Sonar image retrieval method and system based on deep wavelet convolution Hash network

Abstract

The invention discloses a deep wavelet convolution Hash network-based sonar image retrieval method and system, and the method comprises the steps: inputting a to-be-retrieved sonar image into a trained deep wavelet convolution Hash network, and obtaining a retrieval result; the deep wavelet convolution Hash network comprises an image preprocessing module, a shared weight network module and a hybrid quantization optimization module; the image preprocessing module is used for carrying out image equalization and image enhancement and outputting a weak enhancement teacher view and a strong enhancement student view; the shared weight network module comprises a wavelet aggregation backbone network and a hierarchical adaptive fusion module, and the hybrid quantization optimization module configures a total loss function and enables the total loss function to be optimal through training of a training sample; and when a to-be-retrieved sonar image is input into the trained deep wavelet convolution Hash network, performing symbol conversion on continuous embedding output by the shared weight network module to obtain a retrieval result. The method has the advantages that the method is obviously superior to an existing advanced method in the aspects of retrieval precision and calculation efficiency.

Description

Technical Field

[0001] This invention relates to sonar image processing technology, specifically to a sonar image retrieval method and system based on a deep wavelet convolutional hash network. Background Technology

[0002] Underwater acoustic imaging (sonar) is a primary sensing modality in turbid underwater environments, enabling exploration and monitoring tasks under conditions of severely limited optical visibility. Recent advances in deep learning have significantly propelled the development of underwater visual analytics tasks, including target classification, target detection, and semantic segmentation. However, with the explosive growth of sonar data acquired by autonomous underwater vehicles (AUVs), efficiently retrieving relevant images from massive databases remains a key bottleneck. This makes sonar image retrieval a pressing task, and compared to mature optical retrieval, the field is still in its infancy. In general visual search scenarios, traditional image retrieval systems typically rely on high-dimensional real-valued feature vectors for similarity measurement. However, performing accurate nearest neighbor searches in such high-dimensional spaces incurs extremely high computational and storage costs, limiting its scalability across large-scale acoustic archives. To address this issue, deep hashing has emerged as a mainstream paradigm. By mapping high-dimensional data to compact binary code, it utilizes efficient bitwise Hamming distance operations to replace expensive Euclidean distance calculations, thereby achieving real-time approximate nearest neighbor searches.

[0003] Despite the rapid development of deep hashing in optics, existing retrieval methods still face three key challenges when applied to sonar images due to the fundamental differences in acoustic imaging mechanisms. First, the significant noise interference and weak contrast of sonar images severely hinder the effective representation of features. Sonar images are not only affected by distance-dependent attenuation and reverberation but are also riddled with severe multiplicative speckle noise. Standard spatial convolutional networks often struggle to distinguish effective texture from high-frequency noise, easily aliasing noise patterns into semantic representations. In contrast, frequency-domain perceptual learning has shown great potential in optical vision due to its advantages in capturing global structure. This paradigm is theoretically more suitable for underwater acoustic environments: since sonar images are essentially mappings of wave signals, their effective structural semantics (usually concentrated in low-frequency components) and multiplicative speckle noise (dominantly high-frequency components) are naturally separable in the frequency domain. However, despite the introduction of spectral analysis, existing frequency-domain methods often neglect domain-specific preprocessing, directly inputting low-contrast raw data into the frequency analysis module. Without enhancement, inherent noise artifacts are often amplified rather than suppressed during frequency decomposition, thus masking subtle target details. Secondly, the unique imaging geometry of forward-looking sonar leads to severe scale blurring. Unlike optical cameras with consistent perspective projection, the scale of a target in a sonar image is determined by its imaging distance, exhibiting significant dynamic changes. Traditional fixed-scale descriptors lack multi-resolution analysis capabilities and cannot effectively capture and align these distance-dependent target features. Finally, regarding optimization strategies, existing methods lack strict constraints on the high intra-class variance of sonar data. They typically rely on generalized quantization losses, failing to explicitly optimize the geometry of the embedding space. This results in generated continuous embeddings often hovering near the decision boundary, remaining ambiguous. Such embeddings are highly susceptible to sign flipping due to strong noise, leading to unstable binary encoding and ultimately a significant decrease in retrieval accuracy. Summary of the Invention

[0004] In view of this, the primary objective of this invention is to provide a sonar image retrieval method based on a deep wavelet convolutional hash network. This method integrates discrete wavelet transform into frequency sensing in the deep learning process, effectively solving the problem of achieving high-precision and high-efficiency indexing in massive, high-noise acoustic databases.

[0005] To achieve the above objectives, the specific technical solution adopted by the present invention is as follows:

[0006] A sonar image retrieval method based on a deep wavelet convolutional hash network is proposed. The key is to input the sonar image to be retrieved into a trained deep wavelet convolutional hash network to obtain the retrieval result. The deep wavelet convolutional hash network includes an image preprocessing module, a shared weight network module, and a hybrid quantization optimization module.

[0007] The image preprocessing module is used to perform image equalization and image enhancement, and outputs a weakly enhanced teacher view and a strongly enhanced student view.

[0008] The shared weight network module includes a wavelet aggregation backbone network and a hierarchical adaptive fusion module. The wavelet aggregation backbone network is configured with multiple stage modules. Each stage module decomposes the input features into low-frequency structural components and multiple high-frequency structural components through discrete wavelet transform. The multiple high-frequency structural components are then convolved and modulated, and feature reconstruction is achieved through inverse wavelet transform. The hierarchical adaptive fusion module adaptively and dynamically fuses the outputs of the last two stage modules of the wavelet aggregation backbone network, and finally continuously embeds the output through a hash layer.

[0009] The hybrid quantization optimization module is used to configure the total loss function based on the first continuous embedding obtained by the weakly enhanced teacher view through the shared weight network module and the second continuous embedding obtained by the strongly enhanced student view through the shared weight network module, and to train the network parameters of the deep wavelet convolutional hash network through training samples to make the total loss function optimal.

[0010] When the sonar image to be retrieved is input into a trained deep wavelet convolutional hash network, the continuous embeddings output by the shared weight network module are symbolically transformed to obtain the retrieval result.

[0011] Optionally, the image preprocessing module performs image equalization using a contrast-limited adaptive histogram equalization method before performing any geometric enhancement.

[0012] Optionally, the wavelet aggregation backbone network adopts the WTConvNeXt backbone network, which is configured with four stages, each of which has a WTConv2d layer.

[0013] Optionally, the adaptive fusion module is used to aggregate the third-stage output features and the fourth-stage output features of the wavelet aggregation backbone network, wherein:

[0014] The third-stage output features and the fourth-stage output features are respectively processed... The convolutional module projects to a unified channel dimension, and then adjusts the feature resolution of the third-stage projection to the resolution of the fourth-stage projection through bilinear interpolation. Then, the features of the two stages are weighted and fused through learnable adaptive weights. Finally, the feature map is refined by the fusion convolutional module and then adaptively recalibrated by the convolutional block attention module.

[0015] Optionally, the fused convolutional module includes two stacked sets of... Convolutional layers, batch normalization layers, and ReLU activation layers.

[0016] Optionally, the total loss function configured in the hybrid quantization optimization module is a weighted fusion of surrogate loss, distillation loss, center loss, and hybrid quantization loss, wherein:

[0017] The agent loss is obtained based on the first continuous embedding statistics obtained by passing the weakly enhanced teacher view through the shared weight network module, and is used to encourage hash codes to cluster tightly around their respective class agents;

[0018] The distillation loss is statistically derived from the first continuous embedding obtained by the weakly enhanced teacher view through the shared weight network module and the second continuous embedding obtained by the strongly enhanced student view through the shared weight network module, and is used to enhance the robustness of the model to different transformations.

[0019] The central loss is obtained based on the first continuous embedding statistics obtained by passing the weakly enhanced teacher view through the shared weight network module, and is used to enhance intra-class compactness and inter-class separability.

[0020] The hybrid quantization loss is derived from the first continuous embedding statistics obtained by passing the weakly enhanced teacher view through the shared weight network module, and combines a rigid amplitude-based quantization loss and a Gaussian-based binary cross-entropy loss.

[0021] Based on the above method, the present invention also provides a sonar image retrieval system based on a deep wavelet convolutional hash network, which is equipped with a sonar image acquisition device and an image retrieval device. The image retrieval device performs sonar image retrieval according to the sonar image retrieval method based on a deep wavelet convolutional hash network described above.

[0022] The beneficial effects of this invention are:

[0023] This invention proposes a deep wavelet convolutional hashing network, the first deep hashing framework specifically tailored for sonar image retrieval, effectively solving the key bottleneck of achieving high-precision and high-efficiency indexing in massive, high-noise acoustic databases. It also constructs a novel frequency-domain sensing architecture, the core of which lies in integrating discrete wavelet transform to explicitly decouple low-frequency structural semantics from high-frequency speckle noise, and designing a hierarchical adaptive fusion module to effectively overcome the inherent scale ambiguity of forward-looking sonar through dynamic aggregation of multi-scale representations. Furthermore, it collaboratively introduces degradation-sensing preprocessing and hybrid quantization optimization strategies to further enhance the robustness and compactness of binary encoding.

[0024] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0026] Figure 1 This is an architecture diagram of a sonar image retrieval system based on a deep wavelet convolutional hash network provided by the present invention;

[0027] Figure 2 This is a 64-bit t-SNE visualization diagram on the UATD test set in a specific embodiment of the present invention;

[0028] Figure 3 The mixing ratio in a specific embodiment of the present invention Sensitivity analysis curve;

[0029] Figure 4 The loss weights in a specific embodiment of the present invention Sensitivity analysis curve;

[0030] Figure 5 The loss weights in a specific embodiment of the present invention Sensitivity analysis curve;

[0031] Figure 6 The loss weights in a specific embodiment of the present invention Sensitivity analysis curves. Detailed Implementation

[0032] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0033] This embodiment defines the following issues for sonar image processing:

[0034] set up Indicates inclusion The training set of sonar images, where each image Corresponding to a label vector , covering There are several target categories. The goal of this framework is to learn a mapping function. Project the input to Widmanian space, generating binary encoding The optimization objective is to minimize the Hamming distance between semantically similar pairs while maximizing the Hamming distance between dissimilar pairs.

[0035] To address the aforementioned problems, this invention proposes a sonar image retrieval method based on a deep wavelet convolutional hash network, such as... Figure 1 As shown, the sonar image to be retrieved is input into a trained deep wavelet convolutional hash network to obtain the retrieval result. The deep wavelet convolutional hash network includes an image preprocessing module, a shared weight network module, and a hybrid quantization optimization module.

[0036] The image preprocessing module is used to perform image equalization and image enhancement, and outputs a weakly enhanced teacher view and a strongly enhanced student view.

[0037] The shared weight network module includes a wavelet aggregation backbone network and a hierarchical adaptive fusion module. The wavelet aggregation backbone network is configured with multiple stage modules. Each stage module decomposes the input features into low-frequency structural components and multiple high-frequency structural components through discrete wavelet transform. The multiple high-frequency structural components are then convolved and modulated, and feature reconstruction is achieved through inverse wavelet transform. The hierarchical adaptive fusion module adaptively and dynamically fuses the outputs of the last two stage modules of the wavelet aggregation backbone network, and finally continuously embeds the output through a hash layer.

[0038] The hybrid quantization optimization module is used to configure the total loss function based on the first continuous embedding obtained by the weakly enhanced teacher view through the shared weight network module and the second continuous embedding obtained by the strongly enhanced student view through the shared weight network module, and to train the network parameters of the deep wavelet convolutional hash network through training samples to make the total loss function optimal.

[0039] When the sonar image to be retrieved is input into a trained deep wavelet convolutional hash network, the continuous embeddings output by the shared weight network module are symbolically transformed to obtain the retrieval result.

[0040] pass Figure 1 As can be seen, the method proposed in this invention (defined as the DWH framework) is built upon a dual-view self-distillation paradigm, comprising a teacher branch that processes the weakly enhanced view and a student branch that processes the strongly enhanced view. This asymmetric design forces the network to learn feature representations robust to severe acoustic distortion. Specifically, the input image first undergoes degradation-aware preprocessing (DAP), followed by the generation of two views: a weak view serving as the pseudo-teacher. and as a student's strong view Both views are fed into the same shared weight backbone network for feature extraction.

[0041] To specifically address the unique challenges of underwater acoustic imaging, namely speckle noise interference and extreme scale blur, we carefully designed three key modules that work together:

[0042] (1) Degradation-Aware Preprocessing (DAP)

[0043] Standard optical enhancements often compromise the acoustic properties of sonar images. To mitigate inherent speckle noise without obscuring fine-grained details, we propose an "enhancement-then-transform" approach. Considering the low-contrast nature of acoustic echoes, we strictly prioritize using Limiting Contrast Adaptive Histogram Equalization (CLAHE) before any geometric enhancements to stretch the dynamic range of target intensity. Given the original image... We excluded the original Gaussian blur operation from the data augmentation pool to preserve key high-frequency acoustic cues (such as object shadows and boundaries), ultimately deriving the augmented view input.

[0044] (2) Wavelet aggregation backbone network (WIB)

[0045] To effectively decouple structural information from noise, this framework employs the WTConvNeXt backbone network, introducing explicit frequency-aware processing into sonar image retrieval tasks for the first time. As shown in Figure 1, the core WTConv2d module utilizes a dual-path parallel processing mechanism. For input features... The wavelet path is first decomposed into low-frequency structural components using Discrete Wavelet Transform (DWT). and a set consisting of three high-frequency subbands (Including details and speckle noise). To enhance effective texture and suppress noise, the module performs convolutional modulation on the high-frequency subbands, then reconstructs the features via inverse wavelet transform (IDWT), and fuses them with the output of the spatial fundamental path:

[0046]

[0047] in, This represents a spatial path that includes standard convolutions and learnable basic scaling. This represents a frequency-domain modulation operation involving depthwise convolution and learnable wavelet scaling. This design fully leverages the separability of the frequency domain to achieve precise suppression of sonar speckle noise while preserving key structural semantics.

[0048] (3) Hierarchical Adaptive Fusion (HAF)

[0049] Forward-looking sonar images possess a unique fan-shaped beam geometry, resulting in severe non-uniform resolution and scale blurring—targets appear very large in the near field and significantly smaller in the far field. Single-scale features struggle to simultaneously capture both large near-field structures and small far-field targets. To address this issue, the HAF module constructs a multi-scale representation by dynamically fusing features from different levels. Specifically, the module aggregates Stage 3 texture features. Semantic features of Stage 4 To eliminate the differences in channel dimensions, we first utilize... Convolution projects the features from the two layers onto a unified channel space, resulting in... and To address the spatial resolution mismatch, we employ bilinear interpolation to bridge the gap between high-resolution and low-resolution data. Downsampled to the same size as Stage 4, resulting in Subsequently, learnable adaptive weights are utilized. The two are then weighted and merged: ;in, and , representing the contribution weights of texture features and semantic features respectively. To ensure balance in the early stages of training, we set both initial values ​​to 0.5 and automatically learn the optimal ratio during training. To further integrate features and enhance non-linear expressive power, the weighted features... Next, a fusion convolution module (Fusion Conv) is used, which consists of two stacked sets of... The network consists of convolutional layers, batch normalization (BN) layers, and ReLU activation layers. Finally, the output features are refined via a convolutional block attention module (CBAM). This mechanism allows the network to adaptively recalibrate the feature responses based on the imaging distance, ensuring that the generated hash codes remain robust across different scanning ranges.

[0050] For the Hybrid Quantization Optimization (HQO) module, the backbone features are projected onto a K-dimensional Hamming space through a hash layer. Following the dual-view paradigm, robust continuous embeddings are generated for both the student view and the teacher view. To optimize these embeddings, we improve upon the Deep Hash Distillation (DHD) framework. While DHD provides a strong benchmark, its probabilistic quantization is unstable in the face of severe sonar speckle noise. Therefore, we propose a compact objective function that replaces redundant constraints with a robust hybrid quantization strategy derived from sensitivity analysis.

[0051] Total loss is defined as: ;

[0052] in It is the loss weight hyperparameter.

[0053] (1) Agency loss To enforce class separability, we employ a surrogate loss, which maintains a learnable representative vector for each class. It encourages hash codes to cluster tightly around their respective class agents:

[0054]

[0055] in Teacher view The first continuous embedding obtained after passing through the shared weight network module Corresponding teacher Category Indicates the first Categories The proxy vector of ) It is a temperature scaling factor. It's a temperature parameter. It is a scaling factor. This represents the cosine similarity.

[0056] (2) Distillation loss To transfer noise robustness from the teacher view (weak enhancement) to the student view (strong enhancement), we minimize the cosine distance between their consecutive embeddings: ;

[0057] Represents student view By minimizing the loss obtained from the second continuous embedding obtained through the shared weight network module, the student view learns to generate hash codes that are more robust to the teacher view, thereby enhancing the model's robustness to various transformations.

[0058] (3) Central loss ( As a complement to the surrogate loss, we introduce a center loss to explicitly optimize the geometry of the hash code space. Unlike the surrogate loss, which focuses on classification, the center loss improves the geometry by bringing samples closer to their class centers. To enhance intra-class compactness and inter-class separation: ;

[0059] in It is a category The learnable center vector, This is the temperature scaling factor. The formula encourages hash codes to be close to their corresponding class centers while being far from other class centers, thus creating a compact intra-class distribution and naturally promoting lower quantization errors.

[0060] (4) Mixed quantization loss Standard soft quantization methods often lack sufficient constraints for noisy sonar data. We propose a hybrid strategy. It adaptively combines a rigid amplitude-based quantization loss with a Gaussian-based binary cross-entropy loss. The former strictly enforces binarization, while the latter ensures optimization stability by modeling the probability likelihood of the hash bits. ;

[0061] in (represented in implementation as) The ) represents the mixing ratio. The CSQ quantization term encourages the absolute value of the hash code after tanh activation to be close to 1. ;

[0062] in Indicates sample The Bit hash code. This formula effectively reduces the deviation from the target magnitude by penalizing the bit hash code. Push towards The bce quantization term of DHD uses a probabilistic method based on a Gaussian distribution:

[0063]

[0064] in It is a probability based on a Gaussian distribution. ), representing the probabilities of being close to +1 and -1, respectively. Represents a binary target. It uses binary cross-entropy loss. This combination yields better retrieval accuracy by balancing strict binarization with gradient smoothness. The CSQ component provides strong constraints to ensure the final quantization quality, while the bce component in DHD provides better optimization stability through probabilistic modeling, especially in the early stages of training.

[0065] To further verify the effectiveness of the present invention, DWH was evaluated on two large-scale acoustic benchmark datasets, which present unique challenges in terms of image quality and target diversity.

[0066] The Underwater Acoustic Target Detection (UATD) dataset is a comprehensive multibeam forward-looking sonar (MFLS) benchmark containing 9200 images acquired by a Tritech Gemini 1200ik sonar. It covers 10 typical underwater categories: spheres, circular cages, cubes, cylinders, human bodies, metal barrels, aircraft, ROVs, square cages, and tires. For robust evaluation, we merged two test sets (Test1 and Test2, each containing 800 images) to construct a unified test set containing 1600 query images. The remaining 7600 images were used as the retrieval database and training set. This dataset serves as a standard benchmark for evaluating retrieval performance under different noise levels and target attitudes.

[0067] The PengCheng dataset (PCL-2021) originates from the 2021 Underwater Acoustic Target Detection Challenge. Collected using a Tritech Gemini 1200i sonar, this dataset contains 4000 images across 8 categories (cube, sphere, cylinder, human body, tire, circular net cage, square net cage, metal barrel). It is characterized by severe speckle noise and complex background interference. We split the dataset into a training / database set containing approximately 3600 images and a test set containing approximately 400 images. Crucially, to mitigate the inherent long-tail distribution problem of multi-label data, we designed a balanced strategy of "single-label priority, multi-label supplementation" when constructing the test set. Specifically, we prioritize single-label samples for each category, supplementing with multi-label samples only when necessary. This strategy ensures that each of the 8 categories is represented by at least 50 independent images containing that specific target, thus guaranteeing statistical significance for the evaluation of each category.

[0068] Evaluation metrics: Our report is [length missing] The mean precision (mAP) of the hash codes is used to comprehensively measure retrieval accuracy.

[0069] All methods are implemented based on the WTConvNeXt-Tiny backbone network and initialized using ImageNet-1K pre-trained weights. Experiments were conducted on a single NVIDIA GeForce RTX 2080 Ti GPU. The training consisted of 150 epochs with a batch size of 64. We used the Adam optimizer with weight decay set to [value missing]. The initial learning rate of the hash layer is set to... The learning rate of the backbone network is scaled by a scaling factor of 0.05 (i.e., ...). We employed a cosine annealing learning rate scheduling strategy with a 10-round linear warm-up phase. To ensure reproducibility, the random seed was fixed at 42, and gradient accumulation with a step size of 2 was applied. Starting from the 4th round, the model was evaluated every 5 rounds, and the checkpoint that produced the highest mAP was selected as the final model.

[0070] We compared DWH with eight representative methods, including general hashing methods (HashNet

[14] , DCH

[15] , GreedyHash

[16] , CSQ

[18] , DHD

[19] ) and retrieval-oriented methods (SPRCH). The mAP results are summarized in Table 1. On UATD, DWH consistently outperformed the strongest benchmark (DHD) across all code lengths, achieving a significant 3.6% improvement at 64 bits. This demonstrates the advantage of our wavelet-based feature modeling in handling acoustic shadowing and low-contrast targets. On PengCheng, DWH achieved an average improvement of 2.1%. Notably, general methods like SPRCH struggled to cope with the severe speckle noise in this dataset, while DWH maintained high accuracy, validating the effectiveness of our hierarchical fusion and robust optimization.

[0071] The specific literature is as follows:

[0072]

[14] H. Liu, R. Wang, S. Shan, and X. Chen, “Deep supervised hashing for fast image retrieval,” in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition, 2016, pp. 2064–2072.

[0073]

[15] X. Wang, Y. Shi, and KM Kitani, “Deep supervised hashing with triplet labels,” in Asian Conference on Computer Vision, 2016, pp. 70–84.

[0074]

[16] Z. Cao, M. Long, J. Wang, and P. S. Yu, “HashNet: Deep learningto hash by continuation,” in Proceedings of the IEEE International Conferenceon Computer Vision, 2017, pp. 5608–5617.

[0075]

[17] Y. Cao, M. Long, B. Liu, and J. Wang, “Deep cauchy hashing forhamming space retrieval,” in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition, 2018, pp. 1229–1237.

[0076]

[18] S. Su, C. Zhang, K. Han, and Y. Tian, “Greedy hash: Towards fastoptimization for accurate hash coding in CNN,” in Advances in NeuralInformation Processing Systems, vol. 31, 2018.

[0077]

[19] L. Yuan et al., “Central similarity quantization for efficientimage and video retrieval,” in Proceedings of the IEEE / CVF Conference onComputer Vision and Pattern Recognition, 2020, pp. 3083–3092.

[0078]

[20] YK Jang, G. Gu, B. Ko, I. Kang, and NI Cho, “Deep hashdistillation for image retrieval,” in Computer Vision--ECCV 2022, 2022, pp.354–371.

[0079]

[21] Z. Lu, L. Jin, Z. Li, and J. Tang, “Self-Paced RelationalContrastive Hashing for Large-Scale Image Retrieval,” IEEE Transactions onMultimedia, vol. 26, pp. 3392–3404, 2024.

[0080] Table 1 Comparison results on the UATD and PengCheng datasets.

[0081]

[0082] To further demonstrate the interpretability and robustness of DWH, we visualized the feature distribution and analyzed the hyperparameter sensitivity.

[0083] t-SNE Visualization: We used t-SNE to visualize the benchmark method (DHD) and the 64-bit hash code generated by our DWH on the UATD test set. For example... Figure 2 As shown, DWH produces more compact intra-class clusters and clearer inter-class boundaries, confirming the effectiveness of our center-based constraints and quantization constraints.

[0084] Hyperparameter sensitivity ( Figures 3-6 As shown): To determine the optimal contribution of each optimization component, we weight the loss... , and and the mixing ratio in the mixed quantification strategy A comprehensive sensitivity analysis was conducted. Experimental results show that the model achieves peak performance when these terms are carefully balanced. Sensitivity analysis results indicate that all loss weights have a significant impact on model performance. The best performance was achieved in the range of 0.3-0.6, proving the effectiveness of the hybrid quantization strategy; , and The optimal weight settings ensure that the model converges stably to a discriminative binary code. Under most bit conditions, performance degrades when the weights are (too small) close to 0 or too large, verifying the key role of these loss terms in model optimization.

[0085] In summary, the method and system proposed in this invention provide a robust framework specifically tailored for large-scale sonar image retrieval. Through WIB and HAF, this invention effectively decouples high-frequency speckle noise from structural semantics and addresses the inherent scale variation problem in acoustic images. Furthermore, we introduce DAP to enhance target visibility and design the HQO strategy. By combining the depth hash distillation framework with center-based constraints and hybrid quantization constraints, HQO ensures the generation of compact and discriminative binary codes. Extensive experiments on the UATD and PengCheng datasets demonstrate that DWH significantly outperforms existing state-of-the-art benchmark methods.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A sonar image retrieval method based on a deep wavelet convolutional hash network, characterized in that, The sonar image to be retrieved is input into a trained deep wavelet convolutional hash network to obtain the retrieval result. The deep wavelet convolutional hash network includes an image preprocessing module, a shared weight network module, and a hybrid quantization optimization module. The image preprocessing module is used to perform image equalization and image enhancement, and outputs a weakly enhanced teacher view and a strongly enhanced student view. The shared weight network module includes a wavelet aggregation backbone network and a hierarchical adaptive fusion module. The wavelet aggregation backbone network is configured with multiple stage modules. Each stage module decomposes the input features into low-frequency structural components and multiple high-frequency structural components through discrete wavelet transform. The multiple high-frequency structural components are then convolved and modulated, and feature reconstruction is achieved through inverse wavelet transform. The hierarchical adaptive fusion module adaptively and dynamically fuses the outputs of the last two stage modules of the wavelet aggregation backbone network, and finally continuously embeds the output through a hash layer. The hybrid quantization optimization module is used to configure the total loss function based on the first continuous embedding obtained by the weakly enhanced teacher view through the shared weight network module and the second continuous embedding obtained by the strongly enhanced student view through the shared weight network module, and to train the network parameters of the deep wavelet convolutional hash network through training samples to make the total loss function optimal. When the sonar image to be retrieved is input into a trained deep wavelet convolutional hash network, the continuous embeddings output by the shared weight network module are symbolically transformed to obtain the retrieval result.

2. The sonar image retrieval method based on a deep wavelet convolutional hash network according to claim 1, characterized in that, The image preprocessing module performs image equalization using a contrast-limited adaptive histogram equalization method before performing any geometric enhancement.

3. The sonar image retrieval method based on a deep wavelet convolutional hash network according to claim 1 or 2, characterized in that, The wavelet aggregation backbone network adopts the WTConvNeXt backbone network, which is configured with four stages, each of which has a WTConv2d layer.

4. The sonar image retrieval method based on a deep wavelet convolutional hash network according to claim 3, characterized in that, The adaptive fusion module is used to aggregate the third-stage output features and the fourth-stage output features of the wavelet aggregation backbone network, wherein: The third-stage output features and the fourth-stage output features are respectively processed... The convolutional module projects to a unified channel dimension, and then adjusts the feature resolution of the third-stage projection to the resolution of the fourth-stage projection through bilinear interpolation. Then, the features of the two stages are weighted and fused through learnable adaptive weights. Finally, the feature map is refined by the fusion convolutional module and then adaptively recalibrated by the convolutional block attention module.

5. The sonar image retrieval method based on a deep wavelet convolutional hash network according to claim 4, characterized in that, The fused convolutional module includes two stacked sets. Convolutional layers, batch normalization layers, and ReLU activation layers.

6. The sonar image retrieval method based on a deep wavelet convolutional hash network according to claim 4 or 5, characterized in that, The total loss function configured in the hybrid quantization optimization module is a weighted fusion of surrogate loss, distillation loss, center loss, and hybrid quantization loss, wherein: The agent loss is obtained based on the first continuous embedding statistics obtained by passing the weakly enhanced teacher view through the shared weight network module, and is used to encourage hash codes to cluster tightly around their respective class agents; The distillation loss is statistically derived from the first continuous embedding obtained by the weakly enhanced teacher view through the shared weight network module and the second continuous embedding obtained by the strongly enhanced student view through the shared weight network module, and is used to enhance the robustness of the model to different transformations. The central loss is obtained based on the first continuous embedding statistics obtained by passing the weakly enhanced teacher view through the shared weight network module, and is used to enhance intra-class compactness and inter-class separability. The hybrid quantization loss is derived from the first continuous embedding statistics obtained by passing the weakly enhanced teacher view through the shared weight network module, and combines a rigid amplitude-based quantization loss and a Gaussian-based binary cross-entropy loss.

7. A sonar image retrieval system based on a deep wavelet convolutional hash network, characterized in that, The device is equipped with a sonar image acquisition unit and an image retrieval device, wherein the image retrieval device performs sonar image retrieval according to the sonar image retrieval method based on a deep wavelet convolutional hash network as described in any one of claims 1-6.

Images