A method and system for 3D segmentation of bubble plume targets in multibeam water column images

Abstract

This invention relates to the field of marine mapping and underwater target segmentation technology, and provides a method and system for 3D segmentation of bubble plume targets in multibeam water column images. Based on the standard 3D U-NET model, the method includes: replacing each layer encoder of the 3D U-NET model with a residual 3D encoder, employing a hierarchical structure of cascaded downsampling paths, downsampling level by level, and progressively extracting multi-scale features; embedding a 3D sparse self-attention visual transformer at the bottleneck layer output of the 3D U-NET model to integrate spatial context information and perform global feature enhancement on the deep semantic features after multiple downsampling; and embedding a dual-gated attention fusion unit at each connection point between the encoder and decoder of the 3D U-NET model, except for the bottleneck layer, to perform gated attention modeling of bidirectional features along the channel dimension. This method improves the accuracy of 3D segmentation of bubble plume targets in multibeam water column images while maintaining low overhead.

Description

Technical Field

[0001] This invention relates to the field of marine mapping and underwater target segmentation technology, specifically to a method and system for three-dimensional segmentation of bubble plume targets in multibeam water column images. Background Technology

[0002] Submarine cold seeps release large amounts of methane, some of which is oxidized to carbon dioxide during its ascent, thus affecting the ocean's carbon chemical balance. Bubble plumes are key carriers of greenhouse gases like methane from the seabed into the ocean and atmosphere. Multibeam sonar systems (MBES) can acquire information on the backscatter intensity of targets in the seabed. Due to differences in acoustic impedance, the echo intensity of bubble plumes is typically 20-30 dB higher than the background intensity of the water body, but lower than the echo intensity at the seabed surface. Based on this characteristic, the spatiotemporal state of bubble plumes can be analyzed using multibeam water column image data. However, under current technological conditions, there are significant bottlenecks in automated detection methods for methane seepage and quantitative calculation of seepage volume, seriously hindering efficient marine exploration and resource analysis.

[0003] To address the aforementioned issue of complex data processing, intelligent processing of bubble plume data from multibeam water column images can improve efficiency. Early research focused on thresholding TA and DA images using mean and standard deviation, then taking the intersection and using morphological features to filter bubble plumes. Automatic detection of bubble plumes was achieved by combining sector CFAR thresholding with morphological constraints. With the development of machine learning, subsequent research has focused on fully utilizing Haar-like features and Local Binary Patterns (LBP) features to leverage the grayscale variations and texture information of the target plume, combined with an AdaBoost classifier, to achieve automatic detection of bubble plumes. To address the low accuracy of recognizing tilted, weak bubble plumes, a method using Bag-of-Words (BOVW) and employing a quadratic SVM to nonlinearly partition the bag-of-words features was proposed to improve detection accuracy in complex backgrounds.

[0004] In recent years, due to the application of deep learning in computer vision, methods for automatically analyzing and detecting multibeam water column images using artificial intelligence have gradually emerged. The paper "Automatic segmentation of gas plumes from multibeam water column images using a U-shape network" constructs an AP-UNet network by adding CBAM and PPM modules to the traditional U-Net network, achieving automatic segmentation and extraction of bubble plumes from multibeam water column images. This method significantly improves the semantic segmentation accuracy of bubble plumes. The paper "Deep learning-based sequential processing of multibeam echosounder images for automated detection of seafloor gas seep occurrence" emphasizes the importance of sequence analysis of multibeam water column images and proposes a method using 2D convolution combined with a BiLSTM model to achieve rapid screening and preliminary detection of bubble seepage while considering spatial information. However, fully utilizing the spatial continuity across slices of water column images to ensure the geometric coherence of volumetric data is of great significance for subsequent fields such as gas flux estimation and flow velocity analysis. Therefore, compared to classification and target detection of 3D information, 3D segmentation of bubble plumes in 3D multibeam water column images is more meaningful. "3D U-Net: Learning dense volumetric segmentation from sparse annotation" comprehensively extends the mature 2D U-Net architecture to 3D space, replacing all the original 2D operations with 3D convolution, 3D max pooling, and 3D upconvolution, enabling it to directly process volumetric data. This model has been widely used in the field of medical image processing. "Research on Liver Tumor CT Image Segmentation Method Based on 3DUNet" utilizes an improved 3DU-Net model to segment liver tumor CT images with good generalization performance. "Cerebrovascular segmentation algorithm based on multi-scale 3DUnet" addresses the characteristics of brain blood vessels, such as small diameter, numerous branches, and complex vascular paths, by combining multi-scale modules on the 3D U-Net baseline to achieve accurate segmentation of small brain blood vessels in CAT images.However, the 3D U-Net model relies on CNNs, and its dependence on local feature extraction inherently limits its ability to capture the global background. This correlation needs to span the entire liver region. Even if 3D CNN models such as 3D U-Net expand the receptive field by stacking convolutional layers, they still cannot effectively cover the global range due to low efficiency of receptive field expansion and gradient vanishing problems, leading to missed small lesions or misjudgment of tumor boundaries during segmentation. To overcome this limitation, the TransUNet model used in "Transunet: Transformers make strong encoders formedical image segmentation" combines the Transformer architecture to introduce a global attention mechanism, achieving a breakthrough in segmentation accuracy. The TransUNet encoder uses a transformer model, and its computational complexity increases quadratically with the increase of sequence length. This makes the Transformer model resource-intensive and memory-intensive for high-resolution images, which greatly limits the practical application of the model. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to improve the accuracy of three-dimensional segmentation of bubble plume targets in multibeam water column images while keeping costs low.

[0006] The present invention solves the above-mentioned technical problems through the following technical means:

[0007] This invention provides a 3D segmentation method for bubble plume targets in multibeam water column images, based on the standard 3D U-NET model, including the following steps: S1. A residual 3D encoder is used to replace the encoder of each layer of the 3D U-NET model. A hierarchical structure with cascaded downsampling paths is adopted to downsample step by step and extract multi-scale features gradually. S2. Embed a 3D sparse self-attention visual transformer at the bottleneck layer output of the 3D U-NET model to integrate spatial context information and perform global feature enhancement on the deep semantic features after multiple downsampling. S3. At each connection point of the encoder and decoder of the 3D U-NET model except for the bottleneck layer, a dual-gated attention fusion unit is embedded to perform gated attention modeling of bidirectional features along the channel dimension.

[0008] Furthermore, the residual 3D encoder consists of several residual blocks, each of which is composed of a residual mapping path and an identity mapping path. The residual mapping path performs a stepwise nonlinear transformation on the input features through two layers of 3D convolution. The identity mapping path achieves cross-layer transfer and information reuse of the input features through direct connection or channel mapping. The residual mapping path and the identity mapping path are fused at the output end.

[0009] Furthermore, the construction of the three-dimensional sparse self-attention visual transformer includes the following steps: S21. Obtain the deep low-resolution 3D feature map from the bottleneck layer. Expanding into a sequence, the original three-dimensional feature representation is converted into a set of token sequences. ,in, Indicates batch size. Indicates the number of feature channels. Indicates the number of tokens in the sequence; S22, through a learnable linear mapping matrix Generate query vectors respectively Key vector Sum value vector And use the self-attention mechanism to calculate the correlation weights between spatial features; S23. Based on the correlation weight, output the enhanced 3D feature map. .

[0010] Furthermore, the correlation weights between the computational spatial features are as follows:

[0011] in, It is the dimension of each attention head.

[0012] Furthermore, the construction of the dual-gated attention fusion processor includes the following steps: S31, Encoder features and decoder features Mapped to dimensions respectively intermediate features and ; S32. Add the mapped intermediate features and obtain the fused features using the ReLU activation function. ; to integrate features The positive attention map is obtained by performing convolution operations and processing with activation functions. and reverse attention graph ; S33. Spatial modulation of intermediate features is performed using the generated attention map, and a positive attention map is used. To enhance the spatial features of the encoder Using reverse attention graphs To refine the semantic features of the decoder ; S34, will and Addition to obtain fusion features The number of channels projected back to the original encoder features To obtain the final output .

[0013] Furthermore, the intermediate features and The calculation is as follows:

[0014]

[0015] in, For batch normalization operations, It is a 1×1×1 convolution.

[0016] Furthermore, the positive attention map and reverse attention graph The calculation is as follows:

[0017]

[0018] in, , is the activation function. Forward convolution, This is an inverse convolution.

[0019] Furthermore, the features and The calculation is as follows:

[0020]

[0021] in, For Hadama accumulation.

[0022] Furthermore, the final output The calculation is as follows:

[0023]

[0024] in, It is a 1×1×1 convolution.

[0025] This invention also provides a three-dimensional segmentation system for bubble plume targets in multibeam water column images. The system executes the above-described method during operation and includes the following modules: The feature extraction module is used to replace the encoder of each layer of the 3D U-NET model with a residual 3D encoder. It adopts a hierarchical structure of cascaded downsampling paths, downsamples step by step, and gradually extracts multi-scale features. The feature enhancement module is used to embed a 3D sparse self-attention visual transformer at the bottleneck layer output of the 3D U-NET model, integrate spatial context information, and perform global feature enhancement on the deep semantic features after multiple downsampling. The bidirectional information interaction module is used to embed a dual-gated attention fusion unit at the encoder and decoder connection of the 3D U-NET model to perform gated attention modeling of bidirectional features along the channel dimension.

[0026] The advantages of this invention are: This invention improves upon the standard 3D U-NET by employing a residual 3D encoder as the backbone for enhancing feature representation during information propagation. It also extracts long-range spatially dependent plume information by embedding a 3D sparse self-attention visual transformer at the network bottleneck, endowing the model with global perception capabilities to better understand the attribution relationship between distant bubble clusters and the plume main body. Simultaneously, a dual-gated attention fusion processor is added after skip links and upsampling in the network decoder, making the model more focused on the plume body, sharpening its boundaries while suppressing background noise in shallow spatial information. Balancing performance and effectiveness, this invention achieves 3D segmentation of bubble plumes in complex noisy environments, improving the accuracy of 3D segmentation of bubble plume targets in multibeam water column images, and demonstrating the potential of artificial intelligence in marine mapping. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the architecture of the ESDR-UNet3D model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of a residual three-dimensional encoder according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture of the three-dimensional sparse self-attention visual transformer according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the architecture of the dual-gated attention fusion processor according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0029] Example 1 This embodiment provides a 3D segmentation method for bubble plume targets in multibeam water column images, based on an overall improvement of the existing standard 3D U-Net model. The standard 3D U-Net model is a fully convolutional neural network with an encoder-decoder structure, specifically designed for end-to-end segmentation tasks of 3D volumetric data. Its overall architecture has a U-shaped symmetrical structure and mainly consists of the following four parts: (1) Encoder path, used to extract multi-scale spatial features of the input 3D volume data layer by layer. It consists of multiple stacked 3D convolutional blocks, each containing two consecutive 3×3×3 3D convolutional layers. Each convolution is followed by a ReLU activation function and a 2×2×2 max pooling layer for downsampling, which halves the feature map size while doubling the number of channels. Through four downsampling operations, multi-level feature representations from low-level edge features to high-level semantic features are extracted step by step.

[0030] (2) Decoder path, used to restore the spatial resolution of the feature map and achieve pixel-level precise localization. It consists of multiple stacked 3D deconvolution blocks, each containing a 2×2×2 three-dimensional deconvolution layer for upsampling, doubling the size of the feature map while halving the number of channels. It fuses features with the corresponding layer of the encoder through skip connections. Two consecutive 3×3×3 three-dimensional convolution layers are connected to the ReLU activation function. Through four upsampling operations, the resolution of the input data is gradually restored.

[0031] (3) Skip connections are used to concatenate the feature maps of each layer in the encoder with the upsampled feature maps of the corresponding layers in the decoder along the channel dimension. This compensates for the spatial detail information lost during downsampling, realizes the fusion of deep semantic features and shallow positional features, and improves the localization accuracy of segmentation boundaries.

[0032] (4) Output layer: The last layer uses a 1×1×1 convolution to map the number of feature channels to the number of target categories. Combined with the Softmax activation function or the Sigmoid activation function, the probability map of each voxel belonging to each category is output, thus achieving segmentation.

[0033] However, the standard 3D U-NET model has shortcomings when facing the 3D segmentation task of bubble plume targets in multibeam water column images, including: difficulty in handling problems such as low signal-to-noise ratio, edge discontinuity and complex 3D spatial structure in multibeam water column images; difficulty in effectively handling the long-distance spatial dependence of bubble plume in multibeam water column images; and difficulty in handling the blurred and diffuse boundaries of bubble plume in multibeam water column images.

[0034] Therefore, this embodiment provides an image segmentation model based on the standard 3D U-NET for bubble plume flow in multibeam water column images, referred to as the ESDR-UNet3D model (Enhanced Sparse Attention and Dual-fused Residual UNet3D). The overall architecture of the model is as follows: Figure 1 As shown, a 5-layer structure design is adopted, and the specific construction method includes the following steps: S1. In multibeam water column images, plume targets are characterized by low signal-to-noise ratio, discontinuous edges, and complex three-dimensional spatial structures. The fully connected layers at the ends of traditional convolutional networks compress spatial information, leading to the loss of three-dimensional features. Although fully convolutional networks can preserve spatial information, the standard 3D U-Net encoder still has limitations in processing plume tasks. Specifically, features in low signal-to-noise ratio regions are easily weakened during deep encoding, while shallow detail information is gradually lost during multiple downsampling processes, thus affecting the accurate characterization of target structure and edge features. Therefore, a residual three-dimensional encoder is used to replace each layer encoder of the 3D U-NET model. A hierarchical structure with cascaded downsampling paths is adopted, downsampling level by level to gradually extract multi-scale features. The architecture of the residual three-dimensional encoder is as follows: Figure 2 As shown, it consists of several residual blocks, each composed of a residual mapping path and an identity mapping path. The residual mapping path performs a stepwise nonlinear transformation on the input features through two layers of 3D convolution to enhance feature discrimination capability. The identity mapping path achieves cross-layer transfer and information reuse of input features through direct connections or channel mapping. The residual mapping path and the identity mapping path are fused at the output to achieve feature reconstruction and enhancement. For example, the mapping of the residual path is... The mapping of the identity path is The output of the residual block is The residual learning process can then be represented as:

[0035]

[0036] Through the aforementioned structural design, shallow spatial details are effectively preserved during deep feature propagation, while high-level semantic information is gradually enhanced, thereby improving the network's ability to model complex 3D structures. Furthermore, residual connections alleviate the gradient decay problem in deep networks through cross-layer information transfer and promote the effective fusion of original acoustic textures and high-level semantic features. Compared to the standard 3D U-Net encoder, the proposed coding structure exhibits stronger feature preservation capabilities under low signal-to-noise ratio conditions and provides richer semantic and more powerful feature representations for subsequent 3D segmentation tasks.

[0037] S2. Seafloor bubble plumes exhibit a feather-like structure, typically perpendicular to the seabed, reaching heights of hundreds to thousands of meters. From images, acoustic energy attenuates with distance, and bubbles dissolve, merge, and oscillate during their ascent, resulting in extremely weak signals at the top and edges of the plume, indistinguishable from background noise. While traditional convolutional neural networks demonstrate good local feature extraction capabilities in image segmentation, their limited receptive field hinders their ability to model long-distance spatial dependencies. When capturing structural targets with significant spatial continuity, such as bubble plumes, traditional methods are prone to structural breaks or boundary loss during segmentation, particularly in the weak signal region at the top of the plume. This leads to information loss, resulting in an underestimation of plume length and volume, thus affecting subsequent computational accuracy. Therefore, a three-dimensional sparse self-attention visual transformer is embedded at the output of the bottleneck layer (layer 5 in this embodiment) of the 3D U-NET model, with the architecture as follows: Figure 3 As shown, by integrating spatial context information, global feature enhancement is performed on the deep semantic features after multiple downsampling, improving the overall recognition accuracy of the feather flow structure. The core idea of ​​constructing the 3D sparse self-attention visual transformer is to introduce the self-attention mechanism of the Transformer into the 3D visual feature modeling process, and to compensate for the shortcomings of convolutional networks in long-range dependency modeling by establishing global dependencies between different spatial locations. The 3D sparse self-attention visual transformer is embedded in the bottleneck layer of the network (layer 5 in this embodiment, e.g., ...). Figure 1 As shown in the figure, global feature enhancement is performed on the deep semantic features after multiple downsampling to improve the network's ability to express complex 3D structural targets. The specific steps include: S21. Obtain the deep low-resolution 3D feature map from the bottleneck layer. Expanding into a sequence, the original three-dimensional feature representation is converted into a set of token sequences. ,in, Indicates batch size. Indicates the number of feature channels. Indicates the number of tokens in the sequence; S22, through a learnable linear mapping matrix Generate query vectors respectively Key vector Sum value vector The correlation weights between spatial features are calculated using a self-attention mechanism; the calculation method is as follows:

[0038] in, This is the dimension of each attention head. Scaling factor. This self-attention calculation process controls the numerical range of the inner product result, thereby improving numerical stability during training and avoiding the gradient vanishing problem. Through this process, the model can dynamically weight features at different spatial locations based on global context information, thus achieving effective aggregation of global feature information and more accurately characterizing the long-range structural relationships of bubble plumes in three-dimensional space.

[0039] S23. Based on the correlation weight, output the enhanced 3D feature map. .

[0040] However, directly calculating global self-attention on 3D volume data has high computational complexity, with a time complexity of O(n log n). ,in and These represent the depth, height, and width of the feature map, respectively. This represents the dimension of the token. As the resolution of the 3D features increases, the computational scale of self-attention will grow rapidly, resulting in high computational overhead.

[0041] To reduce computational overhead while maintaining model expressiveness, this embodiment employs a lightweight design for the 3D sparse self-attention visual transformer structure and network embedding strategy. At the structural level, a compact encoding structure is adopted to reduce model complexity. Specifically, only a single-layer encoding structure is retained, the number of attention heads is set to 2, and the MLP scaling ratio is set to 2. This effectively reduces model parameters and computational overhead while maintaining global modeling capabilities, making it more suitable for applications with relatively limited data scale, such as feather currents in 3D oceans. At the embedding strategy level, the 3D sparse self-attention visual transformer is embedded in the bottleneck layer of the network, rather than being introduced into the high-resolution feature layer. This embedding strategy is designed based on two main considerations. First, thanks to resolution compression during the encoding stage, the bottleneck layer feature map has undergone multiple downsamplings, significantly reducing its spatial resolution. This effectively reduces the number of tokens involved in attention computation, thereby significantly reducing the complexity of 3D self-attention computation. Second, after multiple convolutions and downsampling, the low-resolution features of the bottleneck layer have a larger receptive field, enabling the integration of a wider range of spatial contextual information. For bubble plumes in multibeam water column images, they typically exhibit a continuous, ribbon-like structure in three-dimensional space, spanning a large spatial scale along the vertical direction. Therefore, introducing a 3D sparse self-attention visual transformer into the bottleneck layer can more effectively capture the overall spatial structural features of the plume, thereby enhancing the model's ability to model its vertically continuous morphology. In summary, through a lightweight self-attention structure design and a bottleneck layer embedding strategy, global modeling of the overall spatial structure of the plume is achieved while effectively controlling the computational complexity of 3D attention, thus providing a more complete feature representation with global semantic information for the subsequent decoding stage.

[0042] S3. In the U-shaped network of the standard 3D U-NET network, the encoding layer provides rich spatial information while the decoding layer provides rich semantic information. However, its framework directly connects encoder features to the decoder and uses skip connections to compensate for the loss of spatial details in deeper layers. This often fails to fully utilize the rich semantic features of different layers, especially in the 3D segmentation task of multibeam water column images. Due to the extreme class imbalance, weak boundary of data targets, and strong noise in multibeam water column images, there are significant semantic gaps between different regions. Therefore, a dual-gated attention fusion unit is embedded at each connection point of the encoder and decoder in the 3D U-NET model, except for the bottleneck layer. Its architecture is as follows: Figure 4 As shown, gated attention modeling of bidirectional features along the channel dimension enhances the complementarity and information interaction capabilities between features, preserving the multi-scale feature space structure provided by the encoder structure, as well as the semantic information contained in the decoder. The first branch of the dual-gated attention fusion derives from the semantic information of the decoder, filtering out target-related regions from the spatial details of the encoder, thereby strengthening weak boundary targets and suppressing irrelevant strong noise regions. The other branch is dominated by the spatial information of skip connections, using rich spatial context to calibrate the semantic features of the decoder, extracting semantic information highly consistent with the spatial distribution of the target, thus providing richer and more accurate feature representations for skip connections and enhancing the integrity of the target structure. The construction of the dual-gated attention fusion derives includes the following steps: S31, Encoder features and decoder features Mapped to dimensions respectively intermediate features and The calculation is as follows:

[0043]

[0044] in, For batch normalization operations, It is a 1×1×1 convolution.

[0045] S32. Add the mapped intermediate features and obtain the fused features using the ReLU activation function. ; to integrate features The positive attention map is obtained by performing convolution operations and processing with activation functions. and reverse attention graph The calculation is as follows:

[0046]

[0047] in, , is the activation function. Forward convolution, For inverse convolution, the parameters are dynamically learned based on the different forward and inverse data during model training.

[0048] S33. Spatial modulation of intermediate features is performed using the generated attention map, and a positive attention map is used. To enhance the spatial features of the encoder Using reverse attention graphs To refine the semantic features of the decoder The calculation is as follows:

[0049]

[0050] in, For Hadama accumulation.

[0051] S34, will and Addition to obtain fusion features The number of channels projected back to the original encoder features To obtain the final output The calculation is as follows:

[0052]

[0053] in, It is a 1×1×1 convolution.

[0054] Unlike traditional attention mechanisms that enhance features only in a single direction (usually from encoder to decoder), the dual-gated attention fusion fusion proposed in this embodiment constructs a bidirectional information interaction mechanism between the encoder and decoder, enabling semantic information to guide spatial details and spatial information to perform reverse calibration of semantic features. This allows for a more thorough exploration of the complementary relationships between cross-layer features and improves the feature fusion quality of complex 3D structures.

[0055] This embodiment uses DSC ( The DSC, IoU, and VOE scores are used as evaluation criteria for 3D segmentation models. Higher DSC and IoU scores indicate better segmentation performance, while lower VOE scores indicate better performance. The proposed ESDR-UNet3D model was quantitatively evaluated using a constructed dataset. Experimental results showed a DSC of 0.721, an IoU of 0.571, and a VOE of 0.429. Compared to CNN-based 3D segmentation models such as FCN3D, SegNet3D, VNet3D, and 3D U-Net, the DSC performance was improved by 42.4%, 31.3%, 22.4%, and 10.1%, respectively. Compared to the Transformer-based 3D segmentation model UNETR++, the DSC performance was improved by 9.9%. Comparative experiments demonstrate that the ESDR-UNet3D model provided in this embodiment has significant advantages in both performance and efficiency compared to mainstream 3D segmentation models for multibeam water column image bubble plume data.

[0056] Example 2 Based on the same inventive concept, this embodiment also provides a three-dimensional segmentation system for bubble plume targets in multibeam water column images. When the system is running, it executes the method described in Embodiment 1, including the following modules: The feature extraction module is used to replace the encoder of each layer of the 3D U-NET model with a residual 3D encoder. It adopts a hierarchical structure of cascaded downsampling paths, downsamples step by step, and gradually extracts multi-scale features. The feature enhancement module is used to embed a 3D sparse self-attention visual transformer at the bottleneck layer output of the 3D U-NET model, integrate spatial context information, and perform global feature enhancement on the deep semantic features after multiple downsampling. The bidirectional information interaction module is used to embed a dual-gated attention fusion unit at the encoder and decoder connection of the 3D U-NET model to perform gated attention modeling of bidirectional features along the channel dimension.

[0057] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A 3D segmentation method for bubble plume targets in multibeam water column images, based on the standard 3D U-NET model, characterized in that... Includes the following steps: S1. A residual 3D encoder is used to replace the encoder of each layer of the 3D U-NET model. A hierarchical structure with cascaded downsampling paths is adopted to downsample step by step and extract multi-scale features gradually. S2. Embed a 3D sparse self-attention visual transformer at the bottleneck layer output of the 3D U-NET model to integrate spatial context information and perform global feature enhancement on the deep semantic features after multiple downsampling. S3. At each connection point of the encoder and decoder of the 3D U-NET model except for the bottleneck layer, a dual-gated attention fusion unit is embedded to perform gated attention modeling of bidirectional features along the channel dimension.

2. The three-dimensional segmentation method for bubble plume flow targets in multibeam water column images according to claim 1, characterized in that, The residual 3D encoder consists of several residual blocks, each composed of a residual mapping path and an identity mapping path. The residual mapping path performs a stepwise nonlinear transformation on the input features through two layers of 3D convolution. The identity mapping path achieves cross-layer transfer and information reuse of the input features through direct connection or channel mapping. The residual mapping path and the identity mapping path are fused at the output.

3. The three-dimensional segmentation method for bubble plume flow targets in multibeam water column images according to claim 1, characterized in that, The construction of the three-dimensional sparse self-attention visual transformer includes the following steps: S21. Obtain the deep low-resolution 3D feature map from the bottleneck layer. Expanding into a sequence, the original three-dimensional feature representation is converted into a set of token sequences. ,in, Indicates batch size. Indicates the number of feature channels. Indicates the number of tokens in the sequence; S22, through a learnable linear mapping matrix Generate query vectors respectively Key vector Sum value vector And use the self-attention mechanism to calculate the correlation weights between spatial features; S23. Based on the correlation weight, output the enhanced 3D feature map. .

4. The three-dimensional segmentation method for bubble plume flow targets in multibeam water column images according to claim 3, characterized in that, The correlation weights between the computational spatial features are as follows: in, It is the dimension of each attention head.

5. The three-dimensional segmentation method for bubble plume targets in multibeam water column images according to claim 1, characterized in that, The construction of the dual-gated attention fusion processor includes the following steps: S31, Encoder features and decoder features Mapped to dimensions respectively intermediate features and , The number of channels mapped to the decoder. It is the product of depth, height, and width; S32. Add the mapped intermediate features and obtain the fused features using the ReLU activation function. ; to integrate features The positive attention map is obtained by performing convolution operations and processing with activation functions. and reverse attention graph ; S33. Spatial modulation of intermediate features is performed using the generated attention map, and a positive attention map is used. To enhance the spatial features of the encoder Using reverse attention graphs To refine the semantic features of the decoder ; S34, will and Addition to obtain fusion features The number of channels projected back to the original encoder features To obtain the final output .

6. The three-dimensional segmentation method for bubble plume targets in multibeam water column images according to claim 5, characterized in that, The intermediate features and The calculation is as follows: in, For batch normalization operations, It is a 1×1×1 convolution.

7. The three-dimensional segmentation method for bubble plume targets in multibeam water column images according to claim 6, characterized in that, The positive attention map and reverse attention graph The calculation is as follows: in, , is the activation function. Forward convolution, This is an inverse convolution.

8. The three-dimensional segmentation method for bubble plume flow targets in multibeam water column images according to claim 7, characterized in that, The features and The calculation is as follows: in, For Hadama accumulation.

9. The three-dimensional segmentation method for bubble plume targets in multibeam water column images according to claim 8, characterized in that, The final output The calculation is as follows: in, It is a 1×1×1 convolution.

10. A three-dimensional segmentation system for bubble plume-like flow targets in multibeam water column images, characterized in that, Includes the following modules: The feature extraction module is used to replace the encoder of each layer of the 3D U-NET model with a residual 3D encoder. It adopts a hierarchical structure of cascaded downsampling paths, downsamples step by step, and gradually extracts multi-scale features. The feature enhancement module is used to embed a 3D sparse self-attention visual transformer at the bottleneck layer output of the 3D U-NET model, integrate spatial context information, and perform global feature enhancement on the deep semantic features after multiple downsampling. The bidirectional information interaction module is used to embed a dual-gated attention fusion unit at the encoder and decoder connection of the 3D U-NET model to perform gated attention modeling of bidirectional features along the channel dimension.

Images