CN-121999348-A - Unsupervised underwater target searching method based on Gaussian mixture and cross-domain comparison collaboration

CN121999348ACN 121999348 ACN121999348 ACN 121999348ACN-121999348-A

Abstract

The invention discloses an unsupervised underwater target searching method based on Gaussian mixture and cross-domain contrast collaboration, and relates to the technical field of underwater target searching. The method mainly comprises the steps of capturing distribution conditions of characteristic statistics of a source domain and a target domain, initializing a Gaussian mixture model, providing initial values close to real distribution of data for mean values, weights and covariance of the Gaussian mixture model through clustering, calculating weight and covariance component parameters, screening a most representative target domain statistical mode, combining the characteristic statistics of the source domain and the optimal target domain statistics, mixing through mixed weights of Beta distributed sampling, designing a mask structure, extracting advanced characteristics of the source domain, screening the most unique style characteristics of the target domain in each batch, and constructing cross-domain difficult sample comparison loss. According to the invention, a statistical feature mixer based on a Gaussian mixture model is constructed, and cross-domain difficult samples are designed to compare learning losses at the same time, so that the problem of searching inter-domain style difference of an underwater target in an unsupervised domain self-adaption mode is solved.

Inventors

WANG HUIBING
SUN HAOLIN
WANG JIE
LI SHENGYAO
ZHANG JIQING
FU XIANPING

Assignees

大连海事大学

Dates

Publication Date: 20260508
Application Date: 20251230

Claims (8)

1. An unsupervised underwater target searching method based on Gaussian mixture and cross-domain contrast collaboration is characterized by comprising the following steps: S1, constructing an underwater target search model, wherein the underwater target search model comprises a feature extraction network, an underwater target detection network and a re-identification network, acquiring a source domain data set, and pre-training the underwater target search model based on the source domain data set to acquire an underwater target search pre-training model; S2, extracting image features of underwater image data in a source domain data set based on an area suggestion network in an underwater target search pre-training model to serve as source domain features, acquiring a target domain data set, and extracting image features of the underwater image data in the target domain data set based on the area suggestion network in the underwater target search pre-training model to serve as target domain features, wherein the target domain data set comprises an underwater image without tag information; s3, clustering is carried out on statistics of the target domain features through k-means clustering, initial values close to real distribution of data are provided for the mean value, the weight and the covariance of the Gaussian mixture model, and therefore initialization of the Gaussian mixture model is achieved; s4, screening the most representative target domain statistics as target domain optimal statistics based on the weight components and fitting components obtained through Gaussian mixture model fitting; S5, combining the source domain characteristic statistic and the target domain optimal statistic, and mixing by means of mixing weights of Beta distributed sampling, so that mixed characteristics are obtained; S6, designing a mask structure, and multiplying the mixed features and the mask element by element to obtain advanced features of the source domain; S7, extracting target domain features from the memory library, respectively calculating domain discrimination entropy of each target domain feature, screening the features with minimum domain discrimination entropy based on calculation results, and taking the features as the most unique style features of the target domain in each batch; S8, splicing the advanced features of the source domain and the most unique style features of the target domain, constructing cross-domain difficult sample contrast loss, improving the cross-domain adaptability of the underwater target search model, and performing underwater target search based on the trained underwater target search model.
2. The method for unsupervised underwater target search based on the synergy of gaussian mixture and cross-domain contrast according to claim 1, wherein the distribution of characteristic statistics is obtained according to the following calculation: Wherein, the Is the characteristic average value of the characteristic, Is the characteristic variance For the pixel values of the feature at channel c, spatial position (h, w), =1E—8 for numerical stability, H is the height of the feature map, W is the width of the feature map.
3. The method for unsupervised underwater target search based on the cooperation of Gaussian mixture and cross-domain contrast according to claim 1, wherein the initialization of the Gaussian mixture model is completed according to the following steps: Assigning samples, for each sample Assigning cluster labels The labels correspond to the cluster centers closest to the sample: Wherein, the In order to represent a single sample to be allocated, Representing the center of the kth cluster, For the sample With the kth cluster center A distance therebetween; Updating the center, and for each cluster k, recalculating the center of the cluster : Wherein, the Is the number of samples for the kth cluster.
4. The method for unsupervised underwater target search based on Gaussian mixture and cross-domain contrast collaboration according to claim 1, wherein the target domain optimal statistics are obtained by screening according to the following steps: first, weight components are calculated: Wherein, the Representing the number of samples contained in the kth cluster of k-means, Representing the total number of samples; Secondly, calculating a covariance matrix, for the kth cluster, firstly subtracting a clustering center from a sample to obtain a centralized sample, then calculating covariance of the sample through matrix multiplication, and adding a unit matrix regularization term: Wherein, the As a regular term of the term, =10 -6 , Is a unit matrix, and the covariance matrix is finally obtained Wherein diagonal elements reflect the degree of discretization of each feature, and non-diagonal elements reflect the correlation between two features; selecting the component index with the largest weight : Representative mean value corresponding to the component with the greatest weight Is the first Mean of individual clusters Representative standard deviation corresponding to the component with the greatest weight Is to the first Covariance matrix of individual clusters The 2 nd diagonal element [ ] 2 , 2 ) To the square of the corresponding value and adding a value stabilizing parameter Obtaining; Acquisition of As target domain optimal statistics.
5. An unsupervised underwater target search method based on the synergy of gaussian mixture and cross-domain contrast according to claim 1, wherein the hybrid features are obtained according to the following steps: Obtaining source domain feature statistics And target domain optimal statistics Mixing weights Where α 1 =0.1, α 2 =2.0, then the mixing statistic is defined as: acquiring source domain features B is the batch quantity, C is the channel number, H×W is the spatial dimension, then the final blending characteristics The method meets the following conditions: 。
6. an unsupervised underwater target search method based on gaussian mixture and cross-domain contrast synergy as claimed in claim 1, wherein the mask structure is constructed according to the following manner: The input features are transformed through a sequence network that contains Dropout, linear layer, batch normalization, and ReLU activation: Wherein, the For input features, B is the batch size, and MLP (& gt) consists of two linear layers And ReLU activation composition, realizing characteristic nonlinear transformation, wherein BN (·) is unbiased batch normalization; The Gumbel noise is introduced to enhance the sampling randomness, a hard mask is generated through Top-k selection, noise conforming to Gumbel distribution is added to the score of each dimension, and the formula is as follows: Wherein, the Is noise following Gumbel distribution, generated by inverse transform sampling: , , Controlling the steepness of the distribution as the temperature parameter, wherein the smaller the value is, the closer the output is to the discrete distribution; scoring noisy Selecting the dimension corresponding to the first k highest scores according to the dimension sequence to generate a binary mask, if the ith dimension of the b sample belongs to Top-k 1, Otherwise 0, wherein I.e. the hard mask that is the final output 。
7. The method for unsupervised underwater target search based on the synergy of Gaussian mixture and cross-domain contrast according to claim 1, wherein the most unique style characteristics of the target domain are obtained according to the following steps: first, target domain features are screened from a memory bank, and a target domain feature subset is defined as: Wherein, the For the subset of the target domain features, , The source domain category number; the number of the target domain features is D, and the feature dimension is D; And then to the target domain feature Calculating information entropy through probability distribution output by the domain classifier: Wherein the method comprises the steps of For the ith feature in the target domain feature subset, [0] For the source domain score to be given, [1] For the target domain score to be a score, Representing parameterized domain classifier functions by Defining the mapping from the feature space to the domain decision space; calculating information entropy and quantifying domain attribution uncertainty of the features according to the following formula: Wherein, the Is characterized by The information entropy of (1) is in the range of [0, log 2], Is a numerical stability term; the optimal feature index is then obtained as: Wherein, the Representing a subset of target domain features A feature index with the minimum medium entropy value; And finally, carrying out global index mapping: Wherein, the Representing optimal features in global memory bank Is used as an index into the database, The most unique style characteristics of the target domain are obtained: 。
8. the method for unsupervised underwater target search based on the synergy of Gaussian mixture and cross-domain contrast according to claim 1, wherein the cross-domain difficult sample contrast loss is constructed according to the following method: Will pass through the mask Resulting source domain advanced features Difficult sample feature with target domain Splicing, mapping to a comparison space through a predictor, and obtaining cross-domain difficult sample characteristics ; Based on source domain invariant features And cross-domain difficulty sample characterization The similar feature aggregation and the heterogeneous feature separation are constrained by the difficult sample comparison loss, wherein the calculation formula of the difficult sample comparison loss is as follows: Wherein, the Representation pair and method Homogeneous positive sample set It is desirable to take the form of a program, Representing the embedding of a positive sample in the contrast space, where the positive sample is of the same genus class as the i-th sample Is also taken from the sample of the (C) in the contrast space And ; Indicating the embedding of samples other than the ith sample itself in the contrast space, Is that And a certain positive sample Is scored by the similarity of the temperature Scaling, the higher the score, the more likely it is that the same class, the denominator of the logarithmic part is Similarity score sum with all other samples, representing Total probability of being homogeneous with arbitrary samples.

Description

Unsupervised underwater target searching method based on Gaussian mixture and cross-domain comparison collaboration Technical Field The invention relates to the technical field of underwater target search, in particular to an unsupervised underwater target search method based on Gaussian mixture and cross-domain contrast collaboration. Background With the rapid development of the fields of ocean resource development, underwater security protection, environmental monitoring and the like, the underwater target search technology has a key application value in a dual-purpose scene of military and civil. However, the complexity of the underwater environment causes the target characteristics to present remarkable scene dependence, and meanwhile, the labeling of large-scale underwater data faces the reality dilemmas of high cost, high professional threshold, high labeling difficulty and the like. Under the background, the unsupervised underwater target searching technology can realize cross-scene target detection and identification without depending on target domain labeling data, and becomes an important direction for breaking through data bottleneck and pushing the technology to be practical. However, in practical application, the technology is always limited by a 'domain difference' between a training scene and a real deployment scene, namely, the differences of environmental factors such as illumination conditions, water components, background noise and the like of different water areas, so that the characteristic distribution deviation of the model is obvious when the model is migrated across the scenes, and the generalization performance and the practical value of the model are seriously restricted from being further improved. At present, the key problem to be solved in the unsupervised domain self-adaptive underwater target searching task is how to enable cross-domain knowledge to be effectively migrated under the condition that target domain labeling information is not available. The current mainstream technical scheme is obviously restricted by inter-domain differences, and mainly because systematic inconsistencies exist in the aspects of low-layer visual characteristics such as illumination, colors and the like of a source domain and a target domain, and the inconsistencies can cause cross-domain deviation of characteristic representation, so that remarkable negative influence is generated on generalization performance of a model. Disclosure of Invention In view of the defects of the prior art, the invention provides an unsupervised domain self-adaptive underwater target searching method based on a Gaussian mixture model. According to the invention, the Gaussian mixture model is used for fitting the statistical distribution of the characteristics of the target domain, capturing the overall style of the target domain, and interpolating the source domain characteristic statistic and the target domain statistic by combining the mixed weight of Beta distribution sampling, so that smooth migration of the characteristic style is realized, and the inter-domain difference is reduced. In addition, the discrimination of the cross-domain characteristics is enhanced by constructing the cross-domain difficult sample contrast loss, and the robustness of the identity recognition is improved by contrast learning constraint characteristic clustering. And finally, the identity recognition capability of the model to the underwater target in the target domain is improved, and the problem of reduced matching precision caused by domain offset is solved. The invention adopts the following technical means: An unsupervised underwater target searching method based on Gaussian mixture and cross-domain contrast collaboration comprises the following steps: S1, constructing an underwater target search model, wherein the underwater target search model comprises a feature extraction network, an underwater target detection network and a re-identification network, acquiring a source domain data set, and pre-training the underwater target search model based on the source domain data set to acquire an underwater target search pre-training model; S2, extracting image features of underwater image data in a source domain data set based on an area suggestion network in an underwater target search pre-training model to serve as source domain features, acquiring a target domain data set, and extracting image features of the underwater image data in the target domain data set based on the area suggestion network in the underwater target search pre-training model to serve as target domain features, wherein the target domain data set comprises an underwater image without tag information; s3, clustering is carried out on statistics of the target domain features through k-means clustering, initial values close to real distribution of data are provided for the mean value, the weight and the covariance of the Gaussian mixture model, and therefore initi