An underwater target perception enhancement method based on target domain observation information guidance
By guiding target domain observation information in underwater target perception, establishing semantic labels and multi-head decoding models, the ill-posed augmentation problem caused by the quality difference between the source and target domains is solved, achieving well-posed augmentation under different attenuation conditions, and improving the perception quality of underwater target detection and autonomous sensing robots.
Patent Information
- Application Number
- CN202411116446.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2026-11-03
- Estimated Expiration
- 2044-08-14
AI Technical Summary
Existing technologies suffer from unsteady underwater target perception enhancement problems due to quality differences between the source and target domains, making it difficult to effectively improve the performance and accuracy of vision tasks under different attenuation conditions.
By establishing semantic labels for fuzzy decay, strong decay, and multiple decay features, a reference ground truth dataset is generated. Low-quality data is generated and features are extracted. A multi-head decoding model is trained, and combined with decay feature pseudo-labels and responders, the enhancement model is guided to generate significant texture features in the target domain, achieving appropriate enhancement effects.
It significantly improves the perception quality of underwater target detection and autonomous sensing robots under different attenuation conditions, solves the ill-posed enhancement problem between the target domain and the source domain, and improves the operational performance of the vision system.
Smart Images
Figure CN118918456B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of target perception enhancement, and specifically relates to an underwater target perception enhancement method guided by target domain observation information. Background Technology
[0002] Under the supervision of prior domain knowledge or data, current algorithms exhibit good perceptual quality enhancement performance within the source domain (training set). However, in practical applications, the diversity of the target domain (operating environment) is a crucial factor affecting the performance and accuracy of visual tasks. When there are quality differences between the source and target domains, the identical distribution assumption is not satisfied, making it difficult for enhancement patterns trained on the source domain to be applied to the target domain, resulting in ill-posed enhancement results.
[0003] To improve the adaptability of underwater target domains under different attenuation conditions, this invention proposes an underwater target perception quality enhancement method guided by target domain observation information. First, considering the differences in attenuation characteristics among different types of water bodies, semantic labels for different attenuation characteristics are established. Second, addressing the lack of high-dimensional semantic labels in target domain attenuation information, an attenuation label generator and responder are established based on clustering technology to guide the enhancement model. Then, at the model level, enhancement modes responding to different semantic labels are designed, and at the data level, reference ground truths corresponding to different semantic labels are established to enrich the source domain information to respond to different attenuation characteristics. Finally, guided by target domain attenuation information, the enhancement method can significantly improve the cross-domain enhancement effect and enhance the practical application value of the enhanced target domain image. Summary of the Invention
[0004] (a) Technical problems to be solved
[0005] The technical problem to be solved by this invention is how to provide an underwater target perception enhancement method guided by target domain observation information, so as to solve the ill-posed enhancement problem caused by the quality difference between the target domain and the source domain.
[0006] (II) Technical Solution
[0007] To address the aforementioned technical problems, this invention proposes an underwater target perception enhancement method guided by target domain observation information, which includes the following steps:
[0008] S1. Attenuation Feature Establishment Steps: Based on the differences in attenuation features of different types of water bodies, semantic labels for fuzzy attenuation, strong attenuation, and multiple attenuation features are established.
[0009] S2. Reference ground value generation steps for different decay characteristics: Extract different features from the reference ground value dataset to obtain reference ground values corresponding to different decay characteristics, thus enriching the source domain information at the data level.
[0010] S3. Low-quality data generation steps: The reference ground truth dataset is subjected to context feature degradation and texture feature degradation to obtain a low-quality dataset;
[0011] S4. Training data generation steps: Perform different feature extraction processes on the dataset with texture feature attenuation to obtain texture feature enhancement results corresponding to different attenuation features, which are used for feature extraction of the data specific to the target scene.
[0012] S5. Multi-head decoding model training steps: During the training process, data augmentation is used to achieve a tight connection between the feature extraction model and the multi-head decoding model at the algorithm level. During the training process, the L1-norm is used as the loss function for training.
[0013] S6. Steps for establishing pseudo-labels for attenuation features. Cluster the training data to establish the prior distribution of attenuation features, statistically analyze the main attenuation features, and generate pseudo-labels to describe the attenuation features of the observed input in the target domain;
[0014] S7. Observation information extraction steps: Establish an attenuation feature generator to output attenuation feature labels based on the attenuation features of the target domain;
[0015] S8. Attenuation Feature Response Step: An attenuation label responder is established based on multi-focus image fusion technology to generate an image and fuse significant texture features, outputting a perceptual quality enhancement result.
[0016] (III) Beneficial Effects
[0017] This invention proposes an underwater target perception enhancement method guided by target domain observation information. This method addresses the ill-posed enhancement problem caused by the quality difference between the target domain and the source domain. By using the attenuation features of the target domain image as semantic labels, this invention guides the source domain model to generate and fuse significant texture and contextual features, achieving well-posed enhancement of underwater images under different attenuation conditions. This method has significant application value in underwater target detection, key object perception, and autonomous sensing robots.
[0018] The method proposed in this invention uses the attenuation features of the target domain image as semantic tags, which can guide the source domain enhancement model to generate images and fuse significant texture features. This achieves adaptive enhancement of the perception quality of underwater observation targets under different attenuation conditions, improves the operational performance of the vision system, and has important application value in underwater target detection, key object perception, autonomous perception robots and other fields. Attached Figure Description
[0019] Figure 1 These are the construction steps of the underwater target perception quality enhancement method guided by target domain observation information according to the present invention;
[0020] Figure 2 This is a model structure diagram for extracting different attenuation features according to the present invention;
[0021] Figure 3 This is a model structure diagram of multi-head decoding according to the present invention;
[0022] Figure 4 This is a model structure for an attenuation tag generator according to the present invention. Detailed Implementation
[0023] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.
[0024] This invention relates to an underwater target perception quality enhancement method guided by target domain observation information. Specifically, it addresses the ill-posed enhancement problem caused by the quality difference between the target domain and the source domain by designing a method that can generate appropriate enhancement results in the target domain under different attenuation conditions according to the actual attenuation state.
[0025] This invention discloses an underwater target perception quality enhancement method guided by target domain observation information. The method includes: (1) Attenuation feature establishment step. Semantic labels for fuzzy attenuation, strong attenuation, and multiple attenuation features are established based on the differences in attenuation features of different types of water bodies. (2) Reference ground value generation step for different attenuation features. Different features are extracted from the original reference ground value dataset to obtain reference ground values corresponding to different attenuation features, enriching the source domain information at the data level. (3) Low-quality data generation step. The original reference ground value dataset is subjected to context feature degradation and texture feature degradation processing to obtain a low-quality dataset. (4) Training data generation step. The texture feature attenuation dataset is subjected to different feature extraction processing to obtain texture feature enhancement results corresponding to different attenuation features, which are used to perform target scene-specific feature extraction on the data. (5) Multi-head decoding model training step. During the training process, the feature extraction model and the multi-head decoding model are tightly connected at the algorithm level by data augmentation to reduce accumulated error. During the training process, the L1 norm is used as the loss function for joint optimization. (6) Attenuation feature pseudo-label establishment step. Cluster the training data, establish the prior distribution of attenuation features, statistically analyze the main attenuation features and generate pseudo-labels to describe the attenuation features of the target domain observation input. (7) Observation information extraction step. Establish an attenuation feature generator and output semantic labels based on the attenuation features of the target domain. (8) Attenuation feature response step. Establish an attenuation label responder, guide the enhancement model to generate images based on semantic labels and fuse significant texture features, and output the perceptual quality enhancement result.
[0026] The purpose of this invention is to provide an underwater target perception quality enhancement method based on target domain observation information, which solves the ill-posed enhancement problem caused by the quality difference between the target domain and the source domain. Under the guidance of target domain observation information, the enhancement model can significantly improve the cross-domain enhancement effect, thereby improving the perception quality of the observed objects in the target domain.
[0027] To achieve the above objectives, this invention proposes an underwater target perception quality enhancement method guided by target domain observation information, the method comprising:
[0028] S1. Attenuation Feature Establishment Steps: Based on the differences in attenuation features of different types of water bodies, semantic labels for fuzzy attenuation, strong attenuation, and multiple attenuation features are established.
[0029] S2. Reference ground value generation steps for different decay characteristics: Extract different features from the reference ground value dataset to obtain reference ground values corresponding to different decay characteristics, thus enriching the source domain information at the data level.
[0030] S3. Low-quality data generation steps: The reference ground truth dataset is subjected to context feature degradation and texture feature degradation to obtain a low-quality dataset;
[0031] S4. Training data generation steps: Perform different feature extraction processes on the dataset with texture feature attenuation to obtain texture feature enhancement results corresponding to different attenuation features, which are used for feature extraction of the data specific to the target scene.
[0032] S5. Multi-head decoding model training steps: During training, data augmentation is used to achieve a tight connection between the feature extraction model and the multi-head decoding model at the algorithm level, reducing accumulated error. The L1-norm is used as the loss function during training.
[0033] S6. Steps for establishing pseudo-labels for attenuation features. Cluster the training data to establish the prior distribution of attenuation features, statistically analyze the main attenuation features, and generate pseudo-labels to describe the attenuation features of the observed input in the target domain;
[0034] S7. Observation information extraction steps: Establish an attenuation feature generator to output attenuation feature labels based on the attenuation features of the target domain;
[0035] S8. Attenuation Feature Response Step: An attenuation label responder is established based on multi-focus image fusion technology to generate an image and fuse significant texture features, outputting a perceptual quality enhancement result.
[0036] Example 1:
[0037] Figure 1This is a training flowchart for a method to enhance the perception quality of underwater observation targets based on observation information, as described in this invention. Figure 1 As shown, the method includes:
[0038] (1) Steps for establishing attenuation features. Based on the differences in attenuation features of different types of water bodies, semantic labels for fuzzy attenuation, strong attenuation, and multiple attenuation features are established, namely L1, L2, and L3.
[0039] In practical implementation, data with different pixel intensities in the training data are labeled, where: I i Strong decay data with a value of →0 are labeled L1, I i →1 fuzzy attenuation data is labeled L2, 0 i Multiple decay data with a value less than 1 are labeled L3, i.e.:
[0040] Label data{I i →0, L1}{i};
[0041] Label data{I i →1, L2}{i};
[0042] Label data{0 i <1, L3}{i}.
[0043] (2) The process of generating reference true values with different attenuation characteristics. For example... Figure 1 As shown, different features are extracted from the reference ground truth dataset Y as needed to obtain the reference ground truth corresponding to different decay features, thus enriching the source domain information at the data level, i.e., f F :Y→{Y L1 Y L2 Y L3}
[0044] In practical implementation, the original reference truth data Y is used as the basis for feature extraction model f. F Feature extraction is performed to obtain reference ground values corresponding to different attenuation features, thus enriching the source domain information at the data level, i.e., f F :Y→{Y L1 Y L2 Y L3};like Figure 2 As shown. First, Y is mapped to the dual space via T1; then, the data of label L1 is normalized in the dual space using Layer. Norm To enhance the latent texture representation in strongly attenuated images, Retinex operations are performed on the data labeled L2 and L3. Retinex To enhance blurred texture details in images with blur attenuation and multiple attenuation; subsequently, multi-scale feature fusion operations are performed simultaneously.Fusion Then, perform the dual inverse mapping T2 to map the data back to the RGB space; subsequently, perform Retinex operations f on the data of labels L1 and L3 in the RGB space. Retinex To avoid artifact noise; finally, the output dataset {Y} L1 Y L2 Y L3}
[0045] (3) Low-quality data generation steps. The reference ground truth dataset is subjected to context feature degradation and texture feature degradation to obtain a low-quality dataset.
[0046] In specific implementation, X∈{X1∪X2} represents the low-quality dataset, where: X1 is generated by the texture feature degenerate G1 from the dataset Y1, where Y1∈Y; X2 is generated by the context feature degenerate G2 from the dataset Y2, where Y2∈Y.
[0047] (4) Training data generation steps. The dataset with texture feature attenuation is processed by different feature extraction processes to obtain texture feature enhancement results corresponding to different attenuation features, which are used to perform feature extraction specific to the target scene.
[0048] In practical implementation, the dataset X1 with texture feature attenuation is based on the feature extraction model f. F Different feature extractions are performed to obtain texture feature enhancement results corresponding to different attenuation features. The processing procedure is as follows: Figure 2 As shown, f F : X1→X'1={X' 1:L1 , X' 1:L2 , X' 1:L3}、x'1∈{X' 1:L1 , X' 1:L2 , X' 1:L3}
[0049] (5) Multi-head decoding model training steps. During the training process, data augmentation is used to implement the feature extraction model f. F With multi-head decoding model f D:Li (f E Tight connections at the algorithmic level are used to reduce accumulated error. During training, the L1 norm is used as the loss function.
[0050] In specific implementation, such as Figure 3 As shown, the dataset {X2∪X'1} is input into the multi-head decoding model f. D:Li (f E ), to obtain the predicted output, then f D:Li (f E ): X'1→P1:Li f D:Li (f E ): X2→P 2:Li f D:Li Based on an encoder-decoder architecture, a max-pooling layer is used to perform 1 / 2 downsampling, a deconvolutional layer is used to perform ×2 upsampling, and a U-NET structure of depth-cascaded layers is used to implement the encoder f. E With different decoders f D:Li For cross-links, the L1 norm is used for model optimization during training, i.e.:
[0051] P 1:Li =f D:Li (f E (X'1,θ E ), s E θ D );
[0052] P 2:Li =f D:Li (f E (X2, θ) E ), s E θ D );
[0053] argmin Loss(Y Li , {P 1:Li P 2:Li})=|Y Li -{P 1:Li P 2:Li}|.
[0054] Among them, L i Used to index different decay feature labels {L1, L2, L3}, θ E Indicates encoder parameters; θ D Indicates decoder parameters; s E Indicates the characteristics of transmission across links; Y Li The reference true values representing different attenuation characteristics are output from step (2); {P 1:Li P 2:Li} represents the predicted output, which is output from step (5).
[0055] (6) Steps for establishing pseudo-labels for attenuation features. Cluster the dataset {X2∪X'1} to establish the prior distribution of attenuation features, then statistically analyze the main attenuation features and generate pseudo-labels to describe the I of the observed input in the target domain. in The attenuation characteristics.
[0056] In specific implementation, such as Figure 4As shown, the prior distribution [ρ1, ρ2, ..., ρ'1] of the decay feature of the dataset {X2∪X'1} (i.e., the source domain data) is established based on the K-Means clustering algorithm. n The clustering algorithm's k-value is set to 3; its main decay characteristics are statistically analyzed and pseudo-labels {Label1, Label2, ..., Label...} are generated. n}, Used to describe the underwater image I of the target area in Attenuation characteristics L in , where n represents the number of source domain images.
[0057] (7) Observation information extraction steps. Establish an attenuation feature generator to output attenuation feature labels based on the attenuation features of the target domain.
[0058] When performing underwater image enhancement tasks in the target domain, such as Figure 4 As shown, the attenuation tag responder G d Extracting I based on K-Means clustering algorithm in attenuation characteristic distribution And in the pseudo-label distribution [ρ1, ρ2, ..., ρ n Using KNN to retrieve and label distributions The three prior distributions ρ with the smallest bias i And based on its significance, I in Attenuation characteristics L in Label for attenuation type i Label i ∈{L1, L2, L3}.
[0059] (8) Attenuation Feature Response Step. An attenuation label responder is established based on multi-focus image fusion technology to generate an image and fuse significant texture features, outputting a perceptual quality enhancement result, thereby obtaining an appropriate enhancement result.
[0060] In specific implementation, such as Figure 1 As shown in (b), the tag responder G is first decayed. r According to I in The tag L in The corresponding trained model f F encoder f E decoder f D:Li Output image I D:Li The image is marked as the fusion base image. Then, based on the principle of salient focus, the texture features of the remaining decoder output images are extracted. Finally, multi-focus fusion is performed to obtain the enhanced result I. output .
[0061] Example 2:
[0062] A method for enhancing the perception quality of underwater observation targets based on observation information includes:
[0063] (1) Establishment of attenuation characteristics.
[0064] (2) Generation of reference true values with different attenuation characteristics.
[0065] (3) Low-quality data generation.
[0066] (4) Training data generation.
[0067] (5) Training of multi-head decoding model.
[0068] (6) Establishment of pseudo-labels for attenuation features.
[0069] (7) Extraction of observation information.
[0070] (8) Attenuation characteristic response.
[0071] Furthermore, in the step of generating reference ground truth values for different attenuation characteristics, corresponding feature extraction models are designed for different attenuation labels. These include:
[0072] The model, based on the duality principle, inverts the dataset Y to highlight degraded texture features. For data with strong degradation, a normalization operation is employed to improve IG. i →0 region pixel intensity to improve potential texture representation; on the other hand, mapping different pixel intensities to the same interval constrains the solution space of the dual function.
[0073] Furthermore, in the multi-head decoder training step, a joint optimization approach is used to train the multi-head decoder to improve model stability. This includes:
[0074] f D A multi-output encoder-decoder f D:Li (f E Composed of, L i ∈{L1, L2, L3} is used to index different attenuation characteristics. For example... Figure 1 As shown, data augmentation is used during training to achieve tight connections at the algorithm level, and the loss function is minimized to reduce the error accumulation caused by model-level cascading in practical applications.
[0075] Furthermore, in the observation information extraction step, a semantic label generator for attenuation features is designed to be used based on the actual observation scenario, including:
[0076] In the source domain (training set {X2∪X'1}), a prior distribution of decay features [ρ1, ρ2, ..., ρ'1] is established based on the K-Means clustering algorithm. n]; Statistically analyze its main decay characteristics and generate pseudo-labels {Label1, Label2, ..., Label n}, Used to describe the underwater image I of the target area in The attenuation characteristics are denoted by n, where n represents the number of source domain images; when performing the task of enhancing underwater images in the target domain, I is extracted based on the K-Means clustering algorithm. in attenuation characteristic distribution And in the pseudo-label distribution [ρ1, ρ2, ..., ρ n Using KNN to retrieve and label distributions The three prior distributions ρ with the smallest bias i And based on its significance, I in Attenuation characteristics L in Labeled as attenuation type Label i Label i ∈{L1, L2, L3}.
[0077] Furthermore, in the attenuation feature response step, an attenuation label responder is designed based on multi-focus image fusion technology to guide the enhancement method to process the attenuation features of the observed object in the actual scene, including:
[0078] Labeling contextual features, constructing salient texture feature maps, and multi-focus fusion. Attenuation label responder G. r According to I in The tag L in , and the corresponding trained decoder f D:Li Output image I D:Li The image is marked as the fusion base, and then the texture features of the remaining decoder output images are extracted based on the principle of salient focus. Finally, multi-focus fusion is performed to obtain the enhanced result.
[0079] The underwater target perception enhancement method proposed in this invention, guided by target domain observation information, can solve the ill-posed enhancement problem caused by the quality difference between the target domain and the source domain. This invention uses the attenuation features of the target domain image as semantic labels to guide the source domain model to generate and fuse significant texture and contextual features, achieving well-posed enhancement of underwater images under different attenuation conditions. This method has significant application value in underwater target detection, key object perception, and autonomous sensing robots.
[0080] The method proposed in this invention uses the attenuation features of the target domain image as semantic tags, which can guide the source domain enhancement model to generate images and fuse significant texture features. This achieves adaptive enhancement of the perception quality of underwater observation targets under different attenuation conditions, improves the operational performance of the vision system, and has important application value in underwater target detection, key object perception, autonomous perception robots and other fields.
[0081] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for enhancing underwater target perception based on target domain observation information, characterized in that, The method includes the following steps: S1. Attenuation Feature Establishment Steps: Based on the differences in attenuation features of different types of water bodies, semantic labels for fuzzy attenuation, strong attenuation, and multiple attenuation features are established. S2. Steps for generating reference ground values with different decay characteristics: Extract different features from the reference ground value dataset to obtain reference ground values corresponding to different decay characteristics, thereby enriching the source domain information at the data level. S3. Low-quality data generation steps: The reference ground truth dataset is subjected to context feature degradation and texture feature degradation to obtain a low-quality dataset; S4. Training data generation steps: Perform different feature extraction processes on the dataset with texture feature attenuation to obtain texture feature enhancement results corresponding to different attenuation features, which are used for feature extraction of the data specific to the target scene. S5. Multi-head decoding model training steps: During the training process, data augmentation is used to achieve a tight connection between the feature extraction model and the multi-head decoding model at the algorithm level. During the training process, the L1-norm is used as the loss function for training. S6. Steps for establishing pseudo-labels for attenuation features: Cluster the training data, establish the prior distribution of attenuation features, statistically analyze the main attenuation features and generate pseudo-labels to describe the attenuation features of the observed input in the target domain. S7. Observation information extraction steps: Establish an attenuation feature generator to output attenuation feature labels based on the attenuation features of the target domain; S8. Attenuation Feature Response Step: An attenuation label responder is established based on multi-focus image fusion technology to generate images and fuse significant texture features, outputting perceptual quality enhancement results; in, S2 includes: based on the original reference truth data Y using a feature extraction model. f F Feature extraction is performed to obtain reference ground values corresponding to different attenuation features, thereby enriching the source domain information at the data level. f F :Y→{Y L1 Y L2 Y L3 First, Y is passed through... T 1. Map to the dual space; then, perform a normalization operation on the data labeled L1 in the dual space. Norm To enhance the latent texture representation in strongly attenuated images, Retinex operations are performed on the data labeled L2 and L3. f Retinex To enhance blurred texture details in images with blur attenuation and multiple attenuation; subsequently, multi-scale feature fusion operations are performed simultaneously. f Fusion And perform dual inverse mapping. T 2. Map the data back to RGB space; then, perform Retinex operations on the data of labels L1 and L3 in RGB space. f Retinex To avoid artifact noise; finally, the output dataset {Y} L1 Y L2 Y L3 }; The S3 includes: using X∈{X1∪X2} to represent the low-quality dataset, where: X1 is calculated from dataset Y1 through a texture feature degenerate. G 1 is generated, where Y1∈Y; X2 is generated from dataset Y2 through a context feature degenerate. G 2 is generated, where Y2∈Y.
2. The underwater target perception enhancement method based on target domain observation information as described in claim 1, characterized in that, S1 includes: labeling data with different pixel intensities in the training data, wherein: I i Strong decay data with a value of →0 are labeled L1, I i →1 fuzzy attenuation data is labeled L2, 0 i Multiple decay data with a value less than 1 are labeled L3, i.e.: Label data{I i →0,L1}{ i }; Label data{I i →1,L2}{ i }; Label data{0 i <1, L3}{ i }。 3. The underwater target perception enhancement method based on target domain observation information as described in claim 1, characterized in that, The S4 includes: applying the texture feature attenuation dataset X1 based on a feature extraction model. f F Different feature extractions are performed to obtain texture feature enhancement results corresponding to different attenuation features; that is, f F X1→X ’ 1={X ’ 1:L1 X ’ 1:L2 X ’ 1:L3 }、 x ’ 1∈{X ’ 1:L1 X ’ 1:L2 X ’ 1:L3 } 4. The underwater target perception enhancement method based on target domain observation information as described in claim 3, characterized in that, S5 includes: dividing the dataset {X2∪X...} ’ 1} Input to the multi-head decoding model f D: Li ( f E ), to obtain the predicted output, then we have f D: Li ( f E ):X ’ 1→P 1: Li , f D: Li ( f E ): X2→P 2: Li ; f D: Li Based on an encoder-decoder architecture, a max pooling layer is used to perform 1 / 2 downsampling, a deconvolutional layer is used to perform ×2 upsampling, and a U-NET structure of depth-cascaded layers is used to implement the encoder. f E With different decoders f D: Li For cross-links, the L1 norm is used for model optimization during training, that is: P 1: Li = f D: Li ( f E (X ’ 1, θ E ), s E , θ D ); P 2: Li = f D: Li ( f E (X2, θ E ), s E , θ D ); argmin Loss(Y Li ,{P 1: Li ,P 2: Li }) = |Y Li - {P 1: Li ,P 2: Li }| Among them, L i Used to index different decay feature labels {L1, L2, L3} θ E Indicates encoder parameters; θ D Indicates decoder parameters; s E Indicates the characteristics of transmission across links; Y Li The reference true values representing different attenuation characteristics are output by step S2; {P 1: Li P 2: Li } represents the predicted output, which is output by step S5.
5. The underwater target perception enhancement method based on target domain observation information as described in claim 4, characterized in that, S6 includes: transferring the source domain dataset {X2∪X} ’ 1) Establishing a prior distribution of decay features based on clustering algorithm The clustering algorithm's k-value is set to 3; its main decay characteristics are statistically analyzed and pseudo-labels {Label1, Label2, ..., Label...} are generated. n }, Label i ∈{L1, L2, L3}, used to describe the underwater image I of the target domain. in Attenuation characteristics L in ,in This indicates the number of source domain images.
6. The underwater target perception enhancement method based on target domain observation information as described in claim 5, characterized in that, S7 includes: when performing the task of enhancing the underwater image of the target domain, attenuating the tag responder G d Extracting I based on clustering algorithm in attenuation characteristic distribution and in pseudo-label distribution Using KNN to retrieve and label distributions The three prior distributions with the smallest bias And based on its significance, I in Attenuation characteristics L in Label for attenuation type i Label i ∈{L1, L2, L3}.
7. The underwater target perception enhancement method based on target domain observation information as described in claim 6, characterized in that, The clustering algorithm is the K-Means clustering algorithm.
8. The underwater target perception enhancement method based on target domain observation information as described in claim 6, characterized in that, The S8 includes: firstly, attenuating the tag responder G r According to I in The tag L in The corresponding trained model f F encoder f E decoder f D: Li Output image I D: Li The image is marked as the fusion base image. Then, based on the principle of salient focus, the texture features of the remaining decoder output images are extracted. Finally, multi-focus fusion is performed to obtain the enhanced result I. output .