Cross-domain underwater target search method based on pseudo tag refinement and confidence cluster memory
By refining the pseudo-labels based on neighborhood consistency and using confidence cluster memory, the problems of pseudo-label instability and memory drift in underwater target search are solved, improving the accuracy and stability of underwater target search and enabling effective detection and re-identification in complex underwater environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN MARITIME UNIVERSITY
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-05
AI Technical Summary
Existing technologies suffer from unstable pseudo-tag quality and unreliable memory writing in underwater target search, resulting in poor cross-domain migration performance. This is especially true in real underwater scenarios where the imaging medium is unstable and the background structure is complex, making pseudo-tags susceptible to noise and memory drifting, which affects the accuracy of detection and re-identification.
We employ a neighborhood consistency pseudo-label refinement and confidence cluster memory approach. We construct a stable candidate set through κ-nearest neighbors and use distance-temperature weighted soft voting to form the neighborhood posterior and correct pseudo-labels. At the same time, we utilize confidence-weighted cluster proxy memory to adaptively adjust sample weights based on the cosine distance between the sample and the prototype and the cluster size, thereby optimizing memory bank updates.
It improves the consistency and purity of pseudo-labels, reduces confirmation bias in pseudo-supervision loops, reduces the risk of memory drift, enhances the detection recall and retrieval discrimination of cross-domain underwater target search, and achieves robust transfer and generalization in unlabeled target domains.
Smart Images
Figure CN121982503A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of underwater target search technology, specifically to a cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory. Background Technology
[0002] In recent years, adaptive underwater target search in unsupervised domains has often adopted a "pseudo-supervised—self-training" closed loop: first, pre-training is performed on a unified detection / re-identification architecture using source domain labeled data; then, candidate boxes and features are generated based on unlabeled data from the target domain; pseudo-labels are obtained through clustering; and these pseudo-labels are used to feed back into the training to gradually approximate the target domain distribution. This approach alleviates the problem of scarce cross-domain annotations to some extent, but it still faces two key bottlenecks when implemented in real underwater scenarios.
[0003] One issue is the unstable quality of pseudo-labels. The target domain is affected by imaging medium, illumination, turbidity, background texture, and occlusion patterns. The feature topology between instances is often sparse and distorted, making the initial clusters obtained by reordering distance + DBSCAN sensitive to hyperparameters: boundary samples and sparse small clusters are easily labeled as noise or merged into incorrect large clusters; "fragmented clusters" with the same identity appearing from different perspectives are difficult to merge.
[0004] Secondly, memory writes are unreliable and prone to drift. Most methods use uniform momentum or equal-weighted writes, treating each sample's contribution to its cluster center as equally important. Large clusters and small clusters have different statistical stability. If equal-weighted writes are used without considering cluster size, the method is easily influenced by random noise, causing prototype bias and training instability. In addition, the existence of hard-to-bear samples amplifies the damage of erroneous writes to the prototype, inducing memory drift. Summary of the Invention
[0005] In view of the shortcomings of existing technologies, this invention proposes a cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory. Addressing the problems of unstable imaging media, complex background structures, and noise-prone pseudo-labels in real underwater scenarios, this method focuses on two main lines: cleaner pseudo-labels and more reliable memory writing, achieving robust transfer and generalization without the need for target domain annotation. Neighborhood consistency pseudo-label refinement: After pre-training a unified detection / re-identification architecture on source domain data, initial pseudo-labels are obtained in the target domain using re-ranked Jaccard distance and density clustering. Subsequently, a stable candidate set is constructed using κ-nearest neighbors, and a Top-K voting neighborhood is selected. Distance-temperature weighted soft voting is used to form the neighborhood posterior, with single-point correction only performed when the posterior is significantly better than the original label prior. This process effectively repairs broken clusters and eliminates mislabeling and noise caused by occasional nearest neighbors, thereby improving the consistency and purity of pseudo-labels. Confidence-weighted cluster proxy memory: To address the pain point of "equal-weighted writing is prone to drift" in the self-training stage, a prototype writing mechanism with adaptive scaling based on sample confidence is proposed. The confidence consists of two pieces of evidence: (1) Distance confidence: The cosine distance between the sample and the prototype proxy is standardized by z-score in the source / target domains respectively, and then mapped by Sigmoid and a lower limit is set to ensure that different domains are comparable on the same standard scale; (2) Size confidence: Based on the cluster size in the memory bank, normalization and logarithmic compression are performed on the source and target domain segments respectively using the 99th percentile to suppress the statistical bias of maximal clusters.
[0006] The technical means employed in this invention are as follows: A cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory includes the following steps: S1. Construct an underwater target search model architecture, wherein the underwater target search model includes a backbone network, a detection head, and a re-identification head; S2. Obtain the source domain underwater image dataset, and pre-train the underwater target search model based on the source domain underwater image dataset to obtain the underwater target search pre-trained model. The source domain underwater image dataset includes labeled underwater image data. S3. Obtain the target domain underwater image dataset and input it along with the source domain underwater image dataset into the underwater target search pre-training model. Extract the identity-level features from the source and target domains respectively, perform clustering on the identity-level features, and write the obtained cluster centers into the cluster proxy memory as the initial prototype. At the same time, count the number of identity-level features corresponding to each identity and register them as the size of the cluster in the gradient-free cluster size buffer. The identity-level features are the features output by the re-identification head in the underwater target search pre-training model. S4. Obtain the region candidate network of the detection head in the underwater target search pre-trained model, input the underwater image data of the target domain into the region candidate network to generate detection boxes, calculate the Jaccard distance of the generated detection boxes, and obtain the initial pseudo-labels according to DBSCAN. S5. Adopt the κ-nearest neighbor strategy, select K nearest neighbors, form the neighborhood posterior by soft voting with distance and temperature weighting, and update the pseudo-labels according to the comparison results of the pseudo-labels of the neighborhood posterior domains, thereby realizing the refinement of pseudo-labels. S6. Use the target domain identity-level features corresponding to the pseudo-labels to train the cluster proxy memory with confidence weighting. The distance confidence is obtained by mapping the cosine distance between the sample and the prototype through the in-domain z-score and sigmoid, and the size confidence is obtained by the cluster size through log compression. The sample weights are adaptively fused into sample weights based on the variance ratio of the distance confidence and the size confidence in the current batch, and the sample weights are optimized and updated. S7. Utilize the underwater target search model trained with cluster proxy memory to perform inference in the target domain and output cross-domain unsupervised underwater target search results.
[0007] Furthermore, the underwater target search model uses ResNet-50 as the backbone network, and the outputs of Stage-4 and Stage-5 are fed into the detection head and the re-identification head respectively after multi-scale alignment.
[0008] Furthermore, a stable candidate set is constructed using κ-nearest neighbors, and the Top-K voting neighborhoods are selected, including determining the subsequent candidate neighborhoods and voting neighborhoods in the following manner:
[0009]
[0010] in, For the κ candidate neighborhoods of sample i, For the i-th candidate box feature in the target domain, k nearest neighbors For the final voting neighborhood, To select the first k functions.
[0011] Furthermore, neighborhood posteriors are formed using distance-temperature weighted soft voting, including: The neighborhood posterior weights and the weighted votes for each label are obtained as follows:
[0012]
[0013] in, Let be the Euclidean distance between sample i and its neighbor sample j in the feature space. The distance temperature is used to adjust the intensity of distant neighbor inhibition. Let j be the soft weight of neighbor j with respect to sample i. For tags The number of votes in the neighborhood, This is the initial pseudo-label for DBSCAN.
[0014] Furthermore, the pseudo-labels are updated based on the comparison results of the pseudo-labels of the neighboring posterior domains, including: Get the maximum vote tag and its strength:
[0015]
[0016] in, For the largest tag, To maximize the strength of the pseudo-label, we check whether the posterior is still better than the prior of the original label, thereby achieving a more refined adjustment of the pseudo-label:
[0017]
[0018] in, This represents the prior weight of the original label. This represents the refined label. If the neighborhood of i is entirely noise or the posterior is insufficient to overturn the prior, then the original value is retained. For DBSCAN initial pseudo-labels, This indicates an indicator function, where the proposition is true if it is 1, and false if it is 0. This indicates a change in the label's proportion. Represents the number of samples in the target domain, only if the following condition is met. Furthermore, the pseudo-label of the sample is changed only when there are non-noise labels in the original pseudo-label and its neighborhood. Updated to Otherwise, it remains unchanged, thus making single-point corrections only when there is sufficient evidence in the neighborhood, thereby improving the consistency and purity of false labels.
[0019] Furthermore, the distance confidence score is obtained according to the following calculation:
[0020] in, As the lower limit of the weight, The slope of the sigmoid function. For the sigmoid function, The input sample instance features are the standardized distances between the target agents in the cluster agent memory; Take the cluster size corresponding to the label Take the 99th percentile on a set of the same size. Returning to the One Foundation:
[0021]
[0022] in, The sample size of the cluster corresponding to the label. For a set of domain-scale data, the scale confidence score is obtained based on the following calculation:
[0023] in, The function is a compression function, and the calculated scale confidence level is then scaled down to [min_w, 1], where This is the lower limit of the weight.
[0024] Furthermore, the sample weights are optimized and updated by adaptively fusing the variance proportions of distance confidence and size confidence in the current batch, including: Calculate the variance within the batch:
[0025]
[0026] The fusion coefficient is given by the variance proportion. And crop it to [0.7, 0.95]:
[0027]
[0028] in, , Let V be the variance of the two confidence levels in this batch. To prevent division by zero anomalies by a minimal constant, The fusion coefficient is... The final sample confidence level.
[0029] Furthermore, training clustered proxy memory using confidence-weighted features based on target domain identity levels also includes: Utilizing the obtained identity-level features For the feature centers of the memory bank class, based on the final sample confidence level and the preset momentum coefficient Calculate update step size And in an exponential sliding manner, the identity-level features are written to the class center features of the corresponding identity in its memory:
[0030] in, For momentum, the default value is 0.2. This represents the actual write strength; the more reliable the signal, the greater the strength. This represents the class center features of the corresponding label in the memory. Using the formula above, when updating the class center features of the corresponding label in the memory using identity-level features, distance confidence and size confidence are introduced to more reliably influence the class center features.
[0031] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces neighborhood consistency pseudo-label refinement. Based on the initial pseudo-labels obtained by reordering Jaccard + DBSCAN, a stable candidate set is first constructed using κ-nearest neighbors. Then, the Top-K voting neighborhoods are selected from this set, and a neighborhood posterior is formed using soft voting weighted by "distance-temperature". The label of a single sample is only corrected when the posterior is significantly better than the original label prior. This mechanism can effectively repair "fragmented clusters" caused by multiple perspectives, reduce the mislabeling rate of boundary samples affected by occasional nearest neighbor interference, continuously "purify" the target domain supervision signal, and reduce confirmation bias in pseudo-supervision closure loops.
[0032] 2. This invention proposes a confidence-weighted cluster proxy memory: distance confidence is obtained using the cosine distance standardized by the z-score of the domain, and size confidence is obtained using the cluster size normalized to the 99th percentile within the domain. These are then adaptively fused into sample weights based on the variance proportions of the two pieces of evidence. The write intensity scales linearly with the confidence level while maintaining L2 normalization, ensuring that "more writes are made to credible samples and fewer writes are made to suspicious samples." Simultaneously, the cluster_size is explicitly maintained and dynamically updated. Combined with unit-scale management and rapid statistical resetting of hard-to-bear samples, this suppresses the long-term contamination of the prototype by erroneous samples and significantly reduces the risk of rolling drift during cross-cycle training. It balances purity and discriminativity, improving retrieval discrimination and detection recall.
[0033] In summary, this application improves intra-class consistency and cross-perspective coherence through pseudo-label refinement, and enhances the contribution of reliable samples and suppresses noise gradients during the confidence memory update stage, thereby simultaneously improving the feature structure of "intra-class compactness versus inter-class separation". It can effectively alleviate mislabeling and drift problems caused by cross-domain distribution bias and background noise under unlabeled target domain conditions. Attached Figure Description
[0034] 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.
[0035] Figure 1This is a flowchart of a cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory, according to an embodiment of the present invention.
[0036] Figure 2 A framework diagram of a cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory in an embodiment of the present invention.
[0037] Figure 3 This is a diagram of the pseudo-label-based refinement module in an embodiment of the present invention.
[0038] Figure 4 This is a diagram of the confidence cluster memory module in an embodiment of the present invention. Detailed Implementation
[0039] 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. 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.
[0040] 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.
[0041] like Figure 1 , Figure 2 As shown, this invention provides a cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory, including the following steps: S1. Construct an underwater target search model architecture, including a backbone network, a detection head, and a re-identification head. This pre-trained network architecture is used to process data from both the source and target domains simultaneously. The learning tasks for the source and target domains are jointly optimized by a feature extractor with shared weights and the re-identification / detection head.
[0042] Specifically, a unified model is adopted, with ResNet-50 as the backbone network, Region Proposal Network (RPN) and Region Interest Head (RoIhead) as detection branches, and canonical perceptual embedding as the Re-identification (Reid) branch. The backbone network outputs two levels of features, feat_res4 and feat_res5, which are fed into the detection head and re-identification head respectively after multi-scale alignment. At the inference end, bounding boxes, scores, and embeddings are returned simultaneously, and non-maximum suppression and Top-K selection are performed.
[0043] The system connects image-level / instance-level discriminators with a consistency path and uses a gradient inversion layer (GRL) to achieve adversarial domain alignment. The image-level and instance-level GRLs have negative weights to invert gradients and confuse the domain, while the consistency GRL has positive weights to measure the consistency loss between the source and target outputs, thus achieving a balance between "discrimination" and "transferability".
[0044] S2. Obtain the source domain underwater image dataset, and pre-train the underwater target search model based on the source domain underwater image dataset to obtain the underwater target search pre-trained model. Here, the source domain underwater image dataset includes labeled underwater image data.
[0045] S3. Obtain the target domain underwater image dataset and input it along with the source domain underwater image dataset into the underwater target search pre-trained model. Extract identity-level features from both the source and target domains, perform clustering on the identity-level features, and write the obtained cluster centers into the cluster proxy memory as initial prototypes. Simultaneously, count the number of identity-level features corresponding to each identity and register them as the cluster size in a gradient-free cluster size buffer. The identity-level features are the features output by the re-identification head in the underwater target search pre-trained model. Specifically, this includes: S301. After supervised pre-training of the model on labeled source domain data, the source domain and target domain images are input into the shared feature extractor and re-identification head to obtain 256-dimensional underwater target re-identification features for each labeled instance, i.e., identity-level features.
[0046] S302. Aggregate instance-level features according to identity labels, calculate the feature mean of each identity as the identity-level class center, and write it into the cluster proxy memory as the initial prototype; at the same time, count the number of identity-level features corresponding to each identity, use it as the size of the cluster, and register it in the gradient-free cluster size buffer for subsequent confidence estimation and memory update based on cluster size.
[0047] S4. Obtain the region candidate network of the detection head in the underwater target search pre-trained model, input the underwater image data of the target domain into the region candidate network to generate detection boxes, calculate the Jaccard distance of the generated detection boxes, and obtain the initial pseudo-labels according to DBSCAN.
[0048] Specifically, the Jaccard distance is reordered for the target domain features and the initial pseudo-labels are obtained using DBSCAN.
[0049] S5. Subsequently, K nearest neighbors are selected from the κ-mutual nearest neighbors, and a neighborhood posterior is formed by soft voting weighted by distance and temperature. The pseudo-label is only corrected at a single point when the posterior is significantly better than the original label prior, thus improving consistency and purity. The pseudo-label refinement module is as follows: Figure 2 As shown. It mainly includes the following steps: S501. Using Euclidean distance as a metric (features have been normalized, equivalent to a monotonic transformation of cosine similarity), find the κ-nearest neighbor candidate set for each sample. Select K nearest neighbors from the κ-nearest neighbors, and form the posterior neighborhood using distance-temperature weighted soft voting. The posterior candidate neighborhood and the voting neighborhood are:
[0050] Sort the neighborhoods by Euclidean distance in ascending order, and select the top K neighborhoods to obtain the voting neighborhoods.
[0051] in, Let κ be the candidate neighborhood of sample i (excluding itself). Features of the i-th candidate box in the target domain. k nearest neighbors For the final voting neighborhood. To select the first k functions.
[0052] S502. Perform temperature-weighted soft voting to form the neighborhood posterior. For each label, assign a weighted vote and ignore noise-1 neighbors.
[0053]
[0054] in, Let be the Euclidean distance between sample i and its neighbor sample j in the feature space, where the features have been normalized. The distance temperature is used to adjust the intensity of distant neighbor inhibition. Let j be the soft weight of neighbor j for sample i. For tags The number of votes in the neighborhood. This is the initial pseudo-label for DBSCAN.
[0055] The "consistent support of neighbors" is quantified using softmax distance weights to avoid hard voting being affected by "a random neighbor"; temperature provides a "softness / hardness knob".
[0056] S503. Finally, obtain the label with the largest vote and its strength. If a label has strong consensus in its neighborhood, its strength will be close to 1; if opinions in the neighborhood are divided, the strength will decrease, indicating "uncertainty".
[0057]
[0058] in, For the largest tag, This determines the strength of the maximum label. It also checks whether the posterior is still superior to the prior of the original label, thus enabling refined adjustment of the pseudo-label. S504, Detailed Output and Statistics:
[0059]
[0060] in, This represents the prior weight of the original label. This represents the refined label. If the neighborhood of i is all noise or the posterior is insufficient to overturn the prior, then the original value is retained. This is the initial pseudo-label for DBSCAN. This indicates an indicator function, where a proposition is true if it is 1, and 0 otherwise. This indicates a change in the label's proportion. Represents the number of samples in the target domain. Only if the following condition is met. Furthermore, the pseudo-label of the sample is changed only when there are non-noise labels in the original pseudo-label and its neighborhood. Updated to Otherwise, it remains unchanged, thus making single-point corrections only when there is sufficient evidence in the neighborhood, thereby improving the consistency and purity of false labels.
[0061] S6. Use the target domain identity-level features corresponding to the pseudo-labels to train the cluster proxy memory with confidence weighting. Obtain the distance confidence by mapping the cosine distance between the sample and the prototype through the in-domain z-score and sigmoid, and obtain the size confidence by the cluster size through logarithmic compression. Adaptively fuse the distance confidence and size confidence into sample weights based on the variance ratio of the current batch, and optimize and update the sample weights.
[0062] Specifically, in the confidence-weighted cluster proxy memory training, the distance confidence is obtained by mapping the cosine distance between the sample (identity-level features) and the prototype (class center features of the memory bank) through in-domain z-score and sigmoid, and the size confidence is obtained by logarithmic compression of the cluster size. The two confidences are adaptively fused into sample weights based on their variance proportions in the current batch for optimization and updating. The confidence cluster memory module is as follows: Figure 3 As shown. It mainly includes the following steps: S601. Give "trustworthy samples" a large write weight and "abnormal samples / large cluster bias" a small write weight; and let the system decide for itself whether "distance confidence" or "scale confidence" is more important. First, perform L2 normalization on the identity-level features within the batch and their corresponding prototypes in the memory, and calculate the cosine distance d.
[0063]
[0064]
[0065]
[0066] in, For sample features, For the label in the memory bank The prototype. , These are the results after L2 normalization for each. The cosine distance between the sample and the prototype.
[0067] S602. Distance distributions differ across domains (style / noise / sharpness); domain-specific maintenance makes "outlier" values comparable. Separately maintain the mean and variance for the source and target domains:
[0068]
[0069]
[0070] in, , These represent the mean and standard deviation of the distance within the corresponding domain. For EMA step size, For numerically stable terms, It is the standardized distance between the input sample instance features and the target agent in the cluster agent memory. The greater the noise, the larger it is.
[0071] S603, Calculate distance confidence:
[0072] in, The lower limit for weights is set to 0.05 by default to avoid the problem of not learning due to excessively low weights. The slope of the sigmoid function. This is the sigmoid function.
[0073] S604. Retrieve the cluster size corresponding to the label. Take the 99th percentile on a set of the same size. Returning to the One Foundation:
[0074]
[0075] in, This represents the number of samples in the cluster corresponding to the label. This is a domain-scale set. This paper is unsupervised, and the clustered proxy memory stores source domain proxy features and target domain proxy features separately. Here, domain segmentation refers to distinguishing between the source and target domains. Furthermore, this application removes 99% quantiles. It is used to suppress oversized clusters.
[0076] use By compressing the large cluster advantage and then truncating to [min_w,1], we obtain the size confidence score:
[0077] in, For compression functions, The lower limit for weights is set to 0.05 by default to avoid the problem of not learning due to excessively low weights.
[0078] S605. After obtaining two corresponding confidence levels, perform adaptive fusion of the two confidence levels: Calculate the intra-batch variance:
[0079]
[0080] The fusion coefficient is given by the variance proportion. And crop it to [0.7, 0.95]:
[0081]
[0082] in, , This represents the variance of the two confidence levels in this batch. To prevent zero anomalies by using extremely small constants. The fusion coefficient is used to prevent one confidence level from completely overwhelming the other. The final sample confidence level.
[0083] S606, with As an adaptive write strength, the identity-level features are slid exponentially toward the class center features of the corresponding identity in its memory and L2 normalized. (Background / Noise) Writing is skipped directly; the prototype is pulled by "trusted samples" to gradually fit the current cluster distribution; in conjunction with scale and distance confidence, it is robust to noise / imbalance.
[0084] in, For momentum, the default value is 0.2. This represents the actual write strength; the more reliable the signal, the greater the strength. This serves as the prototype for the label. Using the above formula, when updating the class center features of the corresponding label in the memory using identity-level features, distance confidence and scale confidence are introduced to more reliably influence the class center features.
[0085] S7. The model performs inference in the target domain and outputs cross-domain unsupervised underwater target search results, achieving robust target discovery and matching in complex aquatic environments. 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. A cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory, characterized in that, Includes the following steps: S1. Construct an underwater target search model architecture, wherein the underwater target search model includes a backbone network, a detection head, and a re-identification head; S2. Obtain the source domain underwater image dataset, and pre-train the underwater target search model based on the source domain underwater image dataset to obtain the underwater target search pre-trained model. The source domain underwater image dataset includes labeled underwater image data. S3. Obtain the target domain underwater image dataset and input it along with the source domain underwater image dataset into the underwater target search pre-training model. Extract the identity-level features from the source and target domains respectively, perform clustering on the identity-level features, and write the obtained cluster centers into the cluster proxy memory as the initial prototype. At the same time, count the number of identity-level features corresponding to each identity and register them as the size of the cluster in the gradient-free cluster size buffer. The identity-level features are the features output by the re-identification head in the underwater target search pre-training model. S4. Obtain the region candidate network of the detection head in the underwater target search pre-trained model, input the underwater image data of the target domain into the region candidate network to generate detection boxes, calculate the Jaccard distance of the generated detection boxes, and obtain the initial pseudo-labels according to DBSCAN. S5. Adopt the κ-nearest neighbor strategy, select K nearest neighbors, form the neighborhood posterior by soft voting with distance and temperature weighting, and update the pseudo-labels according to the comparison results of the pseudo-labels of the neighborhood posterior domains, thereby realizing the refinement of pseudo-labels. S6. Use the target domain identity-level features corresponding to the pseudo-labels to train the cluster proxy memory with confidence weighting. The distance confidence is obtained by mapping the cosine distance between the sample and the prototype through the in-domain z-score and sigmoid, and the size confidence is obtained by the cluster size through log compression. The sample weights are adaptively fused into sample weights based on the variance ratio of the distance confidence and the size confidence in the current batch, and the sample weights are optimized and updated. S7. Utilize the underwater target search model trained with cluster proxy memory to perform inference in the target domain and output cross-domain unsupervised underwater target search results.
2. The cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory according to claim 1, characterized in that, The underwater target search model uses ResNet-50 as the backbone network. The outputs of Stage-4 and Stage-5 are fed into the detection head and the re-identification head after multi-scale alignment.
3. The cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory according to claim 1, characterized in that, Construct a stable candidate set using κ-nearest neighbors and select the Top-K voting neighborhoods, including determining the subsequent candidate neighborhoods and voting neighborhoods according to the following methods: in, For the κ candidate neighborhoods of sample i, For the i-th candidate box feature in the target domain, k nearest neighbors For the final voting neighborhood, To select the first k functions.
4. The cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory according to claim 3, characterized in that, The neighborhood posterior is formed by soft voting weighted by distance and temperature, including: The neighborhood posterior weights and the weighted votes for each label are obtained as follows: in, Let be the Euclidean distance between sample i and its neighbor sample j in the feature space. The distance temperature is used to adjust the intensity of distant neighbor inhibition. Let j be the soft weight of neighbor j with respect to sample i. For tags The number of votes in the neighborhood, This is the initial pseudo-label for DBSCAN.
5. The cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory according to claim 4, characterized in that, The pseudo-labels are updated based on the comparison results of the pseudo-labels of the neighboring posterior domains, including: Get the maximum vote tag and its strength: in, For the largest tag, To maximize the strength of the pseudo-label, we check whether the posterior is still better than the prior of the original label, thereby achieving a more refined adjustment of the pseudo-label: in, This represents the prior weight of the original label. This represents the refined label. If the neighborhood of i is entirely noise or the posterior is insufficient to overturn the prior, then the original value is retained. For DBSCAN initial pseudo-labels, This indicates an indicator function, where the proposition is true if it is 1, and false if it is 0. This indicates a change in the label's proportion. Represents the number of samples in the target domain, only if the following condition is met. Furthermore, the pseudo-label of the sample is changed only when there are non-noise labels in the original pseudo-label and its neighborhood. Updated to Otherwise, it remains unchanged, thus making single-point corrections only when there is sufficient evidence in the neighborhood, thereby improving the consistency and purity of false labels.
6. The cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory according to claim 1, characterized in that, The distance confidence score is calculated as follows: in, As the lower limit of the weight, The slope of the sigmoid function. For the sigmoid function, The input sample instance features are the standardized distances between the target agents in the cluster agent memory; Take the cluster size corresponding to the label Take the 99th percentile on a set of the same size. Returning to the One Foundation: in, The sample size of the cluster corresponding to the label. For a set of domain-scale data, the scale confidence score is obtained based on the following calculation: in, The function is a compression function, and the calculated scale confidence level is then scaled down to [min_w, 1], where This is the lower limit of the weight.
7. The cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory according to claim 6, characterized in that, The sample weights are optimized and updated by adaptively fusing the variance proportions of distance confidence and size confidence in the current batch, including: Calculate the variance within the batch: The fusion coefficient is given by the variance proportion. And crop it to [0.7, 0.95]: in, , Let V be the variance of the two confidence levels in this batch. To prevent division by zero anomalies by a minimal constant, The fusion coefficient is... The final sample confidence level.
8. The cross-domain underwater target search method based on pseudo-label refinement and confidence cluster memory according to claim 7, characterized in that, Cluster proxy memory training using confidence-weighted features of the target domain identity also includes: Utilizing the obtained identity-level features For the feature centers of the memory bank class, based on the final sample confidence level and the preset momentum coefficient Calculate update step size And in an exponential sliding manner, the identity-level features are written to the class center features of the corresponding identity in its memory: in, For momentum, the default value is 0.
2. This represents the actual write strength; the more reliable the signal, the greater the strength. This represents the class center features of the corresponding label in the memory. Using the formula above, when updating the class center features of the corresponding label in the memory using identity-level features, distance confidence and size confidence are introduced to more reliably influence the class center features.