Unsupervised underwater target searching method based on pseudo label robust optimization
By generating and optimizing pseudo-labels in unsupervised underwater target search, and using a robust optimization module to improve the stability and reliability of pseudo-labels, the problem of unstable model training in cross-domain underwater target search is solved, achieving higher cross-domain accuracy and adaptability.
Patent Information
- Application Number
- CN202511877622.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-12
- Publication Date
- 2026-03-06
AI Technical Summary
In unsupervised cross-domain underwater target search, the target domain data lacks manual annotation, has a lot of pseudo-label noise, the model training is unstable, and the distribution difference between the source domain and the target domain is large, which leads to the performance degradation of the model when applied across domains. Traditional feature learning methods have poor robustness.
A pseudo-label-based robust optimization method is adopted. Pseudo-labels are generated in the target domain and robust optimization is performed. The instance consistency adaptive robust optimization module is used to perform batch-level feature correction and cross-batch consistency maintenance on pseudo-labels. Combined with multi-scale robust representation and consistency maintenance unit, the stability and credibility of pseudo-labels are improved.
It effectively reduces the impact of false label noise, improves the cross-domain accuracy and generalization ability of underwater target search, and enhances the model's adaptability and stability in complex underwater environments.
Smart Images
Figure CN121616946A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of underwater target search technology, and specifically to an unsupervised underwater target search method based on pseudo-label robust optimization. Background Technology
[0002] Underwater target search is a marine sensing task that combines underwater target detection and underwater target re-identification. Its goal is to automatically detect and retrieve specified underwater targets (such as specific underwater robots, shipwrecks, oil and gas pipelines, and unusual organisms) in complex underwater scenarios (such as nearshore waters, deep-sea mining areas, and port channels). Due to the high cost of acquiring large-scale cross-scenario underwater annotation data in practical applications (complex underwater environments, high operating costs of detection equipment, and the need for specialized domain knowledge for manual annotation), unsupervised cross-domain underwater target search methods have gradually become a research hotspot in recent years.
[0003] However, unsupervised cross-domain underwater target search generally suffers from the following problems: First, target domain data (such as deep-sea mining areas and turbid nearshore areas) lacks manual annotation and is usually trained using pseudo-labels generated by clustering based on underwater features (such as sonar echo features and optical image texture features). However, pseudo-labels often contain a lot of noise (such as ocean background interference and mis-clustering caused by sensor noise), leading to unstable model training. Second, there are significant distribution differences between the source domain (such as clear shallow seas and laboratory simulated environments) and the target domain (such as different water transparency, water flow velocity, and background noise types). When the model is directly transferred from the source domain to the target domain after training, its performance drops significantly. Third, traditional underwater feature learning methods are difficult to maintain robustness in cross-domain scenarios. Features are easily affected by factors such as changes in underwater illumination, occlusion by suspended particles, and multipath effects of sonar signals, which further exacerbates the inaccuracy of pseudo-labels.
[0004] Therefore, there is an urgent need for an unsupervised underwater target search method that can robustly model underwater target features in cross-domain scenarios, while improving the stability and reliability of pseudo-labels, so as to effectively reduce the impact of noisy pseudo-labels on model training and improve the accuracy and generalization ability of cross-domain underwater target search. Summary of the Invention
[0005] In view of the shortcomings of existing technologies, this invention provides an unsupervised underwater target search method based on pseudo-label robust optimization. The method first pre-trains a detection and re-identification network using source domain labeled data to obtain feature representations with basic discriminative capabilities. Then, it extracts features from unlabeled target domain data, generates pseudo-labels using a clustering algorithm, and performs robust optimization on the pseudo-labels to improve their stability and reliability. During training, the proposed cross-batch robust optimization mechanism effectively suppresses interference from noisy pseudo-labels by continuously and adaptively enhancing and filtering features in each batch. Finally, the optimized features are stored in an online memory module, enabling cross-domain joint training.
[0006] The technical means employed in this invention are as follows: An unsupervised underwater target search method based on pseudo-label robust optimization includes the following steps: S1. Obtain the source domain training dataset, which includes underwater images with manual annotations; S2. Construct an underwater target search model, which includes a feature extraction backbone network, a region candidate network, and a joint detection and re-identification network; perform supervised pre-training on the underwater target search model based on the source domain training dataset, and use the pre-trained underwater target search model as the initial feature extraction network. S3. Obtain the target domain training dataset, which includes real-world underwater images without manual annotation; S4. Use the initial feature extraction network to cluster the target domain training dataset and generate pseudo-labels; S5. The underwater target search model that has been pre-trained is retrained based on the target domain training data with added pseudo-labels. During the retraining process, an instance consistency adaptive robust optimization module is introduced to improve the stability of the pseudo-labels through batch-level feature correction and cross-batch consistency maintenance. The instance consistency adaptive robust optimization module includes a first feature difference perception unit, a second feature difference perception unit, a multi-scale robust representation unit, a consistency maintenance unit, and a robustness enhancement fusion unit. S6. Update the features optimized by the instance consistency adaptive robust optimization module to the online memory, and iteratively train the underwater target search model based on the updated features to optimize the pseudo-label generation process and output the target search results.
[0007] Furthermore, the first feature difference sensing unit is used to correct the differences between target domain samples at different viewpoints and scales, and its output is defined as:
[0008] in, The corrected features are output by the first feature difference sensing unit. These are the original features obtained after processing the target domain samples through a feature extraction backbone network. This is the difference correction function for the first feature difference sensing unit. The adaptive weights for the samples are updated based on the network's loss feedback during training, and ;
[0009] , and These represent the differential attention results calculated along the channel-width direction, the channel-height direction, and the height-width direction, respectively. Among them, differential attention involves recombining input features according to different height and width channel arrangements, so that the network can observe the same input feature from three directions: channel-width, height-channel, and height-width. In each permutation perspective, the corresponding difference attention result is generated based on the local response of the feature in the current direction, and the results of the three directions are averaged element by element to obtain the output feature that integrates the difference information of multiple directions.
[0010] Furthermore, the second feature difference sensing unit is used to correct the differences between target domain samples at different viewpoints and scales, and its output is defined as:
[0011] in, The corrected features are output by the second feature difference perception unit. These are the original features obtained after processing the target domain samples through a feature extraction backbone network. This is the difference correction function for the second feature difference sensing unit. The adaptive weights for the samples are updated based on the network's loss feedback during training, and ;
[0012] , These represent the differential attention results calculated along the channel-width direction and the channel-height direction, respectively. Among them, differential attention involves recombining input features according to different height and width channel arrangements, so that the network can observe the same input feature from the channel-width direction and the height-channel direction. In each permutation perspective, the corresponding difference attention result is generated based on the local response of the feature in the current direction, and the results of the two directions are averaged element by element to obtain the output feature that integrates the difference information of the two directions.
[0013] Furthermore, the multi-scale robust representation unit is used to extract discriminative features of target domain samples at different scales, and its output is defined as:
[0014] in, To distinguish features, For the first k A feature encoder at each scale, w k In order to target the k The learning weights of each scale feature, and satisfying , K The preset number of scales. Features are processed by the feature difference sensing unit. i =1, 2.
[0015] Furthermore, the output characteristics of the consistency preservation unit are defined as follows:
[0016] in, These are the original features obtained after processing the target domain samples through a feature extraction backbone network. Using the sigmoid function, real numbers are compressed to the (0, 1) interval, resulting in normalized attention weights. is the statistical mean of features in the current batch, and s is the unbiased variance of the current channel, used to measure the volatility of the entire channel and as a normalization factor. This is a numerical stability constant to prevent the denominator from being zero. To ensure element-wise multiplication, the attention weights correspond one-to-one with the feature map size; The consistency preservation unit is used to constrain the consistency of the target domain sample feature distribution across batches according to the following loss function:
[0017] in, N This represents the number of samples in the current training batch. For samples in the current batch i Feature statistics mean This represents the statistical mean of features stored in the online memory. for L2 Norm calculation.
[0018] Furthermore, the robustness enhancement fusion unit is used to fuse the outputs of the feature difference perception unit, the multi-scale robust representation unit, and the consistency preservation unit, and its final output is defined as:
[0019] in, The output features of the multi-scale robust representation unit are... The output characteristics of the consistency preservation unit. , is the feature fusion coefficient.
[0020] Furthermore, the process of updating the pseudo-label generation based on the optimized features output by the instance consistency adaptive robust optimization module is achieved through the following formula:
[0021] in, For target domain samples i Updated pseudo tags, This is the feature similarity measurement function. For category c Feature center, The optimized features output by the instance consistency adaptive robust optimization module. This means selecting the category with the highest similarity as the sample. i Pseudo-tags.
[0022] Compared with the prior art, the present invention has the following advantages: This invention proposes an unsupervised underwater target search method based on robust optimization of pseudo-labels. This method generates pseudo-labels on unlabeled data in the target domain and combines a robust optimization mechanism to perform credibility screening, dynamic threshold adjustment, and cross-batch feature enhancement on the pseudo-labels. This effectively alleviates the problems of pseudo-label noise accumulation, blurred class boundaries, and cross-domain feature shift in traditional methods. Unlike existing techniques that rely on single-batch clustering or fixed threshold screening, this invention enables continuous dynamic optimization of pseudo-label quality, significantly improving the discriminativeness and stability of feature representations.
[0023] This invention is effective and robust, thereby promoting the improvement of the application efficiency and adaptability of underwater target search in multiple fields such as marine security, deep-sea resource exploration, and underwater engineering operation and maintenance. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a flowchart of an unsupervised underwater target search method based on pseudo-label robust optimization in an embodiment of the present invention.
[0026] Figure 2 This is a diagram illustrating the overall architecture of the unsupervised underwater target search model in an embodiment of the present invention. Detailed Implementation
[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0029] like Figure 1 As shown, this invention provides an unsupervised underwater target search method based on pseudo-label robust optimization, comprising the following steps: S1. Construct a source domain training dataset, which is a collection of underwater images with manual annotations.
[0030] S101. Preprocess the source domain data, including performing random horizontal flipping, normalization and scaling on the underwater images to reduce low-level differences such as shape angle, scale and underwater lighting.
[0031] S102. Based on this, each source domain sample has a manually labeled identity ID and a corresponding bounding box, forming an "image-box-ID" triplet, which serves as the input for subsequent supervised training.
[0032] S2. Based on the source domain training dataset, the underwater target search model is pre-trained under supervision to obtain the initial feature extraction network.
[0033] S201. Construct a feature extraction backbone network. Input the underwater image into the backbone part (resnet1-resnet4) of the convolutional neural network to extract multi-level semantic features, providing a basic representation for subsequent detection and re-identification.
[0034] S202. Candidate region generation and feature alignment are performed. A region candidate network generates candidate underwater target regions, and the RoI-Align operation maps the candidate bounding box features to a fixed-size feature map to ensure scale invariance and spatial alignment. The region candidate network is the module used for box finding: a 3×3 convolution slides across the backbone network feature map, predicting foreground / background scores for 9 anchor boxes and 4 offsets at each position, outputting approximately 2k candidate boxes for subsequent fine classification.
[0035] S203, Joint Detection and Re-identification Training: The features processed by RoI-Align are input into res5 and the fully connected layer, which output bounding box prediction, class prediction, and object re-identification features, respectively. Supervised training is then performed using a joint loss function to obtain the initial feature extraction network, i.e. Figure 2 The inference network in the dataset. The loss function is calculated as follows:
[0036]
[0037]
[0038]
[0039] in, Indicates regression loss, N It is the number of positive samples. It is a smooth L1 loss. and They are the first i The predicted bounding box and the corresponding ground truth bounding box for each sample; Represents classification loss. It is the first i The predicted classification probability of each sample. These are the corresponding actual label values; This indicates the re-identification loss. The instance representation extracted from the model. express The category center, Representing the i Category centers of each category This represents the temperature factor.
[0040] S3. Construct a target domain training dataset, which is a collection of real-world underwater images without human annotation, including the following steps.
[0041] S301. Perform preprocessing on the real scene image of the target domain, which is the same as the flipping and normalization of the source domain.
[0042] S302. The target domain samples do not contain identity IDs or bounding box labels, forming an unlabeled dataset for pseudo-label generation and subsequent training.
[0043] S4. Cluster the target domain training dataset using the initial feature extraction network to generate pseudo-labels, including the following steps.
[0044] S401. Input the unlabeled underwater image of the target domain into the initial feature extraction network obtained through S2 to obtain a set of sample feature vectors:
[0045] in, N The number of samples in the target domain. d The feature dimension is used. Then, the feature similarity or distance matrix between samples is calculated:
[0046] in, Indicates sample (i,j) The cosine distance between them is used for subsequent clustering.
[0047] S402, Based on the distance matrix D A density-based clustering algorithm is used to divide the target domain samples into multiple clusters, resulting in a set of pseudo-labels.
[0048] in, Clustering functions, such as DBSCAN Clustering, For the sample i The pseudo-label category index is used. Samples belonging to the same cluster in the clustering results are regarded as pseudo-labels with the same identity, thus transforming unsupervised target domain samples into "image-pseudo-label" pairs, laying the foundation for subsequent domain adaptive training.
[0049] S5. In the target domain training process, an instance consistency adaptive robust optimization module is introduced to improve the stability of the pseudo-labels through batch-level feature correction and cross-batch consistency maintenance. This mainly includes the following steps.
[0050] S501. In each batch of input target domain samples, the underwater target features are corrected using the first and second feature difference sensing units respectively.
[0051] Specifically, the output of the first feature difference sensing unit is defined as:
[0052] in, The corrected features are output by the first feature difference sensing unit. These are the original features obtained after processing the target domain samples through a feature extraction backbone network. This is the difference correction function for the first feature difference sensing unit. The adaptive weights for the samples are updated based on the network's loss feedback during training, and ;
[0053] , and These represent the differential attention results calculated along the channel-width direction, the channel-height direction, and the height-width direction, respectively. The differential attention involves recombining the input features according to different height-width channel arrangements, so that the network can observe the same input feature from three directions: channel-width, height-channel, and height-width. Under each arrangement viewpoint, the corresponding differential attention result is generated based on the local response of the feature in the current direction, and the results of the three directions are averaged element-wise to obtain the output feature that integrates differential information from multiple directions.
[0054] Furthermore, the output of the second feature difference sensing unit is defined as:
[0055] in, The corrected features are output by the second feature difference perception unit. These are the original features obtained after processing the target domain samples through a feature extraction backbone network. This is the difference correction function for the second feature difference sensing unit. The adaptive weights for the samples are updated based on the network's loss feedback during training, and ;
[0056] , These represent the differential attention results calculated along the channel-width direction and the channel-height direction, respectively. The differential attention involves recombining the input features according to different height-width channel arrangements, so that the network can observe the same input features from the channel-width direction and the height-channel direction. Under each arrangement viewpoint, the corresponding differential attention result is generated based on the local response of the feature in the current direction, and the results of the two directions are averaged element by element to obtain the output feature that integrates the differential information of the two directions.
[0057] S502. The corrected features are further input into a multi-scale robust representation unit to enhance their discriminative power across different spatial scales. This unit contains feature encoders at multiple scales. The output is:
[0058] in, To distinguish features, For the first k A feature encoder at each scale, w k In order to target the k The learning weights of each scale feature, and satisfying , K The preset number of scales. Features are processed by the feature difference sensing unit. i =1, 2.
[0059] S503. To address the potential instability of pseudo-labels across different batches, a consistency-preserving unit is introduced to suppress noise by constraining the statistical consistency of features across batches. Its constraint loss is defined as:
[0060] in, N This represents the number of samples in the current training batch. For samples in the current batch i Feature statistics mean This represents the statistical mean of features stored in the online memory. for L2 Norm calculation. This constraint ensures that features remain stable across batch updates, thereby improving the robustness of pseudo-labels. The output feature of the consistency-preserving unit is defined as:
[0061] in, These are the original features obtained after processing the target domain samples through a feature extraction backbone network. Using the sigmoid function, real numbers are compressed to the (0, 1) interval, resulting in normalized attention weights. is the statistical mean of features in the current batch, and s is the unbiased variance of the current channel, used to measure the volatility of the entire channel and as a normalization factor. This is a numerical stability constant to prevent the denominator from being zero. To ensure element-wise multiplication, attention weights are matched one-to-one with feature map sizes.
[0062] S504. Finally, the outputs of the above sub-modules are weighted and fused through the robust enhancement fusion unit to obtain the final optimized representation:
[0063] in, The output features of the multi-scale robust representation unit are... The output characteristics of the consistency preservation unit. , is the feature fusion coefficient.
[0064] S6. Update the features optimized by the instance consistency adaptive robust optimization module to the online memory, and iteratively train the underwater target search model based on the updated features.
[0065] S601, For batch-level optimized features According to the corresponding pseudo-tag category Write to the online memory and update the cluster centers of the corresponding categories using an exponential moving average method:
[0066] in, To update the coefficients, Indicates pseudo-label category The cluster center. This update strategy ensures that the pseudo-labels remain stable throughout the iteration process, effectively mitigating the training non-convergence problem caused by pseudo-label noise and feature drift.
[0067] S602. After the memory bank is updated, the model continues to backpropagate and adjust parameters using optimized features as supervision signals, so that the feature extraction network approaches a stable cross-domain representation round by round, thereby continuously improving the accuracy and robustness of the target domain pseudo-labels during iterative training.
[0068] 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An unsupervised underwater target search method based on pseudo-label robust optimization, characterized in that, The method comprises the following steps: S1, obtaining a source domain training data set, wherein the source domain training data set comprises underwater images with artificial labels; S2, constructing an underwater target search model, wherein the underwater target search model comprises a feature extraction backbone network, a region candidate network, and a joint detection and re-identification network; S3, obtaining a target domain training data set, wherein the target domain training data set comprises real scene underwater images without artificial labels; S4, clustering the target domain training data set using the initial feature extraction network to generate pseudo labels; S5, retraining the pre-trained underwater target search model based on the target domain training data with the pseudo labels, wherein an instance consistency adaptive robust optimization module is introduced in the retraining process, the stability of the pseudo labels is improved through batch-level feature correction and cross-batch consistency maintenance, and the instance consistency adaptive robust optimization module comprises a first feature difference perception unit, a second feature difference perception unit, a multi-scale robust representation unit, a consistency maintenance unit, and a robustness enhancement fusion unit; S6, updating the features optimized by the instance consistency adaptive robust optimization module to an online memory bank, iteratively training the underwater target search model based on the updated features, optimizing the generation process of the pseudo labels, and outputting a target search result. The first feature difference perception unit is used to correct the differences of target domain samples under different perspectives and scales, and the output thereof is defined as:
2. The unsupervised underwater target search method based on pseudo label robust optimization according to claim 1, characterized in that, wherein the difference attention comprises recombining the input features according to different high-width channel arrangement sequences, so that the network can observe the same input features from the channel-width direction, the height-channel direction, and the height-width direction; wherein, the corrected feature output by the first feature difference perception unit, the original feature obtained by processing the target domain sample through the feature extraction backbone network, the difference correction function of the first feature difference perception unit, the adaptive weight of the sample, which is updated according to the loss feedback of the network during the training process, and ; 、 and denote the difference attention results computed along the channel-width, channel-height, and height-width directions, respectively; In each arrangement perspective, the corresponding difference attention result is generated according to the local response of the features in the current direction, and the results of the three directions are averaged element by element, so as to obtain the output features that integrate multiple direction difference information. The second feature difference perception unit is used to correct the differences of target domain samples under different perspectives and scales, and the output thereof is defined as:
3. The unsupervised underwater target search method based on pseudo-label robust optimization according to claim 2, characterized in that, wherein the difference attention comprises recombining the input features according to different high-width channel arrangement sequences, so that the network can observe the same input features from the channel-width direction and the height-channel direction; wherein, the corrected feature output by the second feature difference perception unit, the original feature obtained by processing the target domain sample through the feature extraction backbone network, the difference correction function of the second feature difference perception unit, the adaptive weight of the sample, which is updated according to the loss feedback of the network during the training process, and ; , respectively denote the difference attention results calculated along the channel-width direction and the channel-height direction; In each arrangement perspective, the corresponding difference attention result is generated according to the local response of the features in the current direction, and the results of the two directions are averaged element by element, so as to obtain the output features that integrate two direction difference information. The multi-scale robust representation unit is used to extract discriminative features of target domain samples under different scales, and the output thereof is defined as:
4. The unsupervised underwater target search method based on pseudo-label robust optimization according to claim 3, characterized in that, The output feature of the consistency maintenance unit is defined as: wherein, is a discriminative feature, is a first k dimensional feature encoder, w k is a learning weight for the first k dimensional feature, and satisfies , K is a preset number of scales, is a feature processed by the feature difference perception unit, i =1, 2.
5. The unsupervised underwater target search method based on pseudo-label robust optimization according to claim 1, characterized in that, The consistency maintenance unit is used to constrain the consistency of the feature distribution of the target domain samples between batches according to the following loss function: wherein, is the original feature obtained by processing the target domain sample through the feature extraction backbone network, is a Sigmoid function that compresses real numbers into the interval (0, 1) to obtain normalized attention weights, is the feature statistical mean in the current batch, s is the unbiased variance of the current channel, which is used to measure the fluctuation size of the entire channel and is used as a normalization factor, is a numerical stability constant to prevent the denominator from being zero, is an element-wise multiplication to ensure that the attention weights and the feature map size correspond one by one; The robustness enhancement fusion unit is used to fuse the output results of the feature difference perception unit, the multi-scale robust representation unit, and the consistency maintenance unit, and the final output thereof is defined as: wherein, N is the number of samples in the current training batch, is the sample i feature statistics mean, is the statistics mean of the feature stored in the online memory bank, is the L2 norm calculation.
6. The unsupervised underwater target search method based on pseudo-label robust optimization according to claim 1, characterized in that, wherein, are output features of the multi-scale robust representation unit, are output features of the consistency preserving unit, , are feature fusion coefficients.
7. The unsupervised underwater target search method based on pseudo-label robust optimization according to claim 1, characterized in that, According to the generation process of the optimization feature update pseudo label output by the example consistency adaptive robust optimization module, the following formula is used: wherein, target domain samples i updated pseudo labels, is a feature similarity measure function, is a class c feature center, is an optimized feature output by the instance-consistent adaptive robust optimization module, represents taking the class with the maximum similarity as the pseudo label of the sample i .