CN-121999347-A - Unsupervised domain self-adaptive underwater target searching method based on frequency domain transformation

CN121999347ACN 121999347 ACN121999347 ACN 121999347ACN-121999347-A

Abstract

The invention discloses an unsupervised domain self-adaptive underwater target searching method based on frequency domain transformation, and relates to the technical field of underwater target searching. The method mainly comprises the steps of screening out the most similar characteristic pairs from source domain and target domain characteristics, carrying out Fourier transform on the characteristic pairs, extracting amplitude spectrums and phase spectrums of the source domain and the target domain respectively, designing a mask structure, separating low-frequency information and high-frequency information of the source domain and the target domain, carrying out self-adaptive fusion on the low-frequency information of the source domain and the target domain, carrying out inverse Fourier transform on the fused amplitude spectrums and source domain phase spectrums, realizing self-adaptive fusion of spatial domain local discrimination characteristics and frequency domain global structure representation through learnable dual-path weight distribution, and carrying out cross-domain joint training on an underwater target search model, so that the inter-domain offset problem is relieved. The invention constructs the frequency domain mixer and the space-frequency domain mixer from the angle of the frequency domain so as to solve the problem of searching the style difference among the domains of the self-adaptive underwater target of the unsupervised domain.

Inventors

WANG HUIBING
SUN HAOLIN
WANG JIE
ZHAO GUOJIAN
ZHANG JIQING
FU XIANPING

Assignees

大连海事大学

Dates

Publication Date: 20260508
Application Date: 20251230

Claims (8)

1. An unsupervised domain self-adaptive underwater target searching method based on frequency domain transformation is characterized by comprising the following steps: S1, acquiring a source domain underwater image dataset with tag information and a target domain underwater image dataset without tag information; S2, constructing an underwater target search model of a serialization end-to-end network, and pre-training the underwater target search model based on a source domain underwater image data set to obtain an underwater target search pre-training model; S3, processing the source domain image data and the target domain image data based on a feature extraction network in the underwater target search pre-training model respectively so as to obtain source domain features and target domain features; s4, carrying out Fourier transformation on the source domain features and the target domain features in the fusion feature pair, respectively extracting the amplitude spectrum and the phase spectrum of the source domain and the target domain, and simultaneously designing a mask structure to separate low-frequency and high-frequency information of the amplitude spectrum of the source domain and the target domain; S5, constructing a frequency domain mixer to realize self-adaptive fusion of source domain and target domain amplitude spectrum low-frequency information, and simultaneously reserving source domain phase and high-frequency information; s6, performing inverse Fourier transform on the fused amplitude spectrum and the source domain phase spectrum, so as to obtain a frequency domain reconstruction characteristic; S7, constructing a space-frequency domain mixer, and carrying out weight self-adaptive fusion on the source domain features and the frequency domain reconstruction features through the space-frequency domain mixer to obtain mixed features capable of simultaneously representing the characteristics of the space domain and the frequency domain; s8, performing cross-domain joint training on the underwater target search model, and relieving inter-domain offset problems in subsequent tasks.
2. The method for searching the unsupervised domain adaptive underwater target based on the frequency domain transformation according to claim 1, wherein the similarity between each source domain feature and each target domain feature is quantized according to cosine similarity, so that the most similar feature pair is screened out and used as a fusion feature pair, and the calculation formula of the cosine similarity is as follows: Wherein the method comprises the steps of As a source domain feature vector, As a result of the feature vector of the target domain, The inner product is represented by the number of the inner products, Is the L 2 norm.
3. The method for searching the unsupervised domain adaptive underwater target based on the frequency domain transformation according to claim 1, wherein fourier transformation is performed on the source domain features and the target domain features in the fusion feature pair, and an amplitude spectrum and a phase spectrum of the source domain and the target domain are extracted respectively, comprising: Performing a two-dimensional fourier transform to convert the spatial domain features to the frequency domain: Wherein the method comprises the steps of Is a feature of the spatial domain and, Is a complex characteristic of the frequency domain, Is a frequency coordinate, u corresponds to a frequency component in the horizontal direction, v corresponds to a frequency component in the vertical direction, H, W represent the image height and width respectively, Is a complex exponential basis function that is based on, And then separating the amplitude spectrum and the phase spectrum according to the following formula: Wherein, the Is the real part of the system, Is the imaginary part of the component, In order to be an amplitude spectrum, Is a phase spectrum.
4. The method for unsupervised domain adaptive underwater target searching based on frequency domain transformation according to claim 1, wherein the design mask structure separates the source domain and target domain amplitude spectrum low frequency and high frequency information, comprising: Calculating each frequency point in the frequency domain Distance to center (H/2, W/2)), reserving a low frequency region where distance is less than or equal to the threshold: Wherein, the For the generated low frequency mask, T is the proportion of the low frequency part, Is half the length of the diagonal of the frequency domain; separation of high and low frequency characteristics: Wherein, the Representing the low frequency characteristics of the source domain, Representing the high frequency characteristics of the source domain, Representing the low frequency characteristics of the target domain, Representing the high frequency characteristics of the target domain, Is a mask.
5. The unsupervised domain adaptive underwater target searching method based on frequency domain transformation according to claim 4, wherein the adaptive fusion of the source domain and target domain amplitude spectrum low frequency information is realized based on a frequency domain mixer, and the method comprises the following steps: splicing low-frequency components of a source domain and a target domain, and calculating attention weight through a convolution network Fusion of low frequencies: Wherein the attention weight The definition is as follows: Wherein, the For the fused low-frequency amplitude component, For the channel-dimensional stitching operation, In the case of a two-dimensional convolution operator, The function is activated for Hard-Sigmoid, 。
6. The unsupervised domain adaptive underwater target searching method based on the frequency domain transform according to claim 1, wherein the fused amplitude spectrum and source domain phase spectrum are subjected to inverse fourier transform according to the following formula: Wherein, the Is the fusion result of the frequency domain, which is Wherein For the source domain phase spectrum, Representing the features after inverse fourier transformation of the frequency domain features back into the spatial domain.
7. The method for unsupervised domain adaptive underwater target search based on frequency domain transformation according to claim 1, wherein the step of performing weight adaptive fusion on the source domain features and the frequency domain reconstruction features by a space-frequency domain mixer to obtain the mixed features comprises the steps of: first, frequency component decomposition is carried out, and a low-frequency global component and a high-frequency detail component of a characteristic are separated through convolution operation: Wherein: A large-kernel convolution operator with the parameter theta low is represented, used to extract the low-frequency components, As a spatial feature of the source domain, As a feature of the frequency domain mixing, Is a global low frequency component of the source domain spatial feature, Is a global low frequency component of the frequency domain hybrid feature, Is a global high frequency component of the source domain spatial feature, Global high-frequency components which are frequency domain mixing features; Calculating interaction weights of the source domain and the frequency domain features through a dual-channel attention network: Wherein, the The channel dimension stitching operation is represented as such, And An attention network of "source domain- > frequency domain" and "frequency domain- > source domain" respectively, The method is characterized in that a function is activated for Sigmoid, and the weight value range is ensured to be 0, 1; respectively performing weighted fusion on the global high-frequency component and the global low-frequency component: Wherein, the Representing element-by-element multiplication (Hadamard product); And finally, carrying out feature fusion, and obtaining a mixed feature by dynamically balancing high-frequency and low-frequency fusion results through a learnable coefficient: Wherein, the The contribution ratio of the low frequency components is controlled for the learnable parameters.
8. The method for searching the unsupervised domain self-adaptive underwater target based on the frequency domain transformation according to claim 7, wherein the attention network is realized by a multi-scale gating structure, and the specific steps are as follows: Global low frequency component to source domain features Global low frequency component of frequency domain features Three scale features are generated: wherein AvgPool is an average pooling operation, and global high-frequency information processing is performed in this way; and firstly, channel splicing is carried out on the features of each scale, and then gating weights are generated through convolution operation and Sigmoid: Wherein, the For the connection of the channels, Activating Sigmoid, then up-sampling different scale weights to original size through multi-scale weight fusion, and weighting and summing the different scale weights with large scale weights to obtain weights : Wherein, the Representing the up-sampling of the sample, In order for the parameters to be able to be learned, And (5) weight normalization is ensured.

Description

Unsupervised domain self-adaptive underwater target searching method based on frequency domain transformation Technical Field The invention relates to the technical field of underwater target searching, in particular to an unsupervised domain self-adaptive underwater target searching method based on frequency domain transformation. Background In recent years, with the rapid development of deep learning technology, the underwater target searching technology also has significantly advanced, the detection and identification precision is greatly improved, and a key technical support is provided for exploring the underwater environment. However, in practical applications, the problem of "domain gap" between the training environment and the practical use environment is more and more prominent, and is a problem to be solved. To address this challenge, unsupervised domain adaptive underwater target search techniques have evolved. The core idea is to migrate knowledge learned by the marked data set to a target domain without marked information so as to enhance the adaptability and generalization capability of the model under different environments. At present, a core challenge faced in the task of unsupervised domain adaptive underwater target search is how to achieve efficient cross-domain knowledge migration in the absence of target domain labeling. The current mainstream method is limited by inter-domain differences, and mainly comprises that systematic differences exist between source domains and target domains on low-layer visual characteristics such as illumination, colors and the like, and the differences cause cross-domain deviation of characteristic representation, so that the generalization performance of a model is obviously reduced. 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 frequency domain transformation. From the perspective of a frequency domain, the model is adapted to the style change of a target domain in advance through the frequency domain mixer, and the capture capacity of the space domain to fine granularity characteristics and the adaptation advantage of the frequency domain to macroscopic scene change are fully exerted through the information of the space-frequency domain bridging Fourier and the space domain, so that the inter-domain difference problem is better solved. The invention adopts the following technical means: an unsupervised domain self-adaptive underwater target searching method based on frequency domain transformation comprises the following steps: S1, acquiring a source domain underwater image dataset with tag information and a target domain underwater image dataset without tag information; S2, constructing an underwater target search model of a serialization end-to-end network, and pre-training the underwater target search model based on a source domain underwater image data set to obtain an underwater target search pre-training model; S3, processing the source domain image data and the target domain image data based on a feature extraction network in the underwater target search pre-training model respectively so as to obtain source domain features and target domain features; s4, carrying out Fourier transformation on the source domain features and the target domain features in the fusion feature pair, respectively extracting the amplitude spectrum and the phase spectrum of the source domain and the target domain, and simultaneously designing a mask structure to separate low-frequency and high-frequency information of the amplitude spectrum of the source domain and the target domain; s5, constructing a frequency domain mixer to realize self-adaptive fusion of low-frequency information of amplitude spectrums of a source domain and a target domain, and simultaneously, reserving phase and high-frequency information of the source domain to enable a model to be adapted to style information of the target domain in advance; s6, performing inverse Fourier transform on the fused amplitude spectrum and the source domain phase spectrum, so as to obtain a frequency domain reconstruction characteristic; S7, constructing a space-frequency domain mixer, and carrying out weight self-adaptive fusion on the source domain features and the frequency domain reconstruction features through the space-frequency domain mixer to obtain mixed features capable of simultaneously representing the characteristics of the space domain and the frequency domain; s8, performing cross-domain joint training on the underwater target search model, and relieving inter-domain offset problems in subsequent tasks. Further, the similarity of each source domain feature and each target domain feature is quantified according to the cosine similarity, so that the most similar feature pair is screened out and used as a fusion feature pair, wherein the calculation formula of the cosine similarity is