Unified detection method for heterogeneous sonar based on physical alignment representation learning

Abstract

The application relates to the technical field of deep learning, and discloses a heterogeneous sonar unified detection method based on physical alignment representation learning, which acquires a forward-looking sonar image and a side-scan sonar image; the forward-looking sonar image is input into a distance condition echo normalization module for processing, and normalized features are output; the normalized features and the unprocessed side-scan sonar image are jointly input into a shared backbone network containing a sparse expert routing module to extract hierarchical features, and multi-scale branch features are obtained through a feature pyramid; the side-scan branch features are input into a strip context attention module for processing; after a double-layer calibration of the forward-looking branch features and the side-scan features after the strip interference is suppressed through a synchronization mechanism, the results are input into corresponding detection heads to output results. The application effectively overcomes the physical and reliability mismatch, and significantly improves the underwater target cooperative detection precision, the generalization ability and the model convergence stability.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a unified detection method for heterogeneous sonar based on physical alignment representation learning. Background Technology

[0002] Sonar-based underwater sensing technologies are becoming increasingly important in various marine applications of autonomous underwater vehicles (AUVs), such as underwater exploration, search and rescue operations, and seabed mapping. Forward-looking sonar (FLS) generates real-time two-dimensional acoustic images by emitting a fan-shaped beam forward of the vehicle, providing crucial situational awareness capabilities such as obstacle avoidance and target tracking. Side-scan sonar (SSS) generates high-resolution seabed images through scanning perpendicular to the vehicle's trajectory, making it particularly suitable for large-area surveys. The two technologies have strong complementary advantages.

[0003] There has been much research on sonar-based underwater target detection algorithms. Early and existing methods were mainly designed for single sonar modes.

[0004] In the field of FLS detection, Valdenegro-Toro et al. were the first to apply convolutional neural networks to target scoring of FLS images; subsequent researchers have further explored multi-scale feature extraction and Transformer-based architectures to improve the accuracy of FLS detection under complex acoustic conditions.

[0005] In the interpretation of SSS images, Huo et al. introduced deep transfer learning using semi-synthetic training data for SSS target classification; Tang et al. and Rajani et al. respectively applied semantic segmentation methods and convolutional visual Transformers to SSS data analysis to achieve pixel-level understanding of seabed structures.

[0006] To address the negative transfer problem in multi-task or multi-modal joint modeling, some scholars have introduced multi-task learning and domain adaptation techniques into the field of object detection. For example, Chen et al. proposed the GradNorm gradient normalization algorithm, which automatically balances training by dynamically adjusting gradient magnitudes across tasks; Kendall et al. proposed an uncertainty-based loss weighting method to adaptively adjust task weights. Regarding unified multimodal detection frameworks, Li et al. proposed the unified model SM3Det for multimodal remote sensing object detection, while Wang et al. introduced the unified and efficient multimodal Transformer (UniTR) for bird's-eye view representation.

[0007] Despite some progress made by the above methods, significant limitations still exist in practical applications of heterogeneous sonar collaboration.

[0008] First, most existing methods are limited to a single sonar mode, treating FLS and SSS as independent domains and adopting their own independent model architectures and training processes. This not only increases development and maintenance costs, but also fails to fully utilize the synergistic effects between heterogeneous sonar data.

[0009] Secondly, simply mixing FLS and SSS data for joint training will encounter a serious negative transfer problem. The fundamental challenge lies in the two having drastically different physical imaging degradation patterns (i.e., system heterogeneity): FLS exhibits "distance-dependent echo attenuation" due to propagation loss, resulting in a non-stationary amplitude distribution; while SSS exhibits "strip structured interference" tightly coupled with platform motion and acquisition geometry. Forcing a shared backbone network to simultaneously fit these two conflicting degradation patterns often leads to feature extraction biased towards shortcuts or the generation of spurious activations.

[0010] Furthermore, existing multi-task or domain adaptive unified detection methods (such as loss-weighted and gradient-operated methods) typically assume that modality-specific variations can be implicitly learned by the network, failing to explicitly model and compensate for the inherent physical imaging degradation mechanisms of each sonar system. Simultaneously, existing methods also fail to effectively address the problem of "uneven supervision reliability" caused by differences in noise levels, residual artifacts, and data scale among different modal branches. Summary of the Invention

[0011] This invention aims to at least solve one of the technical problems existing in related technologies. To this end, this invention provides a unified detection method for heterogeneous sonar based on physical alignment representation learning.

[0012] A unified detection method for heterogeneous sonar based on physical alignment representation learning, comprising the following steps: Acquire forward-looking sonar images and side-scan sonar images; The forward-looking sonar image is input into the range-conditional echo normalization module for physical alignment processing, and the normalized forward-looking sonar features are output. The normalized forward-looking sonar features and the side-scan sonar images are input into a shared backbone network to extract hierarchical features, and multi-scale pyramid features are obtained through a multi-scale feature pyramid, which includes forward-looking sonar branch features and side-scan sonar branch features. The shared backbone network includes a sparse expert routing module. The side-scan sonar branch features are input into the strip context attention module for physical alignment processing, and the side-scan sonar features after suppressing strip interference are output. After performing feature layer calibration using a synchronization mechanism on the forward-looking sonar branch features and the side-scan sonar features after suppressing strip interference, the corresponding detection heads are input respectively, and the unified detection results of the forward-looking sonar and the side-scan sonar are output.

[0013] Further, the step of physically aligning the forward-looking sonar image input to the range-conditional echo normalization module and outputting the normalized forward-looking sonar features includes: The forward-looking sonar image is subjected to signed logarithmic compression to suppress extreme outliers, resulting in compressed features; Mean pooling is performed along the orientation dimension of the compressed feature to extract a one-dimensional distance profile; Based on the one-dimensional distance profile, a bounded distance normalization factor is calculated using a lightweight predictor. The distance normalization factor is broadcast along the azimuth dimension and multiplied element-wise with the compressed feature to obtain the normalized forward-looking sonar feature. The lightweight predictor comprises two one-dimensional convolutional networks and an intermediate activation layer, and outputs a bounded calibration signal through the sigmoid activation function at the end to calculate the distance normalization factor.

[0014] Furthermore, the process of extracting hierarchical features from the shared backbone network in conjunction with the sparse expert routing module includes: The backbone network extracts intermediate feature maps; Flatten the spatial dimensions of the intermediate feature map into a token sequence; The similarity score between each token feature in the token sequence and the learnable expert embedding matrix is ​​calculated using a router network. Based on the similarity scores, the top K expert networks with the highest scores are selected for activation; Where K is the preset quantity.

[0015] Furthermore, after selecting and activating a preset number of expert networks with the highest similarity scores, the process further includes: Normalize the routing weights of the selected expert networks; Each token feature is input into the corresponding activated expert network for parallel processing, and the features output by each expert network are weighted and summed according to the normalized routing weights to obtain a hybrid feature recombined output.

[0016] Furthermore, during the model training phase, the sparse expert routing module also introduces load balancing regularization. The load balancing regularization calculates the global cumulative importance and global call load of all expert networks, and then calculates the squared variation coefficient among all expert networks. The squared variation coefficient is then added as a penalty term to the overall training objective to prevent route allocation collapse.

[0017] Furthermore, the overall loss function corresponding to the overall training objective is composed of the detection loss of all branches plus the squared variation coefficient penalty term scaled by a proportional coefficient, and is used to optimize network parameters end-to-end.

[0018] Further, the step of physically aligning the side-scan sonar branch features into the strip context attention module and outputting side-scan sonar features after suppressing strip interference includes: Mean pooling is performed along the cross-track direction of the side-scan sonar branch features to obtain strip-level descriptors; The strip-level descriptors are rearranged along the channel dimension to obtain the channel description vector for each strip; The channel gating weights for each strip are predicted using a multilayer perceptron network based on the channel description vectors. The channel gating weights are broadcast along the cross-track direction and multiplied element-wise with the original side-scan sonar branch features to obtain the side-scan sonar features after suppressing strip interference.

[0019] Furthermore, the feature layer calibration using a synchronization mechanism includes: For the forward-looking sonar branch features and the side-scan sonar features after suppressing strip interference, the mean values ​​are calculated in batch size, spatial dimension and pyramid dimension, respectively, and channel descriptors for the corresponding branches are constructed. Map the channel descriptors of each branch to the scalar reliability scores of their respective branches; Calculate the average reliability score for all branches, and apply near-identical residual scaling based on the ratio of the scalar reliability score to the average reliability score to calibrate the features corresponding to each branch; The step of mapping the channel descriptors of each branch to the scalar reliability scores of their respective branches specifically involves using a multilayer perceptron containing hidden dimensions and activation functions to map the high-dimensional channel descriptors to the scalar reliability scores.

[0020] Furthermore, during the model training phase, the synchronization mechanism also includes optimization layer synchronization processing, which includes: Track the loss terms of different detector heads in each training iteration, and maintain the historical loss evolution value of each loss term using an exponential moving average method; Based on the instantaneous loss value of the current iteration and the corresponding historical loss evolution value, the final optimized weight of each loss term is calculated through the inverse loss score, which is used to adjust the learning rate of the detection head corresponding to each branch.

[0021] Furthermore, the optimization layer synchronization processing also includes learning rate adjustment of the shared backbone network: Calculate the divergence between the current loss distribution and the historical loss distribution; Based on the divergence and the preset bias term, the backbone learning rate of the shared backbone network is dynamically calculated and adjusted. When the current loss distribution deviates significantly from the historical trend, the backbone learning rate is reduced; when the training trend is consistent, the backbone learning rate is increased.

[0022] 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. By introducing “Range Conditional Echo Normalization (RCEN)” and “Strip Contextual Attention (SCA)” in the early and late stages of the network respectively, the range offset of forward-looking sonar and the strip interference of side-scan sonar are specifically compensated, eliminating the underlying physical imaging mechanism conflict between the two modes and significantly improving the accuracy of heterogeneous data collaborative detection (overcoming negative migration).

[0023] 2. Through the coordinated efforts of physical alignment and a unified learning mechanism at the representation level, this model can effectively extract complementary features of dual sonars even on data-limited or small-scale datasets, comprehensively surpassing independently trained single-system baseline models, thus enabling the model to have strong data generalization and small-sample adaptability.

[0024] 3. By introducing "Sparse Expert Routing (SMR)," the network feature space capacity can be allocated on demand. At the same time, the "Reliability Synchronization Mechanism (RSM)" is used to perform global dynamic adjustment at the two levels of feature amplitude and learning rate gradient, which effectively avoids unreliable supervision signal pollution caused by noise in each branch and uneven data scale, thus ensuring the stability of training.

[0025] 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

[0026] 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.

[0027] Figure 1 This is a diagram illustrating the overall architecture of the method of the present invention.

[0028] Figure 2 This is a structural diagram of the Conditional Echo Normalization (RCEN) module.

[0029] Figure 3 This is a structural diagram of the Striped Contextual Attention (SCA) module.

[0030] Figure 4 This is a flowchart of the Sparse Expert Routing (SMR) process. Detailed Implementation

[0031] 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. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0032] This invention addresses the unified target detection problem of heterogeneous sonars (FLS and SSS), taking into account the two major challenges of physical mismatch and reliability mismatch in cross-system fusion, and proposes a unified detection method for heterogeneous sonars (Omni-Sonar) based on physical alignment representation learning.

[0033] This invention designs range-conditional echo normalization (RCEN) and strip contextual attention (SCA) to explicitly compensate for predictable imaging degradation, and combines sparse expert routing (SMR) and reliability synchronization mechanism (RSM) to achieve stable and high-precision cross-system unified representation learning and collaborative detection.

[0034] like Figure 1 The figure shows the overall architecture of the method of the present invention.

[0035] This invention proposes a two-level alignment unified detection method (Omni-Sonar) that considers both physical mismatch and reliability mismatch. Its core steps are "physical alignment first, then unified learning".

[0036] The main steps are as follows: 1. Establish a physical alignment layer to explicitly compensate for system-specific imaging degradation. Physical alignment adapters are inserted for FLS and SSS respectively to resolve predictable distribution shifts caused by different imaging geometries before shared learning.

[0037] 1.1 Range-conditional echo normalization (RCEN) for FLS like Figure 2 As shown, the structure of the Conditional Echo Normalization (RCEN) module is illustrated.

[0038] The echoes in an FLS image at different distances follow different amplitude distributions, introducing a systematic bias. RCEN explicitly models this attenuation characteristic: assuming... Input image for FLS ( For the number of channels, The distance to bin is the number of bins. (This refers to the directional dimension).

[0039] First, sign logarithmic compression is performed to suppress extreme outliers: (1) The 1D range profile is obtained by calculating the average echo amplitude of each range bin: (2) in, Reflects a rough distance condition bias; c, h, and w are the index identifiers for the summation symbol, corresponding to the traversal indexes of dimensions C, H, and W, respectively; c is the summation index for the channel dimension; h is the index for the distance dimension; and w is the summation index for the orientation dimension.

[0040] based on Build a lightweight predictor ; Lightweight predictor Composed of two layers of 1D convolution (kernel size) Hidden passage The activation functions consist of the intermediate ReLU activation function and the final Sigmoid activation function. Output a bounded calibration signal.

[0041] g(h) is a bounded calibration signal, calculated using the following formula: (3) The conservative distance normalization factor is: (4) in Control the calibration intensity. Because... ,have Avoid over-amplification.

[0042] Normalize the broadcast to all pixels at the same distance: (5) Normalized output Feed it into the shared backbone and optimize end-to-end by detecting the loss. The computational overhead introduced by RCEN is negligible. 1D convolution plus broadcast multiplication on length vectors), when It degenerates into an identity mapping.

[0043] It is a three-dimensional tensor, which is the overall representation of the three-dimensional tensor under all c, h, w indices.

[0044] 1.2 Strip Context Attention (SCA) for SSS like Figure 3 As shown, the structure of the striped context attention (SCA) module is illustrated.

[0045] SSS performs imaging by emitting acoustic pulses perpendicular to the platform's trajectory: the height dimension of the feature map corresponds to the direction of the trajectory (platform motion). Width dimension corresponding to cross-track direction (scan width) .

[0046] Strip artifacts (including banded ripples and beam pattern undulations) exhibit strong consistency along the cross-track direction and persist across multiple scales, potentially triggering repetitive striped activations in the characteristic response. To suppress such interference, we propose Striped Contextual Attention (SCA), a lightweight feature domain calibration module.

[0047] SCA is inserted after the FPN / neck output to perform strip conditional calibration on the multi-scale pyramid features.

[0048] set up For SSS at the pyramid level The feature map, where To determine the batch size, SCA first performs mean pooling along the width dimension (i.e., across the track direction, where stripe consistency is stronger) to obtain strip-level descriptors: (6) in, Aggregate each stripe index Channel statistics.

[0049] B represents the batch size, which is the number of sample images in a training batch. C represents the number of channels, specifically the number of convolutional channels in the feature map. For the first The height dimension of the layer feature map defines the stripe index dimension; For the first The width dimension of the layer feature map defines the cross-track direction; b: Batch index, c: Channel index, h: Strip height index, w: Width index.

[0050] This convergence emphasizes the overall consistency of the stripes while downplaying isolated target responses. Rearrange along the channel dimension to obtain the channel description vector for each strip. .

[0051] Drawing inspiration from the squeeze-incentive paradigm, with reduction ratio (This article is under implementation) The two-layer MLP gates each strip prediction channel: (7) in , For learnable weight matrix, It is the ReLU activation function. This is the sigmoid activation function.

[0052] (b,h,:) is the tensor slice index, with a fixed batch b and a fixed strip index h. Iterates through all C channels, i.e., the b-th sample and the h-th strip, with the independent attention weight corresponding to each channel.

[0053] The gating weights are broadcast along the width dimension and multiplied element by element: (8) SCA acts independently on each pyramid level while sharing parameters, consistently suppressing stripe structured interference across multiple scales. (Calibrated features) It is then sent to subsequent modules for testing.

[0054] II. Constructing a unified representation learning layer and performing adaptive capacity allocation After physical alignment, hierarchical features are extracted using a shared backbone network. To handle residual heterogeneity, sparse expert routes (SMR) are embedded in the backbone to perform conditional computation on different local acoustic statistics.

[0055] like Figure 4 The diagram illustrates the Sparse Expert Routing (SMR) process.

[0056] 2.1 Token-level Sparse Expert Hybrid set up For intermediate feature maps, where For batch size, For the number of channels, For spatial dimensions. Flatten the space into... The token is recorded as the first one. Each sample is indexed in space. The token characteristics at the location are SMR maintenance Expert feedforward network Each represents a nonlinear operator corresponding to a class of local acoustic statistics. The router employs cosine similarity gating: token features are projected onto a low-dimensional space via a learnable linear mapping, and route scoring is performed. It is calculated from the temperature-scaled cosine similarity between the projected token and the learnable expert embedding matrix.

[0057] Gaussian noise is injected into logits during training to promote load diversity. The top-scoring logits are selected. expert: (9) Where Top-K is the filtering function, and MoE is the standard operator for sparse routing. b represents the batch sample; t is the current token; e is expert number e; Hyperparameter K: The number of experts activated by each token (a core hyperparameter for sparsity, K≪E, enabling sparse computation). Operation rule: From all E experts, select the top K experts with the highest routing scores sb,t,e, and store their numbers in a set. .

[0058] Effect: Achieve sparse activation - each token will only call K out of the E global experts, significantly reducing the amount of computation.

[0059] Routing weight Normalization: (10) The normalized routing hybrid weights are for the b-th sample, the t-th token, and the e-th expert. b, t, and e represent the numerical constraints of the mixed weights (gating weights) for the b-th sample, the t-th token, and the e-th expert, respectively. j: Iterate through the set of all selected experts for the current token. All expert numbers within.

[0060] right make The output of MoE is: (11) in: Let be the final output feature vector of the b-th sample and the t-th token after SMR sparse expert mixing.

[0061] This format is suitable for sonar images: token-level routing allows different regions (targets, shadows, background textures, strong scattering points) within the same image to be processed by specialized operators, capturing local nonstationarity.

[0062] 2.2 Load Balancing Regularization To prevent route collapse (where most tokens are assigned to a few experts), we introduce load balancing.

[0063] Defining expert importance and workload: (12) The cumulative importance of expert number e in the current batch; This represents the call load for expert e in the current batch; This is an indicator function; 1 indicates that the condition is true, and 0 indicates that the condition is false. Select the Top-K expert set for the t-th token of the b-th sample.

[0064] For each expert, equation (12) e =1, …, E Obtain scalar and They were pieced together according to expert dimensions. E 3D vector Imp = [Imp1, …, Imp] E ] T Load=[Load1, …, Load E ] T Therefore, the coefficient of variation among punishment experts can be expressed as: (13) For any E dimensional vector v, The coefficient of variation is the squared variation. CV is the numerical stability constant. 2 (Imp) and CV 2 (Load) is calculated on this vector to constrain load balancing among experts.

[0065] The overall goal is: (14) in, To detect the loss, Controlling the load balancing strength (set in the experiment of this invention) ).

[0066] SMR improves the characterization capacity of heterogeneous echo statistics, but does not address the reliability imbalance between branches. Therefore, we introduce a reliability synchronization mechanism (RSM) to align branch contributions between the feature layer and the optimization layer.

[0067] III. Reliability Synchronization (RSM) Based on Loss Dynamics and Feature Contributions Due to differences in data noise levels and dataset size across different systems, issues arise regarding the reliability of supervision. This invention addresses this by implementing synchronization at the feature layer and the optimization layer, respectively.

[0068] 3.1 Feature Layer Calibration (RSM-F) Suppose that in one training iteration there are There are 1 valid branch. (Regarding the branch...) Let its multi-scale FPN features be denoted as ,in The number of pyramid levels. RSM-F first calculates the mean of the batch, spatial, and pyramidal dimensions, and then constructs a channel descriptor for each branch: (15) Hidden dimension is Lightweight MLP will Mapped to scalar reliability score: (16) in, Let m be the reliability score of the m-th branch.

[0069] MLP(·) is a lightweight multilayer perceptron with a hidden layer dimension of C / 4, used to map high-dimensional channel descriptors dm to scalars.

[0070] This is the sigmoid activation function.

[0071] Calculate average reliability Apply near-identical residual scaling: (17) in, For the m-th branch, The feature map after hierarchical calibration has the same dimensions as the original feature map. Consistent; The average reliability score for all branches; For a smaller scaling factor (typical value) ).

[0072] Inspired by the near-identity property of residual connections, this design preserves the characteristic geometry while synchronizing branch contributions: relatively more reliable branches are slightly amplified, while unreliable branches are flexibly suppressed. During single-branch inference, the reliability of the current branch is used. Perform self-centered scaling: (18) 3.2 Optimization Layer Synchronization (RSM-O) Feature layer alignment alone is insufficient to address the inconsistent evolution rates of multi-branch loss. RSM-O adjusts the learning rate based on the dynamic implicit reliability of the loss. Similar to the idea of ​​dynamic weight averaging but with finer granularity, RSM-O tracks the evolution of each term of the loss. Let the current iteration be... The loss items are (e.g., classification and regression losses for different detection heads), with update rate Maintenance Index Moving Average (EMA): (19) The historical exponential moving average (EMA) value of the k-th loss term is used to smooth the loss changes over multiple historical iterations. Let be the instantaneous loss value of the k-th loss term in the current iteration; k is the number of a single loss term, and the iteration is k∈{1,2,…,N}. Define inverse loss score Through temperature Convert to normalized weights: (20) The final optimized weight for the k-th loss term; τ is the Softmax temperature coefficient, τ>0, which controls the sharpness of the weight distribution.

[0073] The loss term is grouped by sub-network, and the learning rate multiplier is obtained from the average of the associated weights. Therefore, the unstable loss term ( relatively Larger targets receive smaller weights to avoid unreliable targets over-driving updates.

[0074] For shared backbones, RSM-O measures the difference between the current and historical loss distributions: (twenty one) divergence .

[0075] The core learning rate multiplier is: (twenty two) in This is a bias term used to control the benchmark ratio. When the loss distribution deviates significantly from the historical trend ( When (increases), Reduce to stabilize shared representations; training consistency ( When it is smaller, Improvements are made to accelerate convergence.

[0076] This patent evaluates Omni-Sonar on four sonar target detection benchmark datasets, covering both FLS and SSS sonar systems.

[0077] (1) UATD is a large-scale FLS underwater acoustic target detection dataset containing 9,200 images and 10 target categories. The images are provided in the form of raw echo intensity data, represented in range-azimuth, where the vertical axis and horizontal axis correspond to the range cell and azimuth beam, respectively; (2) SubPipe is the SSS submarine pipeline detection dataset, which contains 4,890 images and a single category (pipeline). (3) MDFLS is a small-scale multi-domain FLS dataset containing 1,868 images and 10 object categories. Unlike UATD, its images are scanned and converted into a fan-shaped representation, presenting fan-shaped images in Cartesian coordinates; (4) SCTD is the SSS classification and object detection dataset. The original dataset contains four categories: person, pipeline, plane, and ship, and contains 488 images.

[0078] This patent employs standard COCO-style evaluation metrics. The primary metric is the mean average precision (mAP), calculated by averaging AP values ​​across IoU thresholds from 0.5 to 0.95 (in steps of 0.05). This patent also reports AP50, AP75, and the overall mAP across FLS and SSS regimes to evaluate uniform detection performance. Efficiency analysis reports the number of model parameters (abbreviated as #P, in millions) and floating-point operations (FLOPs, in billions), with FLOPs calculated using 640×640 input.

[0079] Table 1. Quantitative comparison with existing methods on large-scale datasets.

[0080] Table 2 Quantitative comparison with existing methods on small-scale datasets

[0081] Table 1 presents the quantitative comparison results on large-scale datasets (UATD and SubPipe). Omni-Sonar achieved the best overall performance among unified methods (mAP 55.4%, AP50 96.2%, AP75 55.6%), improving upon the strongest comparison method UniTR by 1.0 percentage point. This advantage is mainly attributed to the specific degradation of explicit modeling mechanisms in RCEN and SCA, modality-adaptive feature extraction achieved by SMR, and simultaneous cross-branch learning dynamics in RSM to suppress negative transfer.

[0082] Table 2 shows the results on small datasets (MDFLS and SCTD). The limited training samples on small datasets place higher demands on the model's generalization ability. Omni-Sonar achieved the best performance, with an mAP of 85.9% on MDFLS and 59.9% on SCTD, for an overall mAP of 79.9%, surpassing other unified methods. This demonstrates that physically aligned representation learning can effectively utilize complementary information even under conditions of limited data.

[0083] We conducted ablation experiments on a large-scale (UATD + SubPipe) dataset to analyze the contribution of each module.

[0084] Table 3 Ablation experiments on module effectiveness (✓ indicates that the module is enabled)

[0085] Removing RCEN resulted in a 0.9% decrease in UATD while slightly improving SubPipe, confirming that RCEN specifically addresses the issue of distance-related intensity variations in FLS. Removing SCA caused a 0.4% decrease in SubPipe, validating its effectiveness in suppressing SSS stripe interference. Removing SMR resulted in the largest decrease in SubPipe (-1.8%), demonstrating its crucial role in multi-source feature routing. Removing RSM caused a decrease in performance on both datasets, highlighting the importance of adaptive training optimization. The complete model achieved the best overall performance, validating the complementary design of all modules.

[0086] Compared with existing underwater target detection algorithms based on a single sonar mode and conventional multi-task joint learning methods, the algorithm proposed in this invention innovatively considers the system heterogeneity problem between forward-looking sonar (FLS) and side-scan sonar (SSS), and constructs a unified detection framework across systems.

[0087] To address the conflict between negative transfer and degradation patterns arising from direct mixed training, this invention does not rely solely on implicit learning but instead designs a principled two-level alignment strategy: At the physical layer, Range Conditional Echo Normalization (RCEN) and Strip Contextual Attention (SCA) modules are designed to explicitly and specifically compensate for the distance-related attenuation of FLS and the strip structured interference of SSS; at the representation learning layer, Sparse Expert Routing (SMR) is introduced for conditional capacity allocation of the feature space, and a synchronization mechanism, the Reliability Synchronization Mechanism (RSM), is designed to align the supervision contributions of different sonar branches at both the feature and gradient levels, thereby stabilizing cross-system optimization. Final experimental results show that the algorithm designed in this invention effectively overcomes negative transfer on both large-scale and small-scale datasets, achieving high-precision collaborative detection of heterogeneous sonar data.

[0088] This invention proposes a principled two-level aligned cross-system unified detection framework, overcoming the negative transfer defect caused by direct data mixing in traditional multi-task learning. The main beneficial effects include: 1. Significantly improves collaborative detection accuracy for heterogeneous data: By explicitly correcting the distance-induced bias of FLS using RCEN and attenuating the strip structure interference of SSS using SCA, physical mismatch is eliminated before shared learning. In joint training tests on large-scale datasets (UATD and SubPipe), the overall mAP reaches 55.4%, which is 1.0 percentage point higher than the current most powerful multimodal unified detection network UniTR, effectively reducing false positives and false negatives in challenging scenarios.

[0089] 2. Strong data generalization and small sample adaptation ability: On small datasets (MDFLS and SCTD), this invention achieved an overall mAP of 79.9%, which comprehensively surpasses the combination of independently trained single-system SOTA models (79.1%), proving that the physical alignment and representation learning mechanism can effectively extract complementary features when data is limited.

[0090] 3. Stable and efficient model convergence: SMR enables on-demand allocation of parameter capacity, and RSM dynamically balances the reliability of supervision at both the feature magnitude and learning rate gradient levels, avoiding the dominance of unreliable noisy data on the shared network, thus ensuring the stability of training and the robustness of network feature extraction.

[0091] 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 unified detection method for heterogeneous sonar based on physical alignment representation learning, characterized in that the steps are as follows: include: Acquire forward-looking sonar images and side-scan sonar images; The forward-looking sonar image is input into the range-conditional echo normalization module for physical alignment processing, and the normalized forward-looking sonar features are output. The normalized forward-looking sonar features and the side-scan sonar images are input into a shared backbone network to extract hierarchical features, and multi-scale pyramid features are obtained through a multi-scale feature pyramid, which includes forward-looking sonar branch features and side-scan sonar branch features. The shared backbone network includes a sparse expert routing module. The side-scan sonar branch features are input into the strip context attention module for physical alignment processing, and the side-scan sonar features after suppressing strip interference are output. After performing feature layer calibration using a synchronization mechanism on the forward-looking sonar branch features and the side-scan sonar features after suppressing strip interference, the corresponding detection heads are input respectively, and the unified detection results of the forward-looking sonar and the side-scan sonar are output. The step of physically aligning the forward-looking sonar image input to the range-conditional echo normalization module and outputting the normalized forward-looking sonar features includes: The forward-looking sonar image is subjected to signed logarithmic compression to suppress extreme outliers, resulting in compressed features; Mean pooling is performed along the orientation dimension of the compressed feature to extract a one-dimensional distance profile; Based on the one-dimensional distance profile, a bounded distance normalization factor is calculated using a lightweight predictor. The distance normalization factor is broadcast along the azimuth dimension and multiplied element-wise with the compressed feature to obtain the normalized forward-looking sonar feature. The lightweight predictor comprises two one-dimensional convolutional networks and an intermediate activation layer, and outputs a bounded calibration signal through the sigmoid activation function at the end to calculate the distance normalization factor. The step of inputting the side-scan sonar branch features into the strip context attention module for physical alignment processing and outputting side-scan sonar features after suppressing strip interference includes: Mean pooling is performed along the cross-track direction of the side-scan sonar branch features to obtain strip-level descriptors; The strip-level descriptors are rearranged along the channel dimension to obtain the channel description vector for each strip; The channel gating weights for each strip are predicted using a multilayer perceptron network based on the channel description vectors. The channel gating weights are broadcast along the cross-track direction and multiplied element-wise with the original side-scan sonar branch features to obtain the side-scan sonar features after suppressing strip interference.

2. The heterogeneous sonar unified detection method based on physical alignment representation learning according to claim 1, characterized in that, The process of extracting hierarchical features using the shared backbone network combined with the sparse expert routing module includes: The backbone network extracts intermediate feature maps; Flatten the spatial dimensions of the intermediate feature map into a token sequence; The similarity score between each token feature in the token sequence and the learnable expert embedding matrix is ​​calculated using a router network. Based on the similarity scores, the top K expert networks with the highest scores are selected for activation; Where K is the preset quantity.

3. The heterogeneous sonar unified detection method based on physical alignment representation learning according to claim 2, characterized in that, After selecting and activating a preset number of expert networks with the highest similarity scores, the process further includes: Normalize the routing weights of the selected expert networks; Each token feature is input into the corresponding activated expert network for parallel processing. The features output by each expert network are then weighted and summed according to the normalized routing weights to obtain a hybrid feature recombined output.

4. The heterogeneous sonar unified detection method based on physical alignment representation learning according to claim 3, characterized in that, During the model training phase, the sparse expert routing module also introduces load balancing regularization; The load balancing regularization calculates the global cumulative importance and global call load of all expert networks, and then calculates the squared variation coefficient among all expert networks. The squared variation coefficient is then added as a penalty term to the overall training objective to prevent route allocation collapse.

5. The heterogeneous sonar unified detection method based on physical alignment representation learning according to claim 4, characterized in that, The overall loss function corresponding to the overall training objective is composed of the detection loss of all branches plus a squared coefficient of variation penalty term scaled by a proportional coefficient, and is used to optimize network parameters end-to-end.

6. The heterogeneous sonar unified detection method based on physical alignment representation learning according to claim 1, characterized in that, The feature layer calibration using a synchronization mechanism includes: For the forward-looking sonar branch features and the side-scan sonar features after suppressing strip interference, the mean values ​​are calculated in batch size, spatial dimension and pyramid dimension, respectively, and channel descriptors for the corresponding branches are constructed. Map the channel descriptors of each branch to the scalar reliability scores of their respective branches; Calculate the average reliability score for all branches, and apply near-identical residual scaling based on the ratio of the scalar reliability score to the average reliability score to calibrate the features corresponding to each branch; The step of mapping the channel descriptors of each branch to the scalar reliability scores of their respective branches specifically involves using a multilayer perceptron containing hidden dimensions and activation functions to map the high-dimensional channel descriptors to the scalar reliability scores.

7. The heterogeneous sonar unified detection method based on physical alignment representation learning according to claim 1, characterized in that, During the model training phase, the synchronization mechanism also includes optimization layer synchronization processing, which includes: Track the loss terms of different detector heads in each training iteration, and maintain the historical loss evolution value of each loss term using an exponential moving average method; Based on the instantaneous loss value of the current iteration and the corresponding historical loss evolution value, the final optimized weight of each loss term is calculated through the inverse loss score, which is used to adjust the learning rate of the detection head corresponding to each branch.

8. The heterogeneous sonar unified detection method based on physical alignment representation learning according to claim 7, characterized in that, The optimization layer synchronization process also includes learning rate adjustment for the shared backbone network: Calculate the divergence between the current loss distribution and the historical loss distribution; Based on the divergence and the preset bias term, the backbone learning rate of the shared backbone network is dynamically calculated and adjusted.

Images