Underwater image recognition method, device and system for data scarcity and domain migration, and storage medium
By utilizing land image data for transfer learning in underwater image recognition, quantifying inter-domain differences and injecting Gaussian noise, and combining gradient inversion and orthogonal constraints, the problems of scarce and poor-quality underwater image data are solved, thereby improving recognition accuracy and model adaptability.
Patent Information
- Application Number
- CN202512023813.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-07
AI Technical Summary
In underwater image recognition, underwater image data is scarce and of poor quality due to the constraints of complex lighting conditions and the influence of the water medium. Existing methods are unable to effectively reduce residual domain differences in cross-domain recognition tasks, resulting in insufficient feature discriminativeness and low recognition accuracy.
By modeling uncertainty based on feature distribution differences and learning personalized features, we utilize land image data for transfer, quantify inter-domain differences and inject Gaussian noise, and combine gradient inversion and orthogonal constraints to enhance the discriminative ability of target domain features.
It effectively reduces prediction uncertainty in cross-domain underwater image recognition, improves the model's generalization ability and recognition accuracy, and adapts to complex underwater environments.
Smart Images
Figure CN121811225A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of underwater image recognition and cross-domain transfer learning technology, specifically to underwater image recognition methods, devices, systems, and storage media for data scarcity and domain offset. Background Technology
[0002] With the rapid development of marine exploration, environmental monitoring, and intelligent sensing technologies, underwater image recognition has demonstrated significant application value in marine scientific research, resource exploration, ecological environment monitoring, and autonomous navigation of underwater robots. However, the complex and ever-changing underwater environment poses severe challenges to image acquisition and recognition. Although data-driven methods, represented by deep learning and support vector machines, have achieved remarkable results in the field of image recognition, their performance is highly dependent on large-scale, high-quality labeled samples and requires that the training and test data follow the same distribution.
[0003] However, in real-world underwater scenarios, underwater images often exhibit degradation phenomena such as color shift, decreased contrast, and blurred / distorted details due to complex lighting conditions, the absorption and scattering of light by the water medium, and equipment performance limitations. Variations in water depth and turbidity further exacerbate uneven lighting and information loss, severely impacting the reliability of feature extraction and target recognition. Furthermore, underwater image acquisition and annotation are costly, and the availability of high-quality annotated samples is extremely limited by the shooting environment and manual annotation conditions. Data scarcity has become a key bottleneck restricting the training effectiveness and generalization performance of deep learning models.
[0004] Domain adaptation methods in transfer learning can achieve knowledge transfer from terrestrial scenes to underwater environments by reducing the distributional differences between the source and target domains, thus alleviating performance bottlenecks caused by insufficient data to some extent. However, existing traditional domain adaptation methods still suffer from insufficient feature discriminative power and high prediction uncertainty due to residual domain differences when dealing with complex cross-domain underwater tasks. How to effectively reduce residual domain differences in cross-domain recognition tasks and improve and mine the model's ability to discriminate target domain image features and recognition accuracy under complex degradation conditions has become a key problem that urgently needs to be solved in current research on domain adaptation for underwater image processing. Summary of the Invention
[0005] This invention, under multiple constraints such as limited underwater image data, poor image quality, and scarce labeled samples, can make full use of existing high-definition land image data for transfer, effectively adapting to underwater environment recognition methods, reducing dependence on underwater data, fully mining the discriminative features of the target domain, and improving the model's generalization ability and adaptability.
[0006] To achieve the above objectives, this invention provides an underwater image recognition method for data scarcity and domain offset, comprising the following steps:
[0007] S1. Acquire labeled source domain images and unlabeled target domain underwater images;
[0008] S2. Based on the source domain image and the target domain underwater image, extract the source domain features and the target domain features, and calculate the feature distribution difference between the two features;
[0009] S3. Based on the difference in feature distribution, uncertainty modeling is performed on the source domain features, and personalized feature learning is performed on the target domain features to obtain the fused feature representation;
[0010] S4. Based on the fused feature representation, jointly train the shared feature extractor, the target domain feature extractor, and the classifier until the model converges to obtain the underwater image recognition model.
[0011] S5. Use an underwater image recognition model to identify underwater images of the target domain and output the category prediction results.
[0012] Preferably, S2 includes:
[0013] The source domain image and the target domain underwater image are input into a shared feature extractor to obtain source domain features and target domain features, respectively.
[0014] Based on the maximum mean difference method, the distribution difference between source domain features and target domain features in the reproducing kernel Hilbert space is calculated as the feature distribution difference.
[0015] Preferably, S3 includes:
[0016] The difference in feature distribution is mapped to the intensity parameter of Gaussian noise;
[0017] Noise is sampled from a Gaussian distribution with intensity parameters as the standard deviation, and this noise is added to the source domain features to obtain the perturbation-induced uncertainty feature representation.
[0018] Preferred methods for personalized feature learning of target domain features include:
[0019] Extract target domain features using a target domain feature extractor;
[0020] Gradient inversion is applied to the target domain features to generate adversarial feature representations, and the distribution difference loss between the adversarial feature representations and the source domain features is calculated.
[0021] Normalize the source domain features and target domain features, and calculate the orthogonal constraint loss between them;
[0022] The target domain features and the uncertainty feature representations are weighted and fused to obtain the personalized enhanced feature representation.
[0023] Preferably, S4 includes:
[0024] The personalized enhanced feature representation is input into the classifier, and the classification loss is calculated.
[0025] A multi-objective optimization function is constructed based on classification loss, feature distribution difference loss, orthogonality constraint loss, and distribution difference loss.
[0026] By iteratively updating the parameters of the shared feature extractor, the target domain feature extractor, and the classifier, the multi-objective optimization function is minimized until the model converges.
[0027] Preferably, S5 includes:
[0028] The underwater image of the target domain is input into the trained underwater image recognition model, features are extracted by the feature extractor, and scores for each category are output by the classifier.
[0029] Softmax normalization is applied to the scores of each category to obtain the probability distribution;
[0030] The category with the highest probability is selected as the final recognition result, and the recognition accuracy is calculated.
[0031] The present invention also provides an underwater image recognition device for data scarcity and domain offset, the device being used to implement the above method, comprising: an acquisition module, a calculation module, a fusion module, a training module, and an output module;
[0032] The acquisition module is used to acquire labeled source domain images and unlabeled target domain underwater images;
[0033] The calculation module is used to extract source domain features and target domain features based on source domain images and target domain underwater images, and to calculate the feature distribution difference between the two features;
[0034] The fusion module is used to perform uncertainty modeling on source domain features based on feature distribution differences, and to perform personalized feature learning on target domain features to obtain the fused feature representation.
[0035] The training module is used to jointly train a shared feature extractor, a target domain feature extractor, and a classifier based on the fused feature representation until the model converges, thus obtaining an underwater image recognition model.
[0036] The output module is used to identify underwater images of the target domain using an underwater image recognition model and output category prediction results.
[0037] The present invention also provides an underwater image recognition system for data scarcity and significant domain offset, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program, when executed by the processor, performs the steps of the above-described underwater image recognition method for data scarcity and significant domain offset.
[0038] The present invention also provides a storage medium storing a computer program, which, when running, executes the steps of the above-described underwater image recognition method for data scarcity and significant domain offset.
[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0040] This invention addresses image recognition tasks in complex waters under multiple constraints, including limited data volume, poor image quality, and scarce labeled samples. It designs a joint feature complementarity mechanism combining passive adaptation and active enhancement. A feature-adaptive uncertainty modeling strategy based on feature distribution differences is introduced, quantifying inter-domain differences as Gaussian noise injected into the feature space during training. This explicitly mitigates the prediction uncertainty caused by significant residual domain distribution differences. Furthermore, a personalized feature learning mechanism is proposed, enhancing the expressive power of target domain-specific features through gradient inversion and orthogonal constraints, actively uncovering the loss of local discriminative features due to image degradation. This invention effectively solves the recognition task of cross-domain migration in underwater images, providing a new approach to addressing the problems of limited target domain data and cross-domain distribution differences. Attached Figure Description
[0041] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;
[0043] Figure 2 This is a dataset constructed according to actual needs in an embodiment of the present invention;
[0044] Figure 3 This is a flowchart illustrating the algorithm for the combined uncertainty modeling and personalized feature learning strategy in an embodiment of the present invention.
[0045] Figure 4 This is a schematic diagram illustrating the underwater image category prediction results and comparative visualization results according to an embodiment of the present invention; Figure 4In the table, (a) represents the prediction result of the method proposed in this invention, (b) represents the prediction result of the source-only method, (c) represents the prediction result of the DAN method, (d) represents the prediction result of the DANN method, (e) represents the prediction result of the JAN method, (f) represents the prediction result of the CAN method, (g) represents the prediction result of the MCD method, (h) represents the prediction result of the MUDA method, (i) represents the prediction result of the FGDA method, and (j) represents the prediction result of the CC-loss method. Detailed Implementation
[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.
[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0048] Example 1:
[0049] S1 acquires labeled source domain images and unlabeled target domain underwater images.
[0050] Against the backdrop of underwater image recognition, this approach aims to fully utilize existing land-based image data for accurate underwater image identification. For example... Figure 2 As shown, labeled land images serve as the source domain, and underwater environment images of the corresponding objects serve as the target domain. Given that underwater images are affected by factors such as light attenuation, scattering interference, and imaging degradation, directly training a deep model can easily lead to unstable features and convergence difficulties. Therefore, this step first involves supervised pre-training of the shared recognition model on the labeled source domain (land scene) data to obtain source domain model parameters with robust feature representation capabilities. Subsequently, these parameters are used to initialize the weights of the target domain feature extractor and the shared classifier, achieving initial alignment between the source and target domains in the feature space.
[0051] Suppose the source domain images with labels and the target domain images without labels are respectively... and ,in, and This represents data from the source and target domains, including information such as cars, ships, containers, and people. Indicates the label corresponding to the image. These represent automobiles, ships, containers, people, airplanes, watches, motorcycles, school bags, etc. In this embodiment, a ResNet50 network is used to design the source domain feature extractor F and the target domain feature extractor, respectively. A classifier C sharing the source domain can be directly used for feature classification in the target domain. To convert the class scores into a probability distribution, a softmax function is typically used in the last layer of the classifier to transform the scores into a probability distribution. The formula is:
[0052]
[0053] In the formula, z i K is the score of the i-th category, and K is the total number of categories. The score is then indexed (to amplify the difference).
[0054] S2. Based on the source domain image and the target domain underwater image, extract the source domain features and the target domain features, and calculate the feature distribution difference between the two features.
[0055] S201. Input a large number of labeled high-resolution land images as the source domain and data-scarce, unlabeled underwater images as the target domain into a shared feature extractor to obtain source domain features and target domain features respectively, and calculate the feature distribution difference. ;
[0056] Specifically, the source and target domain images are input into a shared feature extractor to extract deep semantic features from the source and target domain images, resulting in the following feature formulas for the source and target domains:
[0057]
[0058] .
[0059] To measure the feature differences between two domains, this invention employs a maximum mean difference (MMD)-based approach to calculate the distribution of features in the source and target domains in a high-dimensional space, thereby quantifying the degree of inter-domain offset. This difference metric can guide subsequent uncertainty modeling and parameter optimization, enabling the model to gradually reduce the feature distribution differences between the source and target domains under unsupervised conditions, achieving initial cross-domain alignment. This process maintains the discriminative power of source domain features while improving the transferability of target domain features, laying the foundation for subsequent uncertainty modeling and personalized feature learning stages. (Feature distribution difference) The formula is:
[0060]
[0061] in, and Represents the features of the i-th sample in the source and target domains. and Represents the number of samples in the source and target domains. Represents the Regenerative Core Hilbert Space (RKHS). This represents the kernel function (Gaussian kernel) that maps features to RKHS.
[0062] S202. Quantize the feature distribution difference into Gaussian noise z, add it to the source domain features, and obtain the source domain features under the uncertainty modeling strategy U(·). ;
[0063] Specifically, due to the influence of data acquisition environment, image degradation, and fluctuations in noise distribution, certain residual domain differences will still exist. These differences typically manifest as uncertainties in features and prediction results. Explicitly modeling this uncertainty and transforming it into an optimization signal will help improve the model's adaptability and transfer performance in the target domain. In this embodiment, by mapping the difference in feature distribution between the source and target domains to Gaussian noise intensity parameters, the source domain features are perturbed in the feature space to construct an uncertainty modeling strategy U(·). Specifically, from... and The sampling noise term z in a Gaussian distribution with a standard deviation as the benchmark is defined as follows:
[0064]
[0065] Here, θ is a scaling factor used to adjust the sensitivity of noise intensity to domain differences. Applying this noise to the source domain features yields the perturbed adaptive feature strategy:
[0066] .
[0067] This strategy preserves the original semantic features while characterizing the instability caused by domain differences, enabling the model to explicitly perceive and adapt to residual differences during training. In this way, the model can adaptively learn inter-domain residual features, improving the robustness and generalization performance of feature representations and providing stable priors for subsequent feature optimization and domain alignment.
[0068] S203. Input the target domain image into the target domain feature extractor to obtain the target domain features. ;
[0069] Specifically, the target domain feature extractor extracts features from unlabeled underwater images to prepare for obtaining a more personalized target domain feature representation that differs from the source domain. Its feature expression is:
[0070] .
[0071] S3. Based on the differences in feature distribution, uncertainty modeling is performed on the source domain features, and personalized feature learning is performed on the target domain features to obtain the fused feature representation.
[0072] In this embodiment, the algorithm flow for combining uncertainty modeling and personalized feature learning strategies is as follows: Figure 3 As shown.
[0073] S301. Normalize and invert the extracted personalized features, and calculate the orthogonal loss with the source domain features respectively. and distribution differences ;
[0074] Specifically, to further improve the model's ability to actively mine the unique discriminative structures of complex underwater target domains, it is necessary to optimize and align the extracted personalized features to achieve collaborative constraints and distribution consistency in the feature space. In this embodiment, gradient inversion and orthogonal constraints are used for joint optimization to actively enhance the discriminative features in the target domain that distinguish it from the source domain and reduce feature redundancy.
[0075] Specifically, the target domain features extracted in S203 Building upon this foundation, to enhance the difference between this feature and the source domain features, a gradient inversion layer (GRL) is introduced to generate adversarial feature representations. And use MK-MMD to measure its relationship with the source domain features. Distribution differences:
[0076] .
[0077] The core purpose of this step is to utilize Make the target domain feature extractor It can proactively learn personalized feature representations that are significantly different from the feature distribution of the source domain. To reduce feature redundancy and enhance the independence of the feature from the source domain, this embodiment introduces orthogonal constraints.
[0078] First, the source domain features Features of the target domain Perform L2 normalization:
[0079] .
[0080] Then, an orthogonal loss function is constructed to minimize the inner product between the two types of features, thus ensuring that they are orthogonal to each other:
[0081]
[0082] Where B is the batch size. This constraint ensures that the target domain features are not only distinct from the source domain features but also structurally independent, thus providing a more complementary feature representation.
[0083] S302, S203 Adding it to S202 results in a distribution with uncertainty. The personalized features obtained under the personalized feature learning strategy P(·) are obtained. ;
[0084] Specifically, the target domain features obtained in S203 Combined with the uncertainty modeling results of S202, a weighted fusion is performed in the feature space to form a target domain feature representation with distribution uncertainty. Through a personalized feature learning strategy P(·), the model can adaptively enhance the differentiated features of the target domain while maintaining global consistency, making it more consistent with and adaptable to the true distribution characteristics of underwater images. The specific expression is:
[0085]
[0086] Where r is a balancing weighting factor, used to regulate the contribution of uncertainty features and personalized features in the final representation.
[0087] S4. Based on the fused feature representation, jointly train the shared feature extractor, the target domain feature extractor, and the classifier until the model converges, thus obtaining the underwater image recognition model.
[0088] S401, the result obtained in S302 The input is fed into a shared classifier, and the cross-entropy loss between the label and the classification result is calculated. ;
[0089] Specifically, the personalized features obtained from S302 The input is fed into the classifier, where the cross-entropy loss is calculated using the true labels from the source domain and the model's predictions to optimize the classification boundary and improve the model's discriminative performance. (Classifier Loss) It can be represented as:
[0090]
[0091] in, Represents the adaptive feature of the i-th sample in the source domain. Represents the features and label of the i-th sample in the source domain. This represents the classification loss function.
[0092] S402, based on S301 and Jointly train the target domain feature extractor;
[0093] Specifically, the orthogonal loss and distribution difference term calculated in S301 are used together as optimization objectives to jointly train the target domain feature extractor. This process, by simultaneously constraining inter-domain differences and target domain feature independence within the feature space, enables the model to enhance the representational ability of the target domain features while maintaining distribution consistency with the source domain and reducing redundancy. The specific expression is:
[0094]
[0095] S403, based on S2 and S401 Jointly train and share feature extractors and classifiers;
[0096] Specifically, the distribution difference loss calculated in S2 and the cross-entropy loss of the adaptive features in step S401 are used together as optimization objectives to jointly train the shared feature extractor and classifier. This process, by simultaneously constraining feature distribution consistency and classification accuracy, enables the model to achieve synergistic optimization of cross-domain alignment and class distinction within the global feature space, and effectively mitigates the uncertainty caused by residual domain differences. The specific expression is:
[0097]
[0098] S5. Use an underwater image recognition model to identify underwater images of the target domain and output the category prediction results.
[0099] S501. Repeat S2-S4 above to update the shared model and target domain feature extractor until the underwater image recognition model is trained.
[0100] Specifically, during repeated iterations from S2 to S4, the parameters of the shared feature extractor, target domain feature extractor, and classifier are continuously updated through iterative training. This allows the model to gradually reduce inter-domain distribution differences and enhance feature discrimination capabilities in each optimization round. As training progresses, the feature spaces of the source and target domains tend to converge, and the model's representation ability and recognition stability for underwater images continuously improve, ultimately resulting in a converged underwater image recognition model.
[0101] S502. Use the trained underwater image recognition model to recognize the target domain image.
[0102] Specifically, the feature extractor transforms the input data into high-dimensional features, and the classifier outputs a class score based on these features. The visualization results of underwater image class prediction are shown below. Figure 4 As shown in Table 1, the comparison with other methods is as follows.
[0103] To convert category scores into a probability distribution, a softmax function is typically used in the last layer of the classifier to transform the scores into a probability distribution, selecting the category with the highest probability as the predicted label.
[0104] ;
[0105] Where i is the sample number.
[0106] accuracy The calculation formula is:
[0107] ;
[0108] in, It is the number of samples that the model correctly predicted. This is the total number of samples in the test dataset.
[0109] Table 1
[0110] .
[0111] The overall process of this embodiment is as follows: Figure 1 As shown.
[0112] Example 2:
[0113] This embodiment also provides an underwater image recognition device for data scarcity and domain offset, including: an acquisition module, a calculation module, a fusion module, a training module, and an output module; the acquisition module is used to acquire labeled source domain images and unlabeled target domain underwater images; the calculation module is used to extract source domain features and target domain features based on the source domain images and target domain underwater images, and calculate the feature distribution difference between the two features; the fusion module is used to perform uncertainty modeling on the source domain features based on the feature distribution difference, and to perform personalized feature learning on the target domain features to obtain a fused feature representation; the training module is used to jointly train a shared feature extractor, a target domain feature extractor, and a classifier based on the fused feature representation until the model converges to obtain an underwater image recognition model; the output module is used to use the underwater image recognition model to recognize the target domain underwater image and output the category prediction result.
[0114] Example 3:
[0115] This embodiment also provides an underwater image recognition system for data scarcity and significant domain offset, including: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes the steps of a cross-domain underwater image recognition method when run by the processor.
[0116] Example 4:
[0117] A storage medium storing a computer program that, when running, executes steps of an underwater image recognition method oriented towards data scarcity and significant domain offset.
[0118] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for underwater image recognition addressing data scarcity and domain offset, characterized by the following steps: include: S1. Acquire labeled source domain images and unlabeled target domain underwater images; S2. Based on the source domain image and the target domain underwater image, extract the source domain features and the target domain features, and calculate the feature distribution difference between the two features; S3. Based on the difference in feature distribution, uncertainty modeling is performed on the source domain features, and personalized feature learning is performed on the target domain features to obtain the fused feature representation; S4. Based on the fused feature representation, jointly train the shared feature extractor, the target domain feature extractor, and the classifier until the model converges to obtain the underwater image recognition model. S5. Use an underwater image recognition model to identify underwater images of the target domain and output the category prediction results.
2. The underwater image recognition method for data scarcity and domain offset according to claim 1, characterized in that, S2 include: The source domain image and the target domain underwater image are input into a shared feature extractor to obtain source domain features and target domain features, respectively. Based on the maximum mean difference method, the distribution difference between source domain features and target domain features in the reproducing kernel Hilbert space is calculated as the feature distribution difference.
3. The underwater image recognition method for data scarcity and domain offset according to claim 2, characterized in that, S3 includes: The difference in feature distribution is mapped to the intensity parameter of Gaussian noise; Noise is sampled from a Gaussian distribution with intensity parameters as the standard deviation, and this noise is added to the source domain features to obtain the perturbation-induced uncertainty feature representation.
4. The underwater image recognition method for data scarcity and domain offset according to claim 3, characterized in that, Methods for personalized feature learning of target domain features include: Extract target domain features using a target domain feature extractor; Gradient inversion is applied to the target domain features to generate adversarial feature representations, and the distribution difference loss between the adversarial feature representations and the source domain features is calculated. Normalize the source domain features and target domain features, and calculate the orthogonal constraint loss between them; The target domain features and the uncertainty feature representations are weighted and fused to obtain the personalized enhanced feature representation.
5. The underwater image recognition method for data scarcity and domain offset according to claim 4, characterized in that, S4 includes: The personalized enhanced feature representation is input into the classifier, and the classification loss is calculated. A multi-objective optimization function is constructed based on classification loss, feature distribution difference loss, orthogonality constraint loss, and distribution difference loss. By iteratively updating the parameters of the shared feature extractor, the target domain feature extractor, and the classifier, the multi-objective optimization function is minimized until the model converges.
6. The underwater image recognition method for data scarcity and domain offset according to claim 1, characterized in that, S5 include: The underwater image of the target domain is input into the trained underwater image recognition model, features are extracted by the feature extractor, and scores for each category are output by the classifier. Softmax normalization is applied to the scores of each category to obtain the probability distribution; The category with the highest probability is selected as the final recognition result, and the recognition accuracy is calculated.
7. An underwater image recognition device for data scarcity and domain offset, the device being used to implement the method according to any one of claims 1-6, characterized in that, include: The module comprises an acquisition module, a calculation module, a fusion module, a training module, and an output module. The acquisition module is used to acquire labeled source domain images and unlabeled target domain underwater images; The calculation module is used to extract source domain features and target domain features based on source domain images and target domain underwater images, and to calculate the feature distribution difference between the two features; The fusion module is used to perform uncertainty modeling on source domain features based on feature distribution differences, and to perform personalized feature learning on target domain features to obtain the fused feature representation. The training module is used to jointly train a shared feature extractor, a target domain feature extractor, and a classifier based on the fused feature representation until the model converges, thus obtaining an underwater image recognition model. The output module is used to identify underwater images of the target domain using an underwater image recognition model and output category prediction results.
8. An underwater image recognition system for situations with scarce data and significant domain offset, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program, when executed by the processor, performs the steps of underwater image recognition oriented towards data scarcity and domain offset as described in any one of claims 1-6.
9. A storage medium, characterized in that, The storage medium stores a computer program that, when running, executes the steps of the underwater image recognition method for data scarcity and significant domain offset as described in any one of claims 1-6.