CN-121982503-A - Cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memory

CN121982503ACN 121982503 ACN121982503 ACN 121982503ACN-121982503-A

Abstract

The invention discloses a cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memory, and relates to the technical field of underwater target searching. The method comprises the steps of constructing an underwater target search architecture to form a unified model by a feature extractor and a re-identification/detection head which share weights. And introducing confidence weighting cluster proxy memory in training, obtaining distance confidence by using the distance between the sample and the class center cosine, obtaining scale confidence by combining the cluster scale, and adaptively fusing the scale confidence into sample updating weight according to the variance ratio. On the premise of ensuring compatibility with the existing self-adaptive training pipeline, the method purifies the pseudo tag, writes the prototype in a steady mode, effectively inhibits noise accumulation and model drift, and is suitable for cross-domain underwater target searching tasks.

Inventors

WANG HUIBING
LIU XINGCHEN
ZHAO GUOJIAN
ZHANG JIQING
FU XIANPING
FAN JIARUI

Assignees

大连海事大学

Dates

Publication Date: 20260505
Application Date: 20251230

Claims (8)

1. A cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memorization is characterized by comprising the following steps: s1, constructing an underwater target search model framework, wherein the underwater target search model comprises a main network, a detection head and a re-identification head; S2, acquiring a source domain underwater image dataset, pre-training an underwater target search model based on the source domain underwater image dataset, and acquiring an underwater target search pre-training model, wherein the source domain underwater image dataset comprises underwater image data with labels; S3, acquiring a target domain underwater image dataset, inputting the target domain underwater image dataset and the source domain underwater image dataset into an underwater target searching pre-training model, respectively extracting identity-level features of a source domain and a target domain, carrying out clustering treatment on the identity-level features, writing an obtained clustering center into a cluster agent memory library as an initial prototype, and simultaneously counting the number of the identity-level features corresponding to each identity as the scale of the cluster to be registered in a gradient-free cluster scale buffer zone; S4, acquiring a region candidate network of a detection head in the underwater target search pre-training model, inputting target domain underwater image data into the region candidate network to generate a detection frame, and acquiring an initial pseudo tag according to DBSCAN after Jaccard distance calculation of the generated detection frame; S5, adopting a kappa-mutual neighbor strategy, selecting K neighbors, forming a neighbor posterior by soft voting weighted by distance temperature, and updating the pseudo tag according to the comparison result of the pseudo tag in the neighbor posterior domain so as to realize the refinement of the pseudo tag; S6, performing confidence weighted cluster agent memory training by utilizing target domain identity level features corresponding to pseudo labels, mapping a cosine distance between a sample and a prototype through intra-domain z-score and Sigmoid to obtain distance confidence, performing logarithmic compression on cluster scale to obtain scale confidence, and adaptively fusing the variance ratio of the distance confidence and the scale confidence in a current batch into sample weight according to the variance ratio of the distance confidence and the scale confidence to perform sample weight optimization updating; S7, reasoning in a target domain by using a model underwater target search model for completing cluster agent memory training, and outputting a cross-domain unsupervised underwater target search result.
2. The cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memory according to claim 1, wherein the underwater target searching model takes ResNet-50 as a main network, and after Stage-4 and Stage-5 outputs are aligned in multiple scales, the underwater target searching model is fed into a detection head and a re-identification head respectively.
3. The method for cross-domain underwater target search based on pseudo tag refinement and confidence cluster memorization according to claim 1, wherein the stable candidate set is constructed by kappa-mutual neighbor and Top-K voting neighborhood is selected, comprising the following steps of determining the post candidate neighborhood and the voting neighborhood according to the following modes: Wherein, the For the k candidate neighbors of sample i, For the i-th candidate box feature of the target domain, As a k-nearest neighbor, For the final voting neighborhood, To select the first k functions.
4. A method of cross-domain underwater target search based on pseudo tag refinement and confidence cluster memorization as claimed in claim 3, wherein forming a neighborhood posterior with distance temperature weighted soft votes comprises: obtaining the neighborhood posterior weight and the weighted ticket number of each tag according to the following mode: Wherein, the For euclidean distance in feature space between sample i and neighbor sample j, For the distance temperature, for adjusting the far-neighbor inhibition intensity, For soft weights of neighbor j to sample i, Is a label Is used for counting the number of the neighbor votes, The pseudo tag is initialized for DBSCAN.
5. The cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memorization of claim 4, wherein updating the pseudo tag according to the comparison result of the pseudo tag in the neighborhood posterior domain comprises: obtaining the maximum ticket label and the intensity thereof: Wherein, the As a result of the fact that the label is the largest, The strength of the maximum label is checked to determine whether the posterior is superior to the prior of the original label, and then the refinement adjustment of the pseudo label is realized: Wherein, the Representing the prior weight of the original tag, Representing the refined label, if the neighborhood of i is all noise or the posterior is insufficient to override the prior, keeping the original value, For the DBSCAN initial pseudo tag, Indicating the function, proposing to take a 1 true, otherwise 0, Indicating that the ratio of the labels is changed, Representing the number of target domain samples, only if satisfied And when the original pseudo tag and the adjacent non-noise tag exist in the neighborhood thereof, the pseudo tag of the sample is processed by Updated to Otherwise, the label is kept unchanged, so that single-point correction is performed only when the neighborhood evidence is sufficient, and consistency and purity of the pseudo label are improved.
6. The cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memorization according to claim 1, wherein the distance confidence is obtained according to the following calculation: Wherein, the As a lower limit of the weight of the vehicle, For the slope of sigmoid, As a function of the sigmoid, The standardized distance between the sample instance features and the target agents in the cluster agent memory bank is input; Taking the cluster scale corresponding to the label 99% Quantile on the same domain size set The normalization is as follows: Wherein, the For the sample size of the corresponding cluster of labels, For the domain-size set, the size confidence is obtained according to the following calculation: Wherein, the The calculated scale confidence is then staged to [ min_w,1] for the compression function, where Is the lower weight limit.
7. The cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memorization of claim 6, wherein the sample weight optimization updating is performed by adaptively fusing the variance ratio of the distance confidence and the scale confidence in the current batch into the sample weight, and comprises the following steps: Calculating variance within the batch: Giving the fusion coefficient by the variance ratio And cut to [0.7,0.95]: Wherein, the , For the variance of the two confidence scores in the batch, To prevent and eliminate zero abnormality with an extremely small constant, In order for the coefficient of fusion to be a function of, Is the final sample confidence.
8. The cross-domain underwater target search method based on pseudo tag refinement and confidence cluster memory of claim 7, wherein the cluster proxy memory training for confidence weighting is performed by using target domain identity level features, further comprising: Using the identity-level features obtained For the feature center of the memory library, according to the final sample confidence and the preset momentum coefficient Calculating update step size And writing the identity-level features into class center features of corresponding identities in a memory bank in an exponential sliding mode: Wherein, the For momentum, default is 0.2. The more trusted the greater is for the actual write strength. Is the central characteristic of the memory library class corresponding to the label. Through the formula, when the identity-level features update the class center features of the corresponding tags of the memory bank, the distance confidence and the scale confidence are introduced to pull the class center features more reliably.

Description

Cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memory Technical Field The invention relates to the technical field of underwater target search, in particular to a cross-domain underwater target search method based on pseudo tag refinement and confidence cluster memory. Background In recent years, self-adaptive underwater target searching in the unsupervised field mostly adopts a pseudo-supervision-self-training closed loop, namely, source domain labeling data is used for pre-training on a unified detection/re-identification framework, candidate frames and features are generated by relying on target domain non-labeling data, pseudo labels are obtained through clustering, and the pseudo labels are used for back-feeding training so as to be gradually distributed close to the target domain. The route relieves the problem of lack of cross-domain annotation to a certain extent, but two types of key bottlenecks are still faced when the real underwater scene lands. One is that the fake label quality is unstable. The target domain is affected by imaging media, illumination, turbidity, background texture and shielding modes, the characteristic topology among examples tends to be sparse and distorted, so that an initial cluster obtained by a reordering distance +DBSCAN is sensitive to super parameters, boundary samples and sparse small clusters are easily marked as noise or are incorporated into wrong large clusters, and broken clusters of the same identity in different view angles are difficult to merge. And the second is that the memory bank writing is not credible and is easy to drift. Most methods adopt unified momentum or equal weight writing, the contribution of each sample to the cluster center is regarded as the same importance, the stability of a large cluster is different from that of a small cluster in the statistical sense, if the cluster scale is not considered, the equal weight writing is easy to be pulled by occasional noise, and prototype bias and training are unstable. In addition, the existence of the difficult negative sample makes the damage of the error writing to the prototype more amplified, and induces the memory bank drift. Disclosure of Invention In view of the defects of the prior art, the invention provides a cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memorization. Aiming at the problems of unstable imaging medium, complex background structure and easy noise of a pseudo tag in a real underwater scene, the method is used for expanding two main lines which are cleaner around the pseudo tag and more reliable in memory writing, and realizes stable migration and generalization on the premise of no need of target domain labeling. And (3) refining the neighborhood consistency pseudo tag, namely after a unified detection/re-identification architecture is pre-trained on source domain data, obtaining an initial pseudo tag in a target domain by using a reordered Jaccard distance and density cluster, then constructing a stable candidate set by using kappa-mutual neighbors, selecting a Top-K voting neighborhood, forming a neighborhood posterior by using soft voting weighted by distance and temperature, and carrying out single-point correction only when the posterior is obviously superior to the prior of the original tag. The process can effectively repair broken clusters and remove false marks and noise caused by sporadic neighbors, so that consistency and purity of the pseudo tag are improved. Confidence weighted cluster agent memorization, namely, a prototype writing mechanism which is self-adaptively scaled according to the confidence of a sample is provided for the pain points which are easy to drift when the equal weight writing is performed in the self-training stage. The confidence level consists of two pieces of evidence, namely (1) the distance confidence level is obtained by respectively carrying out z-score standardization on the cosine distance of a sample and a prototype agent in a source/target domain, then carrying out Sigmoid mapping and setting a lower limit to ensure that different domains are comparable on the same standard scale, and (2) the scale confidence level is obtained by respectively carrying out normalization and logarithmic compression on a source domain segment and a target domain segment by using 99% of division bits based on the cluster scale in a memory bank to inhibit the statistical bias of a very large cluster. The invention adopts the following technical means: a cross-domain underwater target searching method based on pseudo tag refinement and confidence cluster memorization comprises the following steps: s1, constructing an underwater target search model framework, wherein the underwater target search model comprises a main network, a detection head and a re-identification head; S2, acquiring a source domain underwater image dataset, pre-training an un