An underwater image segmentation method, device, medium and product based on a frozen visual base model and a multi-scale convolution prior

Abstract

This application discloses an underwater image segmentation method, device, medium, and product based on a frozen visual foundation model and multi-scale convolutional priors, relating to the field of image segmentation. The method includes: acquiring an underwater image to be processed and preprocessing it to obtain an input image tensor; extracting multi-scale features from the input image tensor using a multi-scale convolutional query generation module to obtain multi-scale convolutional query features; inputting the input image tensor into a frozen pre-trained visual foundation model to extract multi-layer global semantic features; using the multi-scale convolutional query features as query features and the multi-layer global semantic features as value features, performing progressive cross-layer semantic fusion through a multi-layer adaptation fusion module to obtain fused features; decoding the fused features to generate a pixel-level segmentation mask for underwater targets, thus completing underwater image segmentation. This application can achieve accurate segmentation of underwater target regions and boundary contour restoration, providing visual technology support for various underwater engineering applications.

Description

Technical Field

[0001] This application relates to the field of image segmentation, and in particular to an underwater image segmentation method, device, medium, and product based on a frozen visual foundation model and multi-scale convolutional priors. Background Technology

[0002] Underwater image processing plays a crucial role in underwater robot navigation, marine life monitoring, target measurement, and marine resource surveys. However, the complex and unique underwater imaging environment can cause a series of problems, such as significant light attenuation, severe color shifts, scattering interference from suspended particles in the water, and turbidity. These issues further result in low overall image contrast, making it easy for target objects to blend into the background and form visual camouflage. Ultimately, this leads to blurred target boundaries, severe degradation of local details and textures, and makes it difficult to reliably achieve accurate image region segmentation.

[0003] While traditional CNNs (Convolutional Neural Networks) have the ability to extract local features, they are insufficient in long-distance dependencies and global semantic modeling, which can easily lead to fragmented regions or semantic inconsistencies.

[0004] Transformer or visual base models have strong global semantic representation capabilities, but their token representations usually have coarse spatial structures and are insufficient for recovering weak underwater boundaries and fine-grained contours. Direct full-parameter fine-tuning of large-scale visual base models is costly and may destroy the original general semantic capabilities of pre-trained models.

[0005] Therefore, there is a need for an underwater image segmentation scheme that can simultaneously integrate local convolutional spatial priors and frozen base model global semantics, while taking into account segmentation accuracy, boundary localization, and real-time performance. Summary of the Invention

[0006] The purpose of this application is to provide an underwater image segmentation method, device, medium, and product based on a frozen visual foundation model and multi-scale convolutional priors, which can achieve accurate underwater image segmentation and detail restoration, and provide visual technology support for various underwater engineering applications.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an underwater image segmentation method based on a frozen visual foundation model and multi-scale convolutional priors, including: Acquire the underwater image to be processed and preprocess it to obtain the input image tensor; The multi-scale convolution query generation module extracts multi-scale features from the input image tensor to obtain multi-scale convolution query features; the multi-scale convolution query generation module includes a Stem module and a multi-channel attention convolution projection module. The input image tensor is fed into a frozen pre-trained visual base model to extract multi-layer global semantic features; wherein the parameters of the pre-trained visual base model are kept frozen during training. Using the multi-scale convolutional query features as query features and the multi-layer global semantic features as value features, progressive cross-layer semantic fusion is performed through a multi-layer adaptation fusion module to obtain fused features. The fused features are improved by enhancing the receptive field and performing grouping aggregation bridging decoding to generate a pixel-level segmentation mask for underwater targets, thus completing underwater image segmentation.

[0008] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described underwater image segmentation method based on a frozen visual foundation model and multi-scale convolutional priors.

[0009] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described underwater image segmentation method based on a frozen visual foundation model and multi-scale convolutional priors.

[0010] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described underwater image segmentation method based on a frozen visual foundation model and multi-scale convolutional priors.

[0011] According to the specific embodiments provided in this application, this application has the following technical effects: This application employs a pre-trained visual base model with frozen parameters to extract global semantic features, effectively compensating for the shortcomings of traditional convolutional neural networks in long-distance dependency modeling, fragmented segmentation regions, and semantic inconsistencies. It also eliminates the need for large-scale model full-parameter fine-tuning, reducing training costs while retaining the model's original general semantic capabilities. A multi-scale convolutional query generation module extracts multi-scale convolutional query features, fully leveraging the spatial representation advantages of convolutional networks to accurately restore weak boundary contours and fine-grained details in underwater images, improving image detail degradation and target contour blurring caused by the underwater imaging environment. A multi-layer adaptation fusion module progressively fuses local convolutional priors and global semantic features, achieving complementary advantages between the two types of features. This effectively resists underwater scene interference such as light attenuation, water scattering, and water turbidity, overcoming segmentation challenges caused by target visual camouflage. The final output is pixel-level segmentation results, improving underwater target segmentation accuracy and boundary localization. The overall solution balances segmentation accuracy, contour restoration capability, and real-time performance, making it well-suited for practical applications such as underwater robot navigation, marine life monitoring, and marine resource surveys. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating an underwater image segmentation method based on a frozen visual foundation model and multi-scale convolutional prior, provided as an embodiment of this application; Figure 2 This is a detailed flowchart illustrating an underwater image segmentation method based on a frozen vision model and multi-scale convolutional prior, provided as an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] In one exemplary embodiment, such as Figures 1-2 As shown, an underwater image segmentation method based on a frozen visual baseline model and multi-scale convolutional prior is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is described using a server as an example, and includes the following steps S1 to S5.

[0017] S1: Obtain the underwater image to be processed and preprocess it to obtain the input image tensor.

[0018] Underwater images can originate from underwater cameras, underwater robot vision systems, marine monitoring equipment, or pre-stored underwater image datasets.

[0019] Subsequently, the underwater image is preprocessed: the image is adjusted to a preset input size (e.g., 352×352 pixels, which can be adjusted according to actual computing power) and normalized (e.g., subtracting the mean and dividing by the standard deviation) to obtain the input image tensor.

[0020] S2: The input image tensor is subjected to multi-scale feature extraction through the multi-scale convolution query generation module to obtain multi-scale convolution query features; the multi-scale convolution query generation module includes a Stem module and a multi-channel attention convolution projection module.

[0021] This step is implemented through a multi-scale convolutional query generation module, aiming to extract spatially corresponding local features from the input image tensor and transform them into a query format suitable for interaction with the Transformer. Specifically, it includes the following sub-steps: S21: The input image tensor is subjected to preliminary convolution feature extraction and resolution alignment by the Stem module to obtain the first-scale convolution features.

[0022] The Stem module is used to suppress early noise interference in underwater images, preserve edge, texture and local structural information, and provide basic spatial features for subsequent multi-scale convolutional query generation.

[0023] u is the input image tensor; s is the convolution stride; Conv3×3 represents a convolution operation with a kernel size of 3×3, and its weights are learnable parameters; BN represents batch normalization, and its scaling and translation parameters are learnable parameters; ReLU represents the linear rectified activation function; CBR(u;s) is the first-scale convolution feature, which is used for local structure extraction and resolution adjustment.

[0024] S22: By using a multi-channel attention convolution projection module, the first-scale convolution features are downsampled and recalibrated at each level to obtain convolution features at multiple scales.

[0025] The Channel-Attentive Convolutional Projection (CA-CP) module includes strided convolution, global average pooling, global max pooling, sigmoid channel weights, and residual channel modulation, which are used to enhance channel discriminability while preserving spatial correspondence.

[0026] S23: Flatten, hierarchically embed, and concatenate the convolutional features at multiple scales to obtain multi-scale convolutional query features.

[0027] Convolutional features at some or all scales are selected, and learnable hierarchical embeddings are added to each convolutional feature to preserve the hierarchical identifiers and scale-aware information of features at different scales. Then, the spatial dimensions of these features are flattened and concatenated according to the spatial dimensions to form multi-scale convolutional query features. This convolutional query sequence retains the local edges, texture structures, and spatial correspondences generated by the multi-scale convolutional query generation module, and can be used as query features in subsequent semantic interaction processes.

[0028] This represents the convolutional feature at the k-th scale; For learnable hierarchical embedding; Flatten indicates spatial flattening operation; This indicates that features at different scales are concatenated along the sequence dimension; This is a multi-scale convolution query feature.

[0029] S3: Input the input image tensor into the frozen pre-trained visual base model to extract multi-layer global semantic features; wherein, the parameters of the pre-trained visual base model are kept frozen during the training process.

[0030] This step utilizes the general visual semantic knowledge learned from large-scale image data by the frozen pre-trained visual base model to provide global contextual constraints for underwater target area localization. At the same time, since the parameters of the pre-trained visual base model are kept frozen, the increased training cost and catastrophic forgetting risk caused by full parameter fine-tuning can be avoided, and a stable source of semantic features can be provided for the subsequent multi-layer adaptation and fusion module.

[0031] In one implementation, the pre-trained visual base model can employ the DINOv3 backbone network, during which all parameters of the model are frozen and do not participate in gradient updates.

[0032] The feature extraction process is as follows: S31: Divide the input image tensor into multiple image blocks, and map each image block to a visual token through linear projection to obtain an initial token sequence.

[0033] The input image tensor is divided into N non-overlapping image patches, and each image patch is mapped to a fixed-dimensional visual token through a linear projection layer to obtain the initial token sequence.

[0034] S32: Add location encoding to the initial token sequence to obtain an input token sequence containing spatial location information.

[0035] Learnable location codes are added to the initial token sequence so that the token sequence retains spatial location information between different image regions.

[0036] S33: Input the input token sequence into the multi-layer Transformer encoder of the frozen pre-trained visual base model, and extract features layer by layer through self-attention mechanism and feedforward neural network to obtain the output features of each layer of Transformer encoder.

[0037] The token sequence with location encoding is input into the multi-layer Transformer encoder of the DINOv3 backbone network. The long-distance dependency between different regions in the underwater image is modeled through the self-attention mechanism, and the global semantic expression capability is enhanced through the feedforward network.

[0038] S34: Select the output features of the preset level from the output features of each Transformer encoder layer as multi-layer semantic features.

[0039] The outputs of specific intermediate layers in the Transformer encoder are selected as multi-layer semantic features. In this embodiment, the outputs of layers 9, 19, 29, and 39 are selected as multi-layer semantic features. These layers represent visual expressions at different depths, from shallow texture to deep semantics. Shallower features contain relatively rich local texture, edge, and structural information, while deeper features contain stronger target region semantics, contextual relationships, and global perception information.

[0040] S35: Perform dimension alignment processing on the multi-layer semantic features to obtain the multi-layer global semantic features.

[0041] The aforementioned multi-layer semantic features are mapped or rearranged into corresponding two-dimensional semantic feature maps according to their spatial position relationships. The number of channels and spatial resolution are aligned through linear projection, convolutional projection, normalization, or interpolation operations to obtain multi-layer global semantic features that can be used for subsequent fusion.

[0042] S4: Using the multi-scale convolutional query features as query features and the multi-layer global semantic features as value features, progressive cross-layer semantic fusion is performed through the multi-layer adaptation fusion module to obtain fused features.

[0043] This step utilizes a multi-layer adaptation and fusion module to achieve interaction between local convolutional priors and global semantics. Instead of simply concatenating features, this step employs a progressive interaction to guide local convolutional spatial features with global semantics, while avoiding coarse semantics directly overriding weak boundary details. In one implementation, the multi-layer adaptation and fusion module uses a multi-scale deformable attention mechanism, selecting a limited number of sampling points around each query location for semantic aggregation. This reduces the computational complexity of global attention and enhances the matching between local regions and global semantics.

[0044] Specifically, it includes the following sub-steps: S41: Divide the multi-scale convolutional query features into multiple query tokens, each query token containing a feature vector and a normalized two-dimensional reference point.

[0045] S42: For each query token, at each layer of the multi-layer global semantic features, predict the sampling point offset and attention weight by linear projection.

[0046] S43: Based on the normalized two-dimensional reference point and the sampling point offset, determine the spatial position of each sampling point on the corresponding layer of the multi-layer global semantic feature.

[0047] S44: Use bilinear interpolation to extract feature values ​​at each sampling point from the corresponding layer of the multi-layer global semantic features, and then perform a weighted summation of the extracted feature values ​​and the corresponding attention weights to obtain aggregated semantic features.

[0048] S45: Perform a residual concatenation between the aggregated semantic features and the feature vector of the query token to obtain the updated query token.

[0049] S46: Normalize the updated query token and process it with a convolutional feedforward network to obtain the enhanced query sequence.

[0050] S47: The enhanced query sequence is split and rearranged into a two-dimensional feature map according to the original scale to obtain the enhanced convolutional semantic fusion feature.

[0051] S48: The enhanced convolutional semantic fusion feature is fused with the multi-layer global semantic feature to obtain the fused feature.

[0052] S5: Improve the receptive field enhancement and group aggregation bridging decoding processing of the fused features to generate a pixel-level segmentation mask for underwater targets, and complete the underwater image segmentation.

[0053] This step aims to restore spatial resolution and generate the final mask. Specifically, it includes the following sub-steps: S51: Input the fused features into the improved receptive field module, and use multi-branch asymmetric convolution and dilated convolution with different dilation rates to enhance the context features, thereby obtaining context-enhanced features. Use the context-enhanced features as high-level semantic features in the decoding stage.

[0054] The improved receptive field module includes a 1×1 convolution in the first branch, combinations of 1×k and k×1 asymmetric convolutions in the second to fourth branches, and 3×3 dilated convolutions with different dilation rates. The outputs of each branch are concatenated along the channel dimension, fused by 3×3 convolutions, and then output as context-enhanced features via 1×1 convolutional residual connections and ReLU activation. This step expands the receptive field with lower parameter overhead, enabling the model to utilize contextual information surrounding the underwater target and mitigate target-background confusion caused by low contrast, turbidity, and background camouflage.

[0055] This is the output of the first branch; For the output of other branches, n=2,3,4. This represents a 3×3 dilated convolution operation with an inflation rate of d. [·] represents context-enhanced features; [·] represents channel concatenation; δ represents the ReLU activation function.

[0056] S52: Obtain the shallow multi-scale convolutional features generated by the multi-scale convolutional query generation module as low-level spatial features in the decoding stage.

[0057] S53: Generate spatial guidance information representing the response of underwater target areas or boundaries.

[0058] Spatial guidance information refers to the spatial guidance mask generated from intermediate prediction results, high-level semantic responses, or auxiliary segmentation branches during the decoding stage. It is used to represent the region where underwater targets may be located, the boundary response, or the foreground attention region. This spatial guidance information serves as an auxiliary constraint, participating in the fusion with high-level semantic features and low-level spatial features to guide the network to focus on the underwater target region and suppress background interference.

[0059] S54: By using the grouping and aggregation bridging module, the high-level semantic features, the low-level spatial features, and the spatial guidance information are fused to obtain the aggregated features for the decoding stage.

[0060] Specifically, the high-level semantic features are upsampled and channel aligned to obtain aligned high-level semantic features; the aligned high-level semantic features and the low-level spatial features are grouped along the channel dimension to obtain multiple feature groups; the spatial guidance information is concatenated with each feature group to obtain concatenated feature groups; each concatenated feature group is subjected to layer normalization and depthwise separable convolution with different dilation rates to obtain context information at different scales; the context information at different scales is concatenated and the number of channels is restored through 1×1 convolution to obtain aggregated features.

[0061] High-level semantic features; The i-th feature group represents the low-level spatial features; M represents spatial guidance information. These are the aligned high-level semantic features; The aligned high-level semantic features of the i-th feature group; LN represents bilinear interpolation; LN represents layer normalization. The expansion rate is d i 3×3 depthwise separable convolution operation; This represents the contextual information of the i-th feature group; This is an aggregation feature.

[0062] S55: The aggregated features are processed by the decoding output module to generate a pixel-level segmentation mask for the underwater target.

[0063] The decoding output module generates a pixel-level segmentation mask for underwater targets. This segmentation mask can be used for underwater target boundary localization, target area measurement, underwater robot perception, marine life monitoring, or other underwater vision tasks.

[0064] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments. The computer device can be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, memory, and I / O interface are connected via a system bus, and the communication interface is connected to the system bus via the I / O interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores data to be processed. The I / O interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with an external terminal via a network connection. When the computer program is executed by the processor, it implements the steps in the above-described method embodiments.

[0065] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0066] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0067] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0068] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0069] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An underwater image segmentation method based on a frozen vision model and multi-scale convolutional priors, characterized in that, include: Acquire the underwater image to be processed and preprocess it to obtain the input image tensor; The multi-scale convolution query generation module extracts multi-scale features from the input image tensor to obtain multi-scale convolution query features; the multi-scale convolution query generation module includes a Stem module and a multi-channel attention convolution projection module. The input image tensor is fed into a frozen pre-trained visual base model to extract multi-layer global semantic features; wherein the parameters of the pre-trained visual base model are kept frozen during training. Using the multi-scale convolutional query features as query features and the multi-layer global semantic features as value features, progressive cross-layer semantic fusion is performed through a multi-layer adaptation fusion module to obtain fused features. The fused features are improved by enhancing the receptive field and performing grouping, aggregation, bridging, and decoding to generate a pixel-level segmentation mask for underwater targets, thus completing underwater image segmentation.

2. The underwater image segmentation method based on a frozen vision model and multi-scale convolutional prior as described in claim 1, characterized in that, The multi-scale convolutional query generation module extracts multi-scale features from the input image tensor to obtain multi-scale convolutional query features, specifically including: The input image tensor is preliminarily convolutional feature extraction and resolution alignment are performed by the Stem module to obtain the first-scale convolutional features. By using a multi-channel attention convolutional projection module, the first-scale convolutional features are downsampled and recalibrated at each level to obtain convolutional features at multiple scales. Multi-scale convolutional features are flattened, hierarchically embedded, and concatenated to obtain multi-scale convolutional query features.

3. The underwater image segmentation method based on a frozen vision model and multi-scale convolutional prior as described in claim 1, characterized in that, The pre-trained visual base model uses the DINOv3 backbone network.

4. The underwater image segmentation method based on a frozen visual foundation model and multi-scale convolutional prior as described in claim 1, characterized in that, The input image tensor is fed into a frozen pre-trained visual base model to extract multi-layer global semantic features, specifically including: The input image tensor is divided into multiple image blocks, and each image block is mapped to a visual token through linear projection to obtain an initial token sequence; Add location encoding to the initial token sequence to obtain an input token sequence containing spatial location information; The input token sequence is fed into the multi-layer Transformer encoder of the frozen pre-trained visual base model. Layer-by-layer feature extraction is performed through self-attention mechanism and feedforward neural network to obtain the output features of each layer of Transformer encoder. From the output features of each Transformer encoder layer, select the output features of the preset layer as multi-layer semantic features; The multi-layer semantic features are dimensionally aligned to obtain the multi-layer global semantic features.

5. The underwater image segmentation method based on a frozen vision model and multi-scale convolutional prior as described in claim 1, characterized in that, Using the multi-scale convolutional query features as query features and the multi-layer global semantic features as value features, progressive cross-layer semantic fusion is performed through a multi-layer adaptation fusion module to obtain fused features, specifically including: The multi-scale convolutional query features are divided into multiple query tokens, each of which contains a feature vector and a normalized two-dimensional reference point. For each query token, at each layer of the multi-layer global semantic features, the sampling point offset and attention weight are predicted by linear projection; Based on the normalized two-dimensional reference point and the sampling point offset, the spatial position of each sampling point on the corresponding layer of the multi-layer global semantic feature is determined. The feature values ​​at each sampling point are extracted from the corresponding layer of the multi-layer global semantic features using bilinear interpolation, and the extracted feature values ​​are weighted and summed with the corresponding attention weights to obtain aggregated semantic features. The aggregated semantic features are residually concatenated with the feature vector of the query token to obtain the updated query token. The updated query token is normalized and processed by a convolutional feedforward network to obtain the enhanced query sequence; The enhanced query sequence is split and rearranged into a two-dimensional feature map according to the original scale to obtain the enhanced convolutional semantic fusion feature. The enhanced convolutional semantic fusion features are fused with multi-layer global semantic features to obtain fused features.

6. The underwater image segmentation method based on a frozen vision foundation model and multi-scale convolutional prior as described in claim 1, characterized in that, The fused features are subjected to improved receptive field enhancement and group aggregation bridging decoding processing to generate a pixel-level segmentation mask for underwater targets, specifically including: The fused features are input into the improved receptive field module, and contextual features are enhanced using multi-branch asymmetric convolution and dilated convolution with different dilation rates to obtain contextual enhanced features. The contextual enhanced features are then used as high-level semantic features in the decoding stage. The shallow multi-scale convolutional features generated by the multi-scale convolutional query generation module are obtained as low-level spatial features in the decoding stage. Generate spatial guidance information representing the response of underwater target areas or boundaries; By using the grouping aggregation bridging module, the high-level semantic features, the low-level spatial features, and the spatial guidance information are fused to obtain the aggregated features for the decoding stage; The aggregated features are processed by the decoding output module to generate a pixel-level segmentation mask for underwater targets.

7. The underwater image segmentation method based on a frozen vision foundation model and multi-scale convolutional prior as described in claim 6, characterized in that, Through a grouping and aggregation bridging module, high-level semantic features, low-level spatial features, and spatial guidance information are integrated, specifically including: The high-level semantic features are upsampled and channel aligned to obtain aligned high-level semantic features; The aligned high-level semantic features and the low-level spatial features are grouped along the channel dimension to obtain multiple feature groups; The spatial guidance information is concatenated with each feature group to obtain the concatenated feature group; Each concatenated feature group is subjected to layer normalization and depthwise separable convolution with different dilation rates to obtain contextual information at different scales. Contextual information at different scales is concatenated, and the number of channels is restored through 1×1 convolution to obtain aggregated features.

8. A computer device, comprising: The memory, the processor, and the computer program stored in the memory and executable on the processor are characterized in that the processor executes the computer program to implement the underwater image segmentation method based on the frozen visual foundation model and multi-scale convolution prior as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the underwater image segmentation method based on a frozen vision foundation model and a multi-scale convolutional prior, as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the underwater image segmentation method based on a frozen vision foundation model and a multi-scale convolutional prior, as described in any one of claims 1-7.

Images