling the two channels to mutually reinforce each other with position-adaptive weights. The frequency-space cross-gating unit extracts spatial and frequency domain views from the pyramid features; A gate is generated from both the spatial and frequency domain views and applied to the other view; then summed with the pyramid feature to obtain the spatial-frequency domain cross-gated fusion feature, calculated as follows: in, It is a spatial-frequency domain cross-gated fusion feature. Gating coefficients generated for the spatial view. The gating coefficients generated for the frequency domain view. This is a view of the airspace. This is a frequency domain view. Features of a pyramid This is for element-wise multiplication.

[0019] Furthermore, the patch-embedding consists of two 3×3 convolutions with a stride of 2; The backbone network has 4 stages; The detection head is an anchor-frame based detection head.

[0020] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: (1) This invention is the first to design a backbone based on two physical clues that distinguish sonar images from optical images (directional echo-shadow geometry and acoustic spectrum texture), and embeds the two as parallel spatial and frequency channels into each module of the backbone, instead of simply adding sonar modules at the input / output end and keeping the backbone as a general-purpose spatial processor as in existing methods.

[0021] (2) In the spatial channel, this invention proposes SA-SSM, which aligns the scanning direction of the state space model with the acoustic imaging axis (distance axis, azimuth axis), and uses the echo-shadow pairing operator to adaptively aggregate the neighborhood along the distance axis token by token, so that the recursion follows the echo-shadow structure; while the scanning direction of the existing visual SSM (state space model) is fixed on the image row / column, and cannot track the structure.

[0022] (3) In the frequency channel, this invention proposes AFFM, which decomposes the features into non-overlapping frequency sub-bands and applies learnable spatial attention per sub-band and per position, explicitly enhancing the mid-to-high frequencies of target correlation and suppressing low-frequency reverberation; this is a capability that existing pure spatial backbones do not possess.

[0023] (4) In the neck fusion, the present invention proposes AAFA, which replaces the uniform element-by-element addition of the standard feature pyramid with position-adaptive frequency-space cross-gating (FSCG), so that the spatial and frequency channels can enhance each other and not dilute each other in multi-scale propagation.

[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the method provided by the present invention.

[0027] Figure 2 This is a flowchart of the sonar adaptive state-space modeling provided by the present invention.

[0028] Figure 3 This is a flowchart of acoustic frequency characteristic modulation provided by the present invention.

[0029] Figure 4 This is a flowchart of acoustic sensing feature aggregation provided by the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but should not be used to limit the scope of this invention.

[0031] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0032] The following is combined Figures 1 to 4 The present invention will be further described in detail below, describing the sonar target detection method of the present invention that integrates directional echo shadow geometry and acoustic spectrum: In this embodiment, as Figure 1 As shown, this paper presents a sonar target detection method that integrates directional echo shadow geometry and acoustic spectrum. Following the standard "backbone network—neck—detection head" pipeline, unlike existing methods that merge two physical cues into a single spatial path, this method constructs and fuses the spatial and frequency channels as two parallel streams within the backbone and neck. The method includes the following steps: Acquire sonar images.

[0033] Given an input sonar image , , Input the number of channels. For image height, For image width, For the real number domain, a patch-embedding process consisting of two 3×3 convolutions with a stride of 2 is first used to reduce the resolution to one-quarter of the original, before it is input into the backbone network. The backbone network consists of four stages, each outputting a four-scale feature map at strides {4, 8, 16, 32}. , The stages / pyramid levels are numbered, with the channel dimensions increasing progressively to (48, 96, 192, 384) and the depth to (2, 2, 5, 2). Each stage stacks several space-frequency modules, which in turn contain a Sonar-Adaptive State Space Model (SA-SSM), an Acoustic Frequency Feature Modulation (AFFM) model, and a positional multilayer perceptron (MLP). SA-SSM serves as the spatial channel primitive, and the Acoustic Frequency Feature Modulation (AFFM) model serves as the frequency channel primitive. The data is then fed into the neck structure for acoustic sensing feature aggregation (neck cross-channel fusion), and finally, the target detection result is output by an anchor-box-based detection head (Faster R-CNN). The model is trained end-to-end. The sonar images are either forward-looking sonar images or side-scan sonar images.

[0034] No. The calculation process of the space-frequency module in the stage is represented as follows: in, The input characteristics of the space-frequency module, The characteristics are those obtained after SA-SSM residuals. The features are those obtained after AFFM residuals. The final output features after passing through the MLP residuals. For layer normalization, It represents a stochastic depth. This is a sonar adaptive state-space model. It is a position-based multilayer sensor. This is an acoustic frequency characteristic modulation model. AFFM is only used in the first two stages ( The spatial resolution is high here, and the spectral decomposition still has observable structure; the features of the deep stages (stages 3 and 4) are dominated by low frequencies, so they are omitted to save overhead.

[0035] SA-SSM is the spatial channel primitive of the space-frequency module, such as Figure 2 Before the state-space recursion is initiated, SA-SSM first aligns the one-dimensional scanning sequence with the acoustic imaging axis (instead of the image rows / columns used by existing visual SSM), and then uses the echo-shadow pairing operator (ESP) to incorporate the neighborhood of each position along the distance axis, so that the recursion unfolds along the direction of echo and shadow formation.

[0036] Input characteristics of the space-frequency module The flow is divided into a feature flow and a gated flow by linear projection, and the feature flow is depth-separable. Internal features obtained after convolution and SiLU activation , , The internal feature width is used; the gated flow is activated by the SiLU, resulting in multiplicative gating. .

[0037] SA-SSM will The sequence is flattened into three sequences aligned along the acoustic axes and processed separately: a forward scan along the range axis, a reverse scan along the range axis (obtained by flipping the forward sequence), and a scan along the azimuth axis. The range axis is the direction in which the echoes and shadows are arranged in sequence, the forward-looking sonar is along the image height, the side-scan sonar is along the image width, and the azimuth axis is orthogonal to the range axis.

[0038] The two range axis branches (range axis forward scan branch and range axis reverse scan branch) share the same ESP-enhanced features; the azimuth axis scan branch uses the original features without ESP because shadows do not propagate along the azimuth axis.

[0039] The three sequences output from the three branches are processed in parallel by three independent selective SSMs; they are then restored in their respective scan order. The mesh is then restored, where the distance axis reverse scan branch is first flipped and restored.

[0040] The average of the restored meshes of the three branches is then applied, followed by layer normalization and multiplicative gating. Element-wise multiplication is performed, and finally, the original channel dimension of the module is restored by linear projection to obtain the SA-SSM output feature.

[0041] Echo-Shadow Pairing (ESP): In sonar images, each target casts a shadow along the range axis after its echo, with a distance of tens of pixels between them. ESP encodes this spatial relationship along the range axis as a token-level one-dimensional neighborhood aggregation. For distance axis coordinates, For each position, the azimuth coordinates are... , by a lightweight predictor Give one Dimensional aggregate weight With a scalar gate : in, For position The normalized neighborhood aggregation weight at the location, for Output position The original weight (logit) at the location. For position The gating coefficient at the location, for Output position The original gating coefficient at the location; In this embodiment, the neighborhood radius is the distance to each side of the axis. That is, the nine-point neighborhood. For neighborhood offset index; In order to be in Normalize on each offset, making ; It is the sigmoid function; This is the upper limit of the gate. In this embodiment This will keep the injection range within a reasonable range. This represents the normalized neighborhood aggregation weight.

[0042] Accordingly, for each channel Apply aggregation independently. Previously viewed sonar range axis. For example (in the case of side-scan sonar, the range axis is changed to...) The calculation formula is: in, For the channel c and position obtained by aggregation Output token features at the location, For the input tensor at channel c, position The original token feature value on, for The feature value of the neighboring token after a k-shift along the distance axis.

[0043] ESP is an overlay of a distance-axis neighborhood term onto the original token. The context is formed by weighted aggregation (outbound offsets are filled with boundary copies). The bias is initialized to a large negative value, so that at the beginning of training... ESP degenerates into an identity mapping; weights With gate Learned during end-to-end training, but with insufficient background signal. A larger signal can borrow context from its neighboring region; its own signal is already strong in echo and shadow areas. Keep it small. for The original weights output. The gating coefficient, for The output of the original gating coefficients.

[0044] Selective SSM recursion: Each scan branch processes a segment of length... The sequence. For a single channel, let... For sequence position Channel value at that location, , for The corresponding hidden state, , For the state dimension, the recurrence relation is: in, For position The output; This is the discretized state transition matrix. , For a learnable diagonal matrix, For position-by-position step size, ; For discretized input vectors, , For the input projection vector, ; To output the projection vector, ; For direct scalar, . This is the matrix transpose.

[0045] Selective mechanisms enable , , Complete depending on the current position Dimensional token features , : in, For position-by-position step size, , All are learnable projections. The activation function is used to ensure a positive step size. This recursion applies to all... Execution is performed in parallel on multiple channels, allowing the token content to adaptively determine the propagation and forgetting of the state at each step.

[0046] The three independent selective SSMs each have independent parameters: .

[0047] The acoustic frequency characteristic modulation model is a frequency channel primitive of the space-frequency module, such as... Figure 3 After SA-SSM aggregates tokens along the acoustic axis, the same feature map still mixes the target echo and background reverberation together, even though they occupy different frequency bands. AFFM operates directly in the Fourier domain of the features: it decomposes the features into disjoint frequency subbands, modulates each subband independently using learnable spatial attention, and then reconstructs the features.

[0048] Let the input feature map of AFFM be... , The workflow of AFFM is as follows: Spectrum transformation and cascaded subband separation: First, for The spectrum is obtained by performing an orthogonally normalized two-dimensional real Fourier transform (rFFT2D). For a set of strictly increasing cutoff parameters , In this embodiment, Pre-compute binary low-pass mask Deadline The larger the value, the tighter the low-pass region, therefore the masks are nested together. ). For the first A binary low-pass mask. For hierarchical indexes, J represents the total number of cutoff parameters.

[0049] right Peel off layer by layer in sequence: in, This indicates element-wise multiplication. This is the two-dimensional real inverse Fourier transform. The low-frequency component in the spatial domain is obtained by inverse transform after the j-th level mask filtering. For the j-th level high-frequency subband component, The output of the j-th iteration is the updated low-frequency residual feature. The frequency components corresponding to the interval between two adjacent cutoffs. The entire process yields... Disjoint high-frequency subbands With a low-frequency residual , ;when At this point, two non-overlapping high-frequency sub-bands and a low-frequency residual are obtained, covering the entire spectrum without overlap. To distinguish them, the two non-overlapping high-frequency sub-bands are called the high-frequency sub-band and the mid-frequency sub-band, respectively.

[0050] Subband Spatial Attention and Reconstruction: Each high-frequency subband is paired with a 3×3 convolution, and spatial attention weights are calculated from the high-frequency subband itself. : in, For the first 3×3 convolution of high-frequency subbands The sigmoid function has a coefficient of 2, making the weight range... And centered at 1, so that the corresponding sub-band can be amplified or attenuated at each spatial location.

[0051] Modulated subband for Output features It is obtained by summing the modulated high-frequency subband and the unmodulated low-frequency residual: .

[0052] Subband convolution initialization Therefore, AFFM is equivalent to identity mapping at the beginning of training, and subband modulation is learned during training to enhance the mid-to-high frequencies of the target edge and shadow boundary, and suppress the low frequencies of the reverberation and platform noise.

[0053] Acoustic-Aware Feature Aggregation (AAFA): The four-scale feature map output by the backbone simultaneously carries spatial cues from SA-SSM and frequency cues from AFFM. Standard FPN / PANet mixes these two cues using uniform element-wise addition, which dilutes the sparse frequency response. This embodiment designs AAFA in the neck structure, such as... Figure 4 At each pyramid level, this addition is replaced with a frequency-spatial cross-gating (FSCG) unit, which enables the two channels to mutually reinforce each other with adaptive weights based on their positions.

[0054] one Lateral convolution divides each Projected onto unified channel dimension Standard top-down upsampling fusion and Convolutional smoothing yields pyramid features for each stage. . For the backbone network The feature map of the stage, For the first The pyramid characteristics of the stages, in this embodiment, are taken as... .

[0055] Input Frequency-Spatial Cross-Gated Unit (FSCG). FSCG is applied independently at each level (level indices omitted below), from pyramid features. Extract two views (spatial domain branch and frequency domain branch).

[0056] In the spatial branch, the spatial view is obtained by a single layer of depth-separable 3×3 convolution followed by batch normalization: in, This is a view of the airspace. For depthwise separable 3×3 convolution, For batch normalization, It has pyramid-like features.

[0057] In the frequency domain branch, the frequency domain view is first used to generate a position-by-position filter in the spatial domain using a 1×1 convolution. , The real-valued spectral mask is obtained by taking its spectral amplitude and applying it to the sigmoid function. , And then After spectral filtering, the inverse transform is performed back to the spatial domain: in, This is a frequency domain view. It is a two-dimensional real Fourier transform.

[0058] Then, a gating is generated from each of the two views, and the gating from one view is applied to the other view (cross structure): in, It is a spatial-frequency domain cross-gated fusion feature. Gating coefficients generated for the spatial view. The gating coefficients generated for the frequency domain view. For the 1×1 convolution corresponding to the spatial view, This is the 1×1 convolution corresponding to the frequency domain view. Obtained from the spatial domain view, it is used to modulate the frequency domain characteristics, thereby increasing the frequency characteristic contribution where the spatial response is strong; The spatial features obtained from the frequency domain view are used to modulate spatial features, thereby enhancing the contribution of spatial features at the location where the acoustic structure of the spectrum is presented. As a residual term, it preserves a stable identity path for the neck region. Both gated convolutions are initialized to zero at the beginning of training. Starting with balanced mixing, cross-gating is gradually strengthened during training.

[0059] Multiscale The PANet-style bottom-up sampling path is then used to close the pyramid, resulting in multi-scale fusion features, which are then fed into the detection head, which outputs the target detection results.

[0060] Experimental data and effect verification: This embodiment verifies the performance on two sonar target detection datasets, covering two imaging modalities: forward-looking sonar and side-scan sonar. (1) UATD: a large-scale forward-looking sonar benchmark, containing 9200 range-azimuth images and 10 target categories. The official training set contains 7600 images, and the test sets Test1 and Test2 each contain 800 images. (2) MSTD: a multi-source side-scan sonar dataset, containing 835 images of four categories: people, aircraft, ships, and pipelines. The training / test sets contain 667 and 168 images respectively. The detection accuracy adopts the standard COCO style evaluation index. The main index is the mean accuracy (mAP) of the IoU threshold of 0.5~0.95 (step size 0.05), supplemented by AP@50 and AP@75. In terms of efficiency, the number of parameters (#P, millions) and the number of floating-point operations (FLOPs, billions, 800×800 input) are reported. The detector head uses Faster R-CNN, the backbone is initialized with ImageNet-1K pre-trained weights, and the optimizer is AdamW.

[0061] (1) Quantitative comparison with general target detection methods.

[0062] Table 1 presents the comparison results of our proposed method with several representative target detectors on UATD (Test1) and MSSTD. The target detectors used were DINO, RT-DETR, YOLOv12, and RF-DETR. It is evident that our proposed method achieves optimal accuracy on both sonar modes: 41.5% mAP and 86.2% AP@50 on UATD (Test1), and 64.1% mAP and 96.6% AP@50 on MSSTD, significantly outperforming the strongest general-purpose detector. Simultaneously, the number of parameters is only 27.6M, lower than various Transformer-type detectors, and the computational cost remains within a controllable range. This indicates that the accuracy improvement stems from the method design itself, rather than a larger model size.

[0063] Table 1

[0064] (2) Ablation experiments of the three core modules.

[0065] Table 2 presents the ablation results after removing the three core modules (SA-SSM, AFFM, and AAFA) one by one to verify the independent contribution of each module. w / o indicates the module removed. Using the baseline with all three modules removed as a reference, the complete model improved mAP from 38.3% to 41.5% (+3.2%) on UATD (Test1). When removing individual modules one by one, the absence of SA-SSM decreased mAP by 1.6% (contributing the most, as its axis-aligned scan matches the echo-shadow distribution of the sonar target with distance variation in ESP), the absence of AFFM decreased by 1.1% (supplementing frequency domain cues), and the absence of AAFA decreased by 1.0% (maintaining two-channel balance at the neck while improving detection and localization). The order was consistent on MSSTD. This indicates that the three modules provide complementary gains, and their synergistic use is superior to any single use.

[0066] Table 2

[0067] Experimental data can prove that: 1. Significantly improves sonar target detection accuracy. By aligning and pairing echoes and shadows along the acoustic axis using SA-SSM, separating target echoes from background reverberation using AFFM, and maintaining two-channel balance in the neck area using AAFA, this invention achieves 41.5% mAP on the forward-looking sonar benchmark UATD and 64.1% mAP on the side-scan sonar dataset MSSTD, both of which are superior to the target detectors compared, effectively reducing false detections and missed detections in complex acoustic scenarios.

[0068] 2. Balancing accuracy and computational efficiency. This invention constructs its backbone based on a state-space model with linear complexity, requiring only 27.6M parameters, lower than other Transformer-type detectors. Its additional computational load is concentrated on the three-branch scanning of SA-SSM and the frequency domain processing of AFFM / FSCG, making the overall computational cost controllable and suitable for deployment on underwater embedded platforms.

[0069] 3. Good modal transferability and robustness. Consistent performance ranking on forward-looking sonar (UATD) and side-scan sonar (MSSTD) indicates that the space-frequency design of this invention is not limited to a certain imaging geometry; both ESP gating and FSCG gating are trained from identity / equilibrium states and adaptively learned in the end-to-end process, ensuring training stability and robustness of feature extraction.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A sonar target detection method integrating directional echo shadow geometry and acoustic spectrum, characterized in that, include: Acquire sonar images; After sonar images are patch-embedding, they are input into the backbone network to obtain multi-scale feature maps; The aforementioned backbone network comprises multiple stages, each employing stacked space-frequency modules. Within each space-frequency module are, in sequence, a sonar adaptive state-space model, an acoustic frequency feature modulation model, and a position-based multilayer perceptron. The sonar adaptive state-space model aligns the one-dimensional scanning sequence with the acoustic imaging axis, and then incorporates the neighborhood of each position along the distance axis using an echo-shadow pairing operator, allowing the recursion to unfold along the direction formed by the echo and shadow. The acoustic frequency feature modulation model operates in the Fourier domain of the features, decomposing the features into disjoint frequency subbands, and then reconstructing the features. The multi-scale feature map is aggregated with acoustic sensing features through the neck structure and then sent to the detection head to output the target detection result.

2. The sonar target detection method according to claim 1, which integrates directional echo shadow geometry and acoustic spectrum, is characterized in that... The acoustic frequency characteristic modulation model is only enabled in the first two stages of the backbone network.

3. The sonar target detection method according to claim 2, which integrates directional echo shadow geometry and acoustic spectrum, is characterized in that... The working process of the sonar adaptive state-space model is as follows: The input features of the space-frequency module are divided into a feature stream and a gated stream by linear projection; Based on the gated flow, multiplicative gating is obtained; Based on the feature flow, a flattened sequence aligned along the acoustic axis is obtained; after the sequence is enhanced by the echo-shading pairing operator, it is input into the range axis forward scan branch and the range axis reverse scan branch respectively to perform range axis forward scan and range axis reverse scan; the sequence is input into the azimuth axis scan branch to perform azimuth axis scan. The sequences output by the forward scan branch of the range axis, the reverse scan branch of the range axis, and the scan branch of the azimuth axis are respectively passed through a selective state-space model to obtain the restored mesh; The average of the three restored grids is then normalized and multiplied element-wise with a multiplicative gate to obtain the output features of the sonar adaptive state-space model.

4. The sonar target detection method that integrates directional echo shadow geometry and acoustic spectrum as described in claim 3, characterized in that, The range axis is the direction in which echoes and shadows are arranged in sequence. The forward-looking sonar is along the image height, and the side-scan sonar is along the image width. The azimuth axis is orthogonal to the range axis.

5. The sonar target detection method according to claim 3, which integrates directional echo shadow geometry and acoustic spectrum, is characterized in that... The sequences output by the forward scan branch of the range axis, the reverse scan branch of the range axis, and the scan branch of the azimuth axis are restored by independent selective state-space models according to their respective scan order to obtain the restored mesh; The three independent selective state-space models each have independent parameters.

6. The sonar target detection method according to claim 3, which integrates directional echo shadow geometry and acoustic spectrum, is characterized in that... The echo-shading pairing operator encodes spatial relationships along the distance axis as a token-level one-dimensional neighborhood aggregation; The specific process is as follows: For each position, the predictor provides a Dimensional aggregation weights and a gate, The neighborhood radius on each side of the distance axis; For each channel, aggregate the token features at that location independently using aggregate weights and gating.

7. The sonar target detection method as described in claim 1, which integrates directional echo shadow geometry and acoustic spectrum, is characterized in that... The working process of the acoustic frequency characteristic modulation model is as follows: Perform an orthogonally normalized two-dimensional real Fourier transform on the input feature map, and then perform cascaded subband separation to obtain multiple non-overlapping high-frequency subbands and a low-frequency residual. The spatial attention weights are calculated from the high-frequency subband itself, and the high-frequency subband is modulated to obtain the modulated subband; The output characteristics of the acoustic frequency characteristic modulation model are obtained by calculating the sum of the low-frequency residual and all modulated subbands.

8. The sonar target detection method according to claim 7, which integrates directional echo shadow geometry and acoustic spectrum, is characterized in that... The cutoff parameters used for cascaded subband separation are 2 and 4; cascaded subband separation yields two non-overlapping high-frequency subbands and a low-frequency residual.

9. The sonar target detection method according to claim 1, which integrates directional echo shadow geometry and acoustic spectrum, is characterized in that... Each pyramid level of acoustic sensing feature aggregation employs a frequency-space cross-gating unit, enabling the two channels to mutually reinforce each other with adaptive positional weights. The frequency-space cross-gating unit extracts spatial and frequency domain views from the pyramid features; Generate a gate from both the spatial domain view and the frequency domain view, and apply it to the other view; Then, summing the sum with the pyramid features yields the spatial-frequency cross-gated fusion features, calculated using the following formula: in, It is a spatial-frequency domain cross-gated fusion feature. Gating coefficients generated for the spatial view. The gating coefficients generated for the frequency domain view. This is a view of the airspace. This is a frequency domain view. Features of a pyramid This is for element-wise multiplication.

10. The sonar target detection method according to claim 1, which integrates directional echo shadow geometry and acoustic spectrum, is characterized in that... The patch-embedding consists of two 3×3 convolutions with a stride of 2; The backbone network has 4 stages; The detection head is an anchor-frame based detection head.

Images