Ocean sub-mesoscale process instance segmentation method based on frame supervision
By employing a box-supervised instance segmentation method for ocean sub-mesoscale processes, combined with a dual-domain enhanced pyramid network and prior knowledge of ocean dynamics, the problem of low segmentation accuracy for ocean sub-mesoscale processes is solved, achieving high-precision pixel-level segmentation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-10
- Publication Date
- 2026-04-07
AI Technical Summary
Existing box-supervised instance segmentation techniques suffer from low segmentation accuracy in ocean sub-mesoscale processes due to weak and gradual boundary changes that are submerged in complex background noise. Furthermore, the generated pseudo-labels suffer from boundary positioning drift, region breakage, or adhesion issues.
A box-supervised ocean sub-mesoscale process instance segmentation method is adopted. By training teacher and student models, and combining a dual-domain augmented pyramid network, a multi-scale aggregation module, and an instance segmentation head module, the method utilizes prior knowledge of ocean dynamics for pseudo-mask correction and geometric constraints to improve segmentation accuracy.
It achieves high-precision pixel-level instance segmentation with only low-cost bounding box labels, effectively suppresses background noise, enhances weak gradient boundary feature representation, solves the problem of inaccurate boundary localization, and improves the instance segmentation accuracy of marine sub-mesoscale processes.
Smart Images
Figure CN121811047A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of marine remote sensing information processing, and in particular to a method for segmenting marine sub-mesoscale process instances based on frame supervision. Background Technology
[0002] Submesoscale oceanic processes are of great scientific significance for understanding ocean dynamic mechanisms and protecting ecosystems. With the explosive growth of remote sensing data, extracting submesoscale oceanic processes using deep learning techniques has become a research hotspot. However, traditional fully supervised instance segmentation techniques heavily rely on pixel-level fine-grained annotation data, which is not only time-consuming and labor-intensive but also places extremely high demands on the oceanographic expertise of the annotators.
[0003] Existing instance segmentation techniques only require the bounding rectangle of the target to complete box-supervised instance segmentation. This general box-supervised instance segmentation technique has attracted widespread attention due to its low annotation cost advantage. However, existing general box-supervised instance segmentation techniques rely on strong boundary information for segmentation inference. The boundaries of ocean sub-mesoscale processes usually exhibit weak gradient changes and are often submerged in complex ocean background noise, lacking clear boundaries. When this method is directly applied to the extraction of ocean sub-mesoscale processes, it is often difficult to accurately separate the target from the weak gradient background. This results in serious problems such as boundary positioning drift, regional breakage, or adhesion in the generated pseudo-labels, ultimately leading to low accuracy in the segmentation of ocean sub-mesoscale process instances. Summary of the Invention
[0004] This invention addresses the problem of low segmentation accuracy in existing box-supervised instance segmentation methods by proposing a box-supervised method for segmenting marine sub-mesoscale processes.
[0005] A box-supervised instance segmentation method for sub-mesoscale ocean processes includes the following steps:
[0006] The remote sensing image of the chlorophyll concentration to be tested is input into the trained student model to obtain the instance segmentation results of marine sub-mesoscale processes. ;
[0007] The trained student model is obtained through the following method:
[0008] Step 1: Utilize remote sensing images of chlorophyll concentration and bounding boxes of sub-mesoscale oceanic processes. The first training set was formed;
[0009] Step 2: Train the teacher model using the first training set. The trained teacher model outputs an initial set of prediction masks. ;
[0010] Step 3: For the initial prediction mask set Perform mask correction to obtain a pseudo mask. ;
[0011] Step 4: Apply the pseudo-mask And the frame label for oceanic submesoscale processes As a label for chlorophyll concentration remote sensing images, using chlorophyll concentration remote sensing images and pseudomasks And the frame label for oceanic submesoscale processes A second training set is created, and the student model is trained and optimized using the second training set. To obtain a well-trained student model;
[0012] The student model and the teacher model have the same structure.
[0013] Furthermore, the teacher model It includes: backbone network module, dual-domain enhanced pyramid module, multi-scale aggregation module, and instance segmentation header module;
[0014] The backbone network module is used to extract feature maps from remote sensing images of chlorophyll concentration. ;
[0015] in, It is a feature icon number;
[0016] The dual-domain augmentation pyramid module is used to acquire feature maps. Frequency domain characteristics and spatial domain features and will and Fusion to obtain dual-domain enhanced features ;
[0017] The multi-scale aggregation module uses the feature pyramid to perform multi-scale aggregation on all dual-domain enhanced features to obtain multi-scale aggregated features.
[0018] The instance segmentation head module uses multi-scale aggregated features to obtain an initial prediction mask set. And the detection score for each initial prediction mask.
[0019] Furthermore, the dual-domain enhanced pyramid module includes: a frequency domain feature extraction branch, a spatial domain feature extraction branch, and a dual-domain enhanced feature acquisition unit;
[0020] The frequency domain feature extraction branch includes: a fast Fourier transform unit, a feature enhancement unit, and an inverse fast Fourier transform unit;
[0021] The Fast Fourier Transform unit is used to process the feature map. Perform a fast Fourier transform to obtain the amplitude spectrum. and phase spectrum ;
[0022] The feature enhancement unit is used to process the amplitude spectrum respectively. and phase spectrum Perform feature enhancement to obtain the enhanced amplitude spectrum. and the enhanced phase spectrum ;
[0023] The feature enhancement unit includes a first convolutional branch and a second convolutional branch;
[0024] The first and second convolutional branches are the same; each convolutional branch includes two 3×3 convolutional layers, and the input of the second convolutional layer is the output of the first convolutional layer.
[0025] The input to the first convolution branch is the amplitude spectrum. The output is the enhanced amplitude spectrum. ;
[0026] The input to the second convolution branch is the phase spectrum. The output is the enhanced phase spectrum. ;
[0027] The inverse fast Fourier transform unit is used to process the enhanced amplitude spectrum. and the enhanced phase spectrum Perform inverse fast Fourier transform to obtain frequency domain features. ;
[0028] The spatial domain feature extraction branch includes: a gradient optimization unit, a structure preservation unit, and a spatial domain feature acquisition unit;
[0029] The gradient optimization unit is a 3×3 convolution with a stride of 1, used to extract feature maps. gradient features ;
[0030] The structure preservation unit is used to acquire feature maps. Structural features ;
[0031] The spatial domain feature acquisition unit is used to extract gradient features. With structural features Adding them together yields the spatial domain features. ;
[0032] The dual-domain enhanced feature acquisition unit utilizes spatial domain features and frequency domain features Obtain dual-domain enhanced features .
[0033] Furthermore, the structure-preserving unit is used to acquire feature maps. Structural features Specifically:
[0034] First, input features Divided into first feature components with the same number of channels With the second characteristic component ;
[0035] Then, using the first eigencomponent With the second characteristic component Obtaining structural features :
[0036]
[0037] in, It is a scaling factor. It is a bias factor. yes The first characteristic component, yes The second characteristic component, It is an instance normalization process.
[0038] Furthermore, the dual-domain enhanced feature acquisition unit utilizes spatial domain features. and frequency domain features Obtain dual-domain enhanced features Specifically:
[0039]
[0040]
[0041]
[0042]
[0043] in, It is a query vector. It is a key vector. It is a value vector. It is the scaling factor.
[0044] Furthermore, in step three, the initial prediction mask set... Perform mask correction to obtain a pseudo mask. Specifically:
[0045] A1. Obtain the initial prediction mask set Mask confidence score for each initial prediction mask The mask confidence score of each initial prediction mask is calculated. Compared with the preset mask threshold of 0.5, the initial predicted mask that is lower than the preset mask threshold is deleted;
[0046] A2. Obtain the remaining initial prediction mask after processing A1. boundary mass fraction and will Compared with a preset threshold of 0.5, obtain the boundaries of the initial prediction masks corresponding to boundary quality scores higher than the preset threshold and the boundaries corresponding to boundary quality scores lower than the preset threshold. ;
[0047] A3, Yes Perform boundary evolution correction to obtain the evolved corrected version. Boundary, using evolution-corrected The pseudo-mask is obtained by taking the boundary and the initial prediction mask boundary corresponding to the boundary quality score above the preset threshold. .
[0048] Furthermore, the mask confidence score of the initial prediction mask. It can be obtained through the following methods:
[0049]
[0050] in, It is the first The detection score of an initial predicted mask. It is the first In each prediction mask The probability that a location pixel represents a sub-mesoscale process in the ocean. It is an indicator function, when The indicator function value is 1 if the time condition is met, and 0 otherwise. It is a constant.
[0051] Furthermore, the remaining initial prediction mask after A1 processing boundary mass fraction Specifically:
[0052]
[0053] in, It is the remaining initial prediction mask number. It is the first Initial prediction mask gradient magnitude at that point It is the first Pixels in the initial prediction mask Boundary mask at the location, It is a constant.
[0054] Furthermore, the pair in A3 Perform boundary evolution correction to obtain the evolved corrected version. Boundary, using evolution-corrected The pseudo-mask is obtained by taking the boundary and the initial prediction mask boundary corresponding to the boundary quality score above the preset threshold. Specifically:
[0055] A3-1, Obtain The boundary will The boundary is used as the initial time. level set function Initialization evolution time Initialize level set function ;
[0056] A3-2, On Evolutionary Time Corresponding level set function Update, obtain evolution time Corresponding level set function Specifically:
[0057]
[0058] in, , Evolution time , The corresponding level set function, It is evolutionary time. It is the evolutionary time step. yes gradient magnitude, yes The gradient alignment term weights of the gradient vector of the mask confidence score plot. It is the weight of the curvature constraint term. for The gradient vector of the mask confidence score map. It is a curvature constraint term. It is a level set function The boundary normal vector, Let be the modulation velocity field at time t;
[0059] A3-3, Judgment Is it equal to ,like If so, execute A3-4; otherwise, let Then return to A3-2;
[0060] in, It is the total evolutionary time;
[0061] A3-4, Final level set function middle Part of it is the evolved and modified version. The boundary is determined by the initial prediction mask boundary corresponding to the boundary quality score above a preset threshold and the evolved correction. The boundaries form the pseudo-mask boundaries, and then the regions within the pseudo-mask boundaries are filled to obtain the pseudo-mask M. pseudo .
[0062] Furthermore, the student model is trained using the second training set. The following loss function is used:
[0063]
[0064] in, It is the student model prediction mask. The set of boundary pixels, It's a fake mask. The set of boundary pixels, Indicates the number of boundary points. Let x be the square of the Euclidean distance from point x to the nearest point in set B.
[0065] The beneficial effects of this invention are as follows:
[0066] This invention proposes a box-supervised instance segmentation method for sub-mesoscale oceanic processes. For the first time, this invention integrates prior knowledge of ocean dynamics into a box-supervised instance segmentation framework, achieving high-precision pixel-level instance segmentation with only low-cost box labels. First, the proposed dual-domain enhanced feature pyramid network extracts and inter-fuses features in the frequency and spatial domains, effectively suppressing ocean background noise and enhancing the feature representation of weak gradient boundaries in sub-mesoscale oceanic processes. Second, the pseudo-mask correction innovatively introduces ocean dynamics priors such as gradient and curvature, using the level set evolution equation to correct the pseudo-mask, solving the problem of inaccurate boundary localization caused by weak gradients. Finally, a geometrically constrained feature consistency loss function strongly constrains the boundary consistency of the teacher-student architecture. This invention improves the instance segmentation accuracy of sub-mesoscale oceanic processes solely based on box labels. Attached Figure Description
[0067] Figure 1 This is a flowchart of the present invention;
[0068] Figure 2 This is a network structure diagram of the present invention. Detailed Implementation
[0069] Specific implementation method one: as follows Figure 1 and Figure 2As shown in the figure, the specific process of the ocean sub-mesoscale process instance segmentation method based on box supervision in this embodiment is as follows:
[0070] The remote sensing image of the chlorophyll concentration to be tested is input into the trained and optimized student model to obtain the instance segmentation results of marine sub-mesoscale processes. .
[0071] like Figure 2 As shown, the optimized student model is obtained in the following way:
[0072] Step 1: Construct a supervised instance segmentation dataset for ocean sub-mesoscale process frames, and use the supervised instance segmentation dataset for ocean sub-mesoscale process frames as the first training set;
[0073] The ocean sub-mesoscale process bounding box supervised instance segmentation dataset includes: remote sensing images of chlorophyll concentration and bounding box labels of ocean sub-mesoscale processes corresponding to the chlorophyll concentration remote sensing images. ;
[0074] Step 2: Train the teacher model using the first training set to obtain the initial prediction mask set output by the trained teacher model. ;
[0075] Step 3: For the initial prediction mask set Perform mask correction to obtain a pseudo mask. ;
[0076] Step 4: Apply the pseudo-mask And the frame label for oceanic submesoscale processes As a label for chlorophyll concentration remote sensing images, using chlorophyll concentration remote sensing images and pseudomasks And the frame label for oceanic submesoscale processes A second training set is created, and the student model is trained and optimized using the second training set. The trained and optimized student model is obtained, and then the exponential moving average strategy is used based on the student model. Parameter update teacher model Parameters;
[0077] The student model and the teacher model have the same structure.
[0078] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the teacher model... Includes: backbone network module, dual-domain enhanced pyramid module, multi-scale aggregation module, and instance segmentation header module;
[0079] The backbone network module is used to extract multi-scale feature maps from remote sensing images of chlorophyll concentration. ;
[0080] in, These are multi-scale feature maps; this invention uses ResNet101 as the backbone network to obtain four scale feature maps.
[0081] The dual-domain augmentation pyramid module is used to acquire multi-scale feature maps. Frequency domain characteristics and spatial domain features and will and Fusion to obtain dual-domain enhanced features ;
[0082] The dual-domain enhanced pyramid module includes: a frequency domain feature extraction branch, a spatial domain feature extraction branch, and a dual-domain enhanced feature acquisition unit;
[0083] The frequency domain feature extraction branch is used to obtain multi-scale feature maps. Frequency domain characteristics The frequency domain feature extraction branch includes: a fast Fourier transform unit, a feature enhancement unit, and an inverse fast Fourier transform unit.
[0084] The Fast Fourier Transform unit is used to process multi-scale feature maps. Perform a fast Fourier transform to obtain the amplitude spectrum. and phase spectrum ;
[0085] The feature enhancement unit is used to process the amplitude spectrum respectively. and phase spectrum Perform feature enhancement to obtain the enhanced amplitude spectrum. and the enhanced phase spectrum ;
[0086] The feature enhancement unit includes a first convolutional branch and a second convolutional branch;
[0087] The first convolution branch and the second convolution branch are the same;
[0088] The input to the first convolution branch is the amplitude spectrum. The output is the enhanced amplitude spectrum. ;
[0089] The input to the second convolution branch is the phase spectrum. The output is the enhanced phase spectrum. ;
[0090] Each convolutional branch consists of two convolutional layers. The input to the first convolutional layer is the amplitude spectrum or phase spectrum, and the input to the second convolutional layer is the output of the first convolutional layer. The output of the second convolutional layer is the enhanced amplitude spectrum. or enhanced phase spectrum Both convolutional layers are 3×3 kernel layers with a stride of 1.
[0091] The inverse fast Fourier transform unit is used to process the enhanced amplitude spectrum. and the enhanced phase spectrum Perform inverse fast Fourier transform to obtain frequency domain features. ;
[0092] The spatial domain feature extraction branch includes: a gradient optimization unit, a structure preservation unit, and a spatial domain feature acquisition unit;
[0093] The gradient optimization unit is a depthwise separable convolution with a kernel size of 3×3 and a stride of 1, used to extract multi-scale feature maps. gradient features ;
[0094] The structure-preserving unit is used to acquire multi-scale feature maps. Structural features ;
[0095] The structure preservation unit will input features Divided into first feature components with the same number of channels With the second characteristic component For the first characteristic component Execution instance normalization processing For the second characteristic component Perform the identity mapping; then, with the second eigencomponent... Given the condition, generate the affine transformation parameters, i.e., the scaling factors. With bias factor ; using the scaling factor With bias factor The first feature component after instance normalization Perform linear modulation, and then combine with the second characteristic component. Structural features are obtained by splicing and fusion. Its formula is:
[0096]
[0097] in, It is a scaling factor. It is a bias factor. yes The first characteristic component, yes The second characteristic component, It is an instance normalization process;
[0098] The spatial domain feature acquisition unit is used to extract gradient features. With structural features Adding them together yields the spatial domain features. ;
[0099] The dual-domain enhanced feature acquisition unit utilizes a cross-attention mechanism to... Obtain the query vector by performing linear projection. For frequency domain features Perform linear projection mapping to key vectors Sum value vector and utilize , and Obtain dual-domain enhanced features Specifically:
[0100]
[0101]
[0102]
[0103]
[0104] in, It is a query vector. It is a key vector. It is a value vector. It is the scaling factor;
[0105] The dual-domain enhanced feature acquisition unit will acquire the spatial domain features of the current layer. Mapped to a query vector via linear projection The frequency domain features of the current layer Mapped to key vectors via linear projection Sum value vector Next, the query vector is calculated. With key vector The dot product of the transpose of the vector is normalized by dividing by a scaling factor and then processed by the Softmax function to obtain the attention weight map. Finally, the attention weight map is combined with the value vector. Multiply to obtain the fused output features of the current layer. This allows the frequency domain global structural information to guide and enhance the spatial domain detailed features.
[0106] The multi-scale aggregation module uses the feature pyramid to perform multi-scale aggregation on all dual-domain enhanced features to obtain multi-scale aggregated features.
[0107] The multi-scale aggregation module utilizes a top-down feature pyramid path to enhance the dual-domain features of the current layer. Upsampled features of the previous level Multi-scale aggregation is performed to ultimately obtain multi-scale aggregated features;
[0108] The instance segmentation head module uses multi-scale aggregated features to obtain an initial prediction mask set. And the detection score for each initial prediction mask.
[0109] The other steps and parameters are the same as in Specific Implementation Method 1.
[0110] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: the initial prediction mask set in step three... Perform mask correction to obtain a pseudo mask. Specifically:
[0111] A1. Obtain the initial prediction mask set Mask confidence score for each initial prediction mask The mask confidence score of each initial prediction mask is calculated. Compared with the preset mask threshold of 0.5, the initial predicted mask that is lower than the preset mask threshold is deleted;
[0112] No. Initial prediction mask Mask confidence score It can be obtained through the following methods:
[0113]
[0114] in, It is the first The detection score of an initial predicted mask. It is the first In each prediction mask The probability that a location pixel represents a sub-mesoscale process in the ocean. It is an indicator function, when The indicator function value is 1 if the time condition is met, and 0 otherwise. It is a minimal constant that prevents division by zero errors;
[0115] A2. Obtain the remaining initial prediction mask after processing A1. boundary mass fraction and will Compared to a preset threshold of 0.5, the boundaries of the initial prediction mask corresponding to boundary quality scores higher than the preset threshold are partial boundaries of the pseudo mask. The initial prediction mask corresponding to boundary quality scores lower than the preset threshold is then obtained. ;
[0116] No. Boundary quality score of the remaining initial prediction masks Specifically:
[0117]
[0118] in, It is the remaining initial prediction mask number. It is the first Initial prediction mask gradient magnitude at that point It is the first pixels on the boundary in an initial prediction mask Boundary mask at the location, It is a minimal constant that prevents division by zero errors.
[0119] A3. Using the level set evolution equation of gradient and curvature modulation to... Perform boundary evolution correction to obtain the evolved corrected version. Boundary, using evolution-corrected The pseudo-mask is obtained by taking the boundary and the initial prediction mask boundary corresponding to the boundary quality score above the preset threshold. Specifically:
[0120] A3-1, Obtain The boundary will The boundary is used as the initial time. level set function Initialization evolution time Initialize level set function ;
[0121] A3-2, Using the level set evolution equation with gradient and curvature modulation to analyze evolution time Corresponding level set function Update, obtain evolution time Corresponding level set function Specifically:
[0122]
[0123] in, , Evolution time , The corresponding level set function, It is evolutionary time. It is the evolutionary time step. yes gradient magnitude, yes The gradient alignment term weights of the gradient vector of the mask confidence score plot. It is the weight of the curvature constraint term. for The gradient vector of the mask confidence score map. It is a curvature constraint term, representing the average curvature of the current level set function, used to smooth the geometry of the boundary during evolution and to simulate the surface tension characteristics of fluids to suppress noise; It is the current level set function The boundary normal vector drives the boundary to move towards the position of the high-confidence gradient; Let be the modulated velocity field at time t, including the gradient alignment term. With curvature constraint term .
[0124] A3-3, Judgment Is it equal to ,like If so, execute A3-4; otherwise, let Then return to A3-2;
[0125] in, It is the total evolutionary time;
[0126] A3-4, Final level set function middle Part of it is the evolved and modified version. The boundary is determined by the initial prediction mask boundary corresponding to the boundary quality score above a preset threshold and the evolved correction. The boundaries form the pseudo-mask boundaries, and then the regions within the pseudo-mask boundaries are filled to generate a pseudo-mask M with smooth geometric edges that fits the characteristics of sub-mesoscale oceanic processes. pseudo .
[0127] The other steps and parameters are the same as in specific implementation method one or two.
[0128] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: the student model is trained and optimized using the second training set. Specifically:
[0129] Using geometric consistency constraint loss function Backpropagation optimizes the student model ;
[0130] The geometric consistency constraint loss function Specifically:
[0131]
[0132] in, , These are the student model prediction masks. Boundary pixel set and pseudomask The set of boundary pixels, Indicates the number of boundary points. Let x be the square of the Euclidean distance from point x to the nearest point in set B.
[0133] This invention uses a geometric consistency constraint loss function Constraints are imposed on the instance segmentation prediction branch of the student model. Building upon the weakly supervised localization based on bounding boxes, a strong supervisory signal based on high-quality pseudo-labels is further introduced, constructing a multi-layered supervision system. By continuously optimizing the geometric features of instance-level labels during training, the student model is guided to gradually learn more refined boundary information, thereby improving the final segmentation accuracy.
[0134] The other steps and parameters are the same as in any of the specific implementation methods one to three.
[0135] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for instance segmentation of ocean sub-mesoscale processes based on box supervision, characterized in that... The specific process of the method is as follows: The remote sensing image of the chlorophyll concentration to be tested is input into the trained student model to obtain the instance segmentation results of marine sub-mesoscale processes. ; The trained student model is obtained through the following method: Step 1: Utilize remote sensing images of chlorophyll concentration and bounding boxes of sub-mesoscale oceanic processes. The first training set was formed; Step 2: Train the teacher model using the first training set. The trained teacher model outputs an initial set of prediction masks. ; Step 3: For the initial prediction mask set Perform mask correction to obtain a pseudo mask. ; Step 4: Apply the pseudo-mask And the frame label for oceanic submesoscale processes As a label for chlorophyll concentration remote sensing images, using chlorophyll concentration remote sensing images and pseudomasks And the frame label for oceanic submesoscale processes A second training set is created, and the student model is trained and optimized using the second training set. To obtain a well-trained student model; The student model and the teacher model have the same structure.
2. The method for segmenting marine sub-mesoscale processes based on box supervision according to claim 1, characterized in that: The teacher model It includes: backbone network module, dual-domain enhanced pyramid module, multi-scale aggregation module, and instance segmentation header module; The backbone network module is used to extract feature maps from remote sensing images of chlorophyll concentration. ; in, It is a feature icon number; The dual-domain augmentation pyramid module is used to acquire feature maps. Frequency domain characteristics and spatial domain features and will and Fusion to obtain dual-domain enhanced features ; The multi-scale aggregation module uses the feature pyramid to perform multi-scale aggregation on all dual-domain enhanced features to obtain multi-scale aggregated features. The instance segmentation head module uses multi-scale aggregated features to obtain an initial prediction mask set. And the detection score for each initial prediction mask.
3. The method for segmenting marine sub-mesoscale processes based on frame supervision according to claim 2, characterized in that: The dual-domain enhanced pyramid module includes: a frequency domain feature extraction branch, a spatial domain feature extraction branch, and a dual-domain enhanced feature acquisition unit; The frequency domain feature extraction branch includes: a fast Fourier transform unit, a feature enhancement unit, and an inverse fast Fourier transform unit; The Fast Fourier Transform unit is used to process the feature map. Perform a fast Fourier transform to obtain the amplitude spectrum. and phase spectrum ; The feature enhancement unit is used to process the amplitude spectrum respectively. and phase spectrum Perform feature enhancement to obtain the enhanced amplitude spectrum. and the enhanced phase spectrum ; The feature enhancement unit includes a first convolutional branch and a second convolutional branch; The first and second convolutional branches are the same; each convolutional branch includes two 3×3 convolutional layers, and the input of the second convolutional layer is the output of the first convolutional layer. The input to the first convolution branch is the amplitude spectrum. The output is the enhanced amplitude spectrum. ; The input to the second convolution branch is the phase spectrum. The output is the enhanced phase spectrum. ; The inverse fast Fourier transform unit is used to process the enhanced amplitude spectrum. and the enhanced phase spectrum Perform inverse fast Fourier transform to obtain frequency domain features. ; The spatial domain feature extraction branch includes: a gradient optimization unit, a structure preservation unit, and a spatial domain feature acquisition unit; The gradient optimization unit is a 3×3 convolution with a stride of 1, used to extract feature maps. gradient features ; The structure preservation unit is used to acquire feature maps. Structural features ; The spatial domain feature acquisition unit is used to extract gradient features. With structural features Adding them together yields the spatial domain features. ; The dual-domain enhanced feature acquisition unit utilizes spatial domain features and frequency domain features Obtain dual-domain enhanced features .
4. The method for segmenting marine sub-mesoscale processes based on box supervision according to claim 3, characterized in that: The structure preservation unit is used to acquire feature maps. Structural features Specifically: First, input features Divided into first feature components with the same number of channels With the second characteristic component ; Then, using the first eigencomponent With the second characteristic component Obtaining structural features : in, It is a scaling factor. It is a bias factor. yes The first characteristic component, yes The second characteristic component, It is an instance normalization process.
5. The method for segmenting marine sub-mesoscale processes based on box supervision according to claim 4, characterized in that: The dual-domain enhanced feature acquisition unit utilizes spatial domain features and frequency domain features Obtain dual-domain enhanced features Specifically: in, It is a query vector. It is a key vector. It is a value vector. It is the scaling factor.
6. The method for segmenting marine sub-mesoscale processes based on box supervision according to claim 5, characterized in that: Step three involves the initial prediction mask set. Perform mask correction to obtain a pseudo mask. Specifically: A1. Obtain the initial prediction mask set Mask confidence score for each initial prediction mask The mask confidence score of each initial prediction mask is calculated. Compared with the preset mask threshold of 0.5, the initial predicted mask that is lower than the preset mask threshold is deleted; A2. Obtain the remaining initial prediction mask after processing A1. boundary mass fraction and will Compared with a preset threshold of 0.5, obtain the boundaries of the initial prediction masks corresponding to boundary quality scores higher than the preset threshold and the boundaries corresponding to boundary quality scores lower than the preset threshold. ; A3, Yes Perform boundary evolution correction to obtain the evolved corrected version. Boundary, using evolution-corrected The pseudo-mask is obtained by taking the boundary and the initial prediction mask boundary corresponding to the boundary quality score above the preset threshold. .
7. The method for segmenting marine sub-mesoscale processes based on box supervision according to claim 6, characterized in that: Mask confidence score of the initial prediction mask It can be obtained through the following methods: in, It is the first The detection score of an initial predicted mask. It is the first In each prediction mask The probability that a location pixel represents a sub-mesoscale process in the ocean. It is an indicator function, when The indicator function value is 1 if the time condition is met, and 0 otherwise. It is a constant.
8. The method for segmenting marine sub-mesoscale processes based on box supervision according to claim 7, characterized in that: Remaining initial prediction mask after A1 processing boundary mass fraction Specifically: in, It is the remaining initial prediction mask number. It is the first Initial prediction mask gradient magnitude at that point It is the first Pixels in the initial prediction mask Boundary mask at the location, It is a constant.
9. A method for segmenting marine sub-mesoscale processes based on box supervision according to claim 8, characterized in that: The pair in A3 Perform boundary evolution correction to obtain the evolved corrected version. Boundary, using evolution-corrected The pseudo-mask is obtained by taking the boundary and the initial prediction mask boundary corresponding to the boundary quality score above the preset threshold. Specifically: A3-1, Obtain The boundary will The boundary is used as the initial time. level set function Initialization evolution time Initialize level set function ; A3-2, On Evolutionary Time Corresponding level set function Update, obtain evolution time Corresponding level set function Specifically: in, , These are evolution time. , The corresponding level set function, It is evolutionary time. It is the evolutionary time step. yes gradient magnitude, yes The gradient alignment term weights of the gradient vector of the mask confidence score plot. It is the weight of the curvature constraint term. for The gradient vector of the mask confidence score map. It is a curvature constraint term. It is a level set function The boundary normal vector, Let be the modulation velocity field at time t; A3-3, Judgment Is it equal to ,like If so, execute A3-4; otherwise, let Then return to A3-2; in, It is the total evolutionary time; A3-4, Final level set function middle Part of it is the evolved and modified version. The boundary is determined by the initial prediction mask boundary corresponding to the boundary quality score above a preset threshold and the evolved correction. The boundaries form the pseudo-mask boundaries, and then the regions within the pseudo-mask boundaries are filled to obtain the pseudo-mask M. pseudo .
10. A method for segmenting marine sub-mesoscale processes based on box supervision according to claim 9, characterized in that: The student model is trained using the second training set. The following loss function is used: in, It is the student model prediction mask. The set of boundary pixels, It's a fake mask. The set of boundary pixels, Indicates the number of boundary points. Let x be the square of the Euclidean distance from point x to the nearest point in set B.