Underwater target measurement method, system and equipment based on salient region sparse matching

Abstract

The invention discloses an underwater target measurement method, system and device for salient region sparse matching in the technical field of underwater optical imaging and computer vision, and the method comprises the steps: calculating an intensity image and a linear polarization degree image based on a Stokes vector according to an obtained underwater binocular polarization image; according to the intensity image and the linear polarization degree image, performing image fusion by using a polarization image fusion network to obtain a fused image; according to the linear polarization degree image and the fusion image, performing target detection by using a saliency target detection network to obtain a target area mask; extracting feature points in the target area of the underwater binocular polarization image by using a SuperPoint network according to the target area mask, and generating a feature descriptor; according to the feature points and the feature descriptors, binocular image matching is carried out through a SuperGlue network, and matching point pairs are output; and calculating the distance from the target to the polarization binocular camera according to the matching point pairs and the parameters of the polarization binocular camera.

Description

Technical Field

[0001] This invention relates to the fields of underwater optical imaging and computer vision technology, and in particular to a method, system and device for measuring underwater targets with sparse matching of significant regions. Background Technology

[0002] Underwater target measurement has important applications in marine engineering, underwater search and rescue, and marine resource exploration. Its core is to obtain key information such as the three-dimensional size and distance of the target through optical imaging and computer vision technology. However, the underwater environment is complex and variable. Water scattering and absorption effects lead to problems such as low contrast, blurred texture, and color distortion in images, which seriously affect the accuracy of subsequent feature extraction and matching.

[0003] Existing technologies generally suffer from the following problems:

[0004] (1) Traditional underwater target measurement methods mostly rely on single-modal intensity images and use stereo matching technology to calculate parallax to achieve three-dimensional reconstruction. However, in low-texture and high-turbidity scenes, the matching accuracy drops significantly, and mismatches are prone to occur, leading to increased measurement errors.

[0005] (2) Existing feature matching methods are mostly performed within the entire map area, which is easily affected by complex underwater backgrounds and has problems such as large computational load and redundant matching. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of low contrast, sparse texture, and inaccurate measurement in existing underwater environment images, and to provide an underwater target measurement method, system, and device with sparse matching of significant regions, which can effectively improve the accuracy and robustness of underwater target measurement.

[0007] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0008] In a first aspect, the present invention provides a method for measuring underwater targets using sparse matching of salient regions, comprising:

[0009] Based on the acquired underwater binocular polarization images, intensity images and linear polarization images are calculated using Stokes vectors.

[0010] Based on the intensity image and the linear polarization degree image, an image fusion network is used to perform image fusion to obtain a fused image.

[0011] Based on the linear polarization degree image and the fused image, a salient target detection network is used to perform target detection and obtain the target region mask;

[0012] Based on the target region mask, feature points within the target region of the underwater binocular polarization image are extracted using the SuperPoint network and feature descriptors are generated.

[0013] Based on the feature points and feature descriptors, stereo image matching is performed using the SuperGlue network to output matching point pairs.

[0014] Based on the matching point pairs and the polarization binocular camera parameters, the distance from the target to the polarization binocular camera is calculated.

[0015] Optionally, the intensity image is obtained using the following formula:

[0016] ,

[0017] in, Represents an intensity image. This represents an underwater binocular polarization image with a polarization angle of 0 degrees. This represents an underwater binocular polarization image with a polarization angle of 45 degrees. This represents an underwater binocular polarization image with a polarization angle of 90 degrees. This represents an underwater binocular polarization image with a polarization angle of 135 degrees.

[0018] The linear polarization degree image is obtained by the following formula:

[0019] ,

[0020] in, Represents the linear polarization degree image. This represents the first linearly polarized light. This indicates the second linearly polarized light.

[0021] Optionally, the polarization image fusion network includes a first encoder module, a second encoder module, a fusion module, and a decoder module;

[0022] The first encoder module extracts the brightness features of the intensity image through grouped spatial convolution and dense blocks;

[0023] The second encoder module extracts the brightness features of the linear polarization image through grouped spatial convolution and dense blocks;

[0024] The first encoder module and the second encoder module share weights;

[0025] The fusion module adaptively fuses the brightness features of the intensity image and polarization image through channel attention and spatial attention mechanisms to obtain fused features;

[0026] The decoder outputs a fused image by reconstructing the fusion features and combining them with the chromaticity information of the intensity image;

[0027] If the intensity image and linear polarization image are color images, perform the following steps:

[0028] The intensity image and linear polarization image are converted from the RGB color space to the YCrCb color space, and then input into the first encoder module and the second encoder module respectively.

[0029] The fused image is converted from the YCrCb color space to the RGB color space.

[0030] Optionally, the total loss function of the polarization image fusion network is as follows:

[0031] ,

[0032] in, This represents the total loss of the polarization image fusion network. The weighting parameters represent the relationship between multi-scale weighted structural similarity loss and polarization intensity loss. This represents the multi-scale weighted structural similarity loss. Indicates the loss of polarization intensity;

[0033] The multi-scale weighted structural similarity loss It can be obtained through the following formula:

[0034] ,

[0035] in, Indicates the weighting coefficient. Represents the linear polarization degree image. Represents an intensity image. Indicates a fused image. Represents a window. Indicates size is window Down and Local similarity between them Indicates size is window Down and Local similarity between them;

[0036] Local similarity It can be obtained through the following formula:

[0037] ,

[0038] in, Display window Internal belonging to Part of express The mean, Display window Internal belonging to Part of express The mean, express covariance, express covariance, express and covariance, and Indicates a constant;

[0039] Local similarity It can be obtained through the following formula:

[0040] ,

[0041] in, Display window Internal belonging to Part of express The mean, express covariance, express and covariance;

[0042] The weighting coefficient It can be obtained through the following formula:

[0043] ,

[0044] in, , Used to enhance weighting coefficients robustness and stability;

[0045] The polarization intensity loss It can be obtained through the following formula:

[0046] ,

[0047] in, express Norm, and As a weighting factor, , This represents the natural exponential function. A coefficient representing the difference between pixel values ​​in a scaled image. , Indicates the image height. Indicates the image width.

[0048] Optionally, the salient target detection network includes a dual encoder, a polarization guide head (FGH), a cross-modal attention fusion module (CAFM), a cross-level polarization guidance enhancement module (CPGM), and a LawinASPP decoder.

[0049] The dual encoders employ a ResNet18 backbone network, including a five-layer D1 encoder and a five-layer R1 encoder.

[0050] The method of using a salient object detection network for object detection includes:

[0051] The linear polarization images are input into a five-layer D1 encoder to obtain... , , , and ;

[0052] Will , , , and The polarization semantic features are extracted by inputting them into the polarization guide head (PGH) to obtain the polarization semantic features. ;

[0053] The fused image is input into the first layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion to obtain the first-level fused features. ;

[0054] The first level of fusion features The input is fed into the second layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features and the first-level fused features are then combined. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the second-level fused features. ;

[0055] The second-level fusion features The input is fed into the third layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are combined with the second-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the third-level fused features. ;

[0056] The third-level fusion feature The input is fed into the fourth layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are then combined with the third-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the fourth-level fused features. ;

[0057] The fourth-level fusion feature The input is fed into the fifth layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are then combined with the fourth-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the fifth-level fused features. ;

[0058] The first level of fusion features Second-level fusion features Third-level fusion features Fourth-level fusion features and fifth-level fusion features The input is fed into the LauinASPP decoder for target detection, resulting in a target region mask.

[0059] Optionally, the will , , , and The polarization semantic features are extracted by inputting them into the polarization guide head (PGH) to obtain the polarization semantic features. ,include:

[0060] Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ;

[0061] Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ;

[0062] Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ;

[0063] Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain That is, polarization semantic features .

[0064] Optionally, the expression for the cross-modal attention fusion module is as follows:

[0065] ,

[0066] in, This indicates that the cross-modal attention fusion module is in the first... Hierarchical output, Indicates a connection operation. Indicates element-wise multiplication. Note the operations that require attention in the representation space. Indicates the first The output of the layer D1 encoder, Indicates the first The output of layer R1 encoder;

[0067] The expression for the cross-level polarization guidance enhancement module is as follows:

[0068] ,

[0069] in, This indicates that the cross-level polarization guidance enhancement module is in the first stage. Hierarchical output, This represents the activation function. Indicates the first Hierarchical fusion features;

[0070] The expression for the LauinASPP decoder is as follows:

[0071] ,

[0072] in, This represents the output of the LawinASPP decoder. Indicates high-level characteristics, Indicates low-level features, This represents the large window attention space pyramid pooling operation. This represents the convolution operation. Indicates an upsampling operation. This indicates a cross-layer attention operation.

[0073] Optionally, the total loss function of the salient object detection network is as follows:

[0074] ,

[0075] in, This represents the total loss of the salient object detection network. Represents the binary cross-entropy loss. Indicates the crossover and union ratio loss;

[0076] The cross-union loss It can be obtained through the following formula:

[0077] ,

[0078] in, Indicates the width of the underwater binocular polarization image. Indicates the height of the underwater binocular polarization image. Indicates the coordinates of the detected target area. Indicates the actual target area coordinates;

[0079] The binary cross-entropy loss It can be obtained through the following formula:

[0080] ,

[0081] in, Represents a logarithmic function.

[0082] Secondly, the present invention provides an underwater target measurement system with sparse matching in significant regions, comprising:

[0083] The Stokes vector calculation module is used to: calculate intensity images and linear polarization images based on Stokes vectors from the acquired underwater binocular polarization images;

[0084] The image fusion module is used to: perform image fusion using a polarization image fusion network based on the intensity image and the linear polarization image to obtain a fused image;

[0085] A salient target detection module is used to: perform target detection using a salient target detection network based on the linear polarization degree image and the fused image to obtain a target region mask;

[0086] The SuperPoint module is used to: extract feature points within the target area of ​​the underwater binocular polarization image and generate feature descriptors based on the target area mask using the SuperPoint network;

[0087] The image matching module is used to: perform binocular image matching through the SuperGlue network based on the feature points and feature descriptors, and output matching point pairs;

[0088] The distance calculation module is used to calculate the distance from the target to the polarization binocular camera based on the matching point pair and the polarization binocular camera parameters.

[0089] Thirdly, the present invention provides a computer device, comprising:

[0090] Memory, used to store computer instructions;

[0091] A processor for executing the computer instructions to implement the steps of the underwater target measurement method for sparse matching of significant regions as described in any one of the first aspects.

[0092] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0093] 1. The underwater target measurement method using sparse matching in salient regions provided by this invention acquires images at different polarization angles using a polarization camera, and calculates intensity images (S0) and linear polarization degree images (DoLP) as fusion source images. Secondly, it proposes an underwater polarization image target detection method based on an interactive enhancement paradigm, fusing complementary information from intensity and linear polarization degree images through a polarization image fusion network (PIFNet), and using a salient target detection network to locate the target region. Next, based on the salient detection results, sparse feature matching is restricted to salient regions, and efficient and robust feature point matching is achieved by combining SuperPoint and SuperGlue. Finally, based on the matched point pairs and camera parameters, the three-dimensional coordinates of the target are calculated, achieving accurate measurement of size (width, height) and distance. This provides efficient and reliable technical support for underwater target detection and three-dimensional measurement, and can be applied to fields such as marine engineering and underwater robotics.

[0094] 2. The underwater target measurement system with sparse matching of salient regions provided by the present invention achieves accurate measurement of underwater targets by setting up a Stokes vector calculation module, an image fusion module, a salient target detection module, a SuperPoint module, an image matching module, and a distance calculation module, which has practical significance and good application prospects;

[0095] 3. The computer device provided by the present invention can execute the steps of the underwater target measurement method for sparse matching of significant regions provided by the present invention. Attached Figure Description

[0096] Figure 1 A flowchart of an underwater target measurement method for sparse matching of salient regions according to an embodiment of the present invention;

[0097] Figure 2 These are underwater binocular images with different polarization angles provided according to embodiments of the present invention;

[0098] Figure 3 This is a structural diagram of the polarization image fusion network PIFNet provided according to an embodiment of the present invention;

[0099] Figure 4 This is a fusion result diagram of the PIFNet polarization image fusion network provided according to an embodiment of the present invention;

[0100] Figure 5 This is a diagram of the salient target detection network structure provided in an embodiment of the present invention;

[0101] Figure 6 This is a structural diagram of the cross-modal attention fusion module (CAFM) provided according to an embodiment of the present invention;

[0102] Figure 7 This is a diagram illustrating the implementation of alternating row and column attention in the SAA module provided according to an embodiment of the present invention.

[0103] Figure 8 This is a basic unit structure diagram of row and column attention implementation in the SAA module provided according to an embodiment of the present invention;

[0104] Figure 9 This is a network structure diagram of the cross-level polarization guidance enhancement module CPGM provided according to an embodiment of the present invention;

[0105] Figure 10 This is an output diagram of the salient target detection network provided according to an embodiment of the present invention;

[0106] Figure 11 This is a diagram of the SuperPoint network structure provided according to an embodiment of the present invention;

[0107] Figure 12 This is a diagram of the SuperGlue network structure provided according to an embodiment of the present invention;

[0108] Figure 13 This is a schematic diagram of feature matching combining SuperPoint and SuperGlue according to an embodiment of the present invention;

[0109] Figure 14 This is a flowchart illustrating the feature matching process combining SuperPoint and SuperGlue according to an embodiment of the present invention.

[0110] Figure 15 This is a diagram showing the three-dimensional measurement results of an underwater target according to an embodiment of the present invention. Detailed Implementation

[0111] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0112] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0113] Example 1:

[0114] This invention discloses a method for measuring underwater targets using sparse matching in salient regions, with reference to... Figure 1 As shown, the specific steps include the following:

[0115] S1, Based on the acquired underwater binocular polarization image, calculate the intensity image and linear polarization degree image based on the Stokes vector;

[0116] S2, Based on the intensity image and the linear polarization degree image, perform image fusion using a polarization image fusion network to obtain a fused image;

[0117] S3, Based on the linear polarization degree image and the fused image, target detection is performed using a salient target detection network to obtain the target region mask;

[0118] S4. Based on the target area mask, use the SuperPoint network to extract feature points within the target area of ​​the underwater binocular polarization image and generate feature descriptors.

[0119] S5, Based on the feature points and feature descriptors, perform binocular image matching through the SuperGlue network and output matching point pairs;

[0120] S6. Calculate the distance from the target to the polarization binocular camera based on the matched point pair and the polarization binocular camera parameters.

[0121] Specifically, in step S1, underwater binocular images with four different polarization angles are acquired using a polarization binocular camera. , , as well as ,refer to Figure 2 As shown; in this embodiment, the polarization binocular camera model is LUCIDTRI050S-QC; according to the Stokes vector representation method, four... Parameters quantitatively describe the polarization and intensity information of an image:

[0122] ,

[0123] Among them, parameters Total light intensity, representing the intensity image; parameters and Indicates the difference in linearly polarized light; and These represent right-handed and left-handed polarized light, respectively; therefore, the parameters... This represents the difference in circularly polarized light. Under natural conditions, the difference in circularly polarized light can be ignored. Therefore, the intensity image is obtained by the following formula:

[0124] ,

[0125] in, Represents an intensity image. This represents an underwater binocular polarization image with a polarization angle of 0 degrees. This represents an underwater binocular polarization image with a polarization angle of 45 degrees. This represents an underwater binocular polarization image with a polarization angle of 90 degrees. This represents an underwater binocular polarization image with a polarization angle of 135 degrees.

[0126] The linear polarization degree image reflects the proportion of linearly polarized light in the total light intensity and can be used to describe the polarization characteristics of the image. It can be obtained by the following formula:

[0127] ,

[0128] in, Represents the linear polarization degree image. This represents the first linearly polarized light. This indicates the second linearly polarized light.

[0129] In step S2, this embodiment utilizes the polarization image fusion network PIFNet to fuse the linear polarization degree image and the intensity image to generate a fused image; Reference Figure 3As shown, PIFNet adopts an encoder-decoder architecture, including a first encoder module, a second encoder module, a fusion module, and a decoder module. The first and second encoder modules extract deep features through grouped spatial convolution GSConv and dense blocks (DenseBlock), and the weights of the first and second encoder modules are shared. The fusion module adaptively fuses complementary information using channel attention and spatial attention mechanisms. The decoder reconstructs the fused features and combines them with intensity image chromaticity information to output the fused image. It should be noted that for color source images, the RGB color space needs to be converted to the YCrCb color space after the input source image, and the YCrCb color space needs to be converted back to the RGB color space before outputting the fused image.

[0130] (1) Encoder module

[0131] In this embodiment, the two encoder modules share weights and are able to extract deep features from the source image. Each encoder module consists of convolutional layers and dense blocks. The module employs Grouped Spatial Convolution (GSConv) to reduce model complexity while maintaining performance; it uses dense blocks (DB) to extract and preserve features at different levels, thereby reducing the number of network parameters. The convolutional layers use... A non-linear activation function is used, employing reflection padding to preserve the size of the feature map. The dense block comprises three convolutional layers, with the input to each layer being the concatenation of the outputs of all previous convolutional layers; direct interruption effectively preserves and utilizes the features extracted by each convolutional layer while mitigating the vanishing gradient problem. Input image features. and After passing through the encoder module, the features can be obtained. and .

[0132] (2) Fusion module

[0133] The fusion module needs to simultaneously focus on the salient information in both source images. The salient information distributions of the linear polarization degree image and the intensity image are inconsistent. Therefore, channel attention and spatial attention mechanisms are introduced during the fusion process to enhance the representation and utilization of features. The channel attention mechanism first performs global max pooling and global average pooling on the input feature maps, obtaining two 1×1×C spatial context descriptions. These are then concatenated and input into a multilayer perceptron (MLP) with two fully connected layers: the first layer has C / 4 neurons and the second layer has C neurons. The output is activated by sigmoid to obtain the channel attention weight map. The spatial attention mechanism then performs max pooling and average pooling along the channel dimension on the channel-weighted feature map, obtaining two H×W×1 feature maps. These are concatenated and then activated by a 7×7 convolutional layer (stride 1, reflection padding) and sigmoid to generate the spatial attention weight map. Finally, feature refinement is achieved through two weighting operations.

[0134] Unlike channel attention, spatial attention mechanisms focus more on key regions in an image. This involves compressing the refined feature maps along the channel dimension. Two feature descriptions can be obtained. and They are concatenated along the channel dimension and fed into a convolutional layer to obtain a spatial attention map. Through feature maps Spatial attention map Multiplication operations can refine image features and highlight salient information.

[0135] By utilizing the two attention mechanisms described above, feature maps can be refined in both spatial and channel dimensions, enhancing the network's focus on salient information and reducing its focus on non-salient information. Furthermore, the introduction of spatial attention is beneficial for capturing sparse salient information in polarization images. Through these strategies, complementary information can be effectively captured and fused to obtain fused features. .

[0136] (3) Decoder module

[0137] The decoder module contains three convolutional layers, all employing grouped spatial convolution (the number of groups is the same as the encoder), with reflection padding to preserve the feature map size. Each convolutional layer is followed by a ReLU activation function, and the final convolutional layer outputs a single-channel luminance feature map. Finally, the chromaticity information of the reconstructed image and the intensity image are concatenated to obtain a polarization-fused image. During the color space conversion process, the conversion from RGB to YCrCb uses a standard conversion matrix. The Y channel retains luminance information for fusion, while the Cr and Cb channels retain chromaticity information. Before the fused image is output, it is restored to the RGB space through an inverse conversion matrix to ensure color consistency.

[0138] (4) Loss function

[0139] The fused image and the source image should exhibit high structural similarity. Therefore, the network uses multi-scale weighted structural similarity to constrain the structural similarity between the two at multiple scales. The multi-scale weighted structural similarity loss function... The calculation formula is as follows:

[0140] ,

[0141] in, Indicates the weighting coefficient. Represents the linear polarization degree image. Represents an intensity image. Indicates a fused image. Represents a window. Indicates size is window Down and Local similarity between them Indicates size is window Down and Local similarity between them;

[0142] Local similarity It can be obtained through the following formula:

[0143] ,

[0144] in, Display window Internal belonging to Part of express The mean, Display window Internal belonging to Part of express The mean, express covariance, express covariance, express and covariance, and Represent constants used to maintain stability, denoted as . and ;

[0145] Local similarity It can be obtained through the following formula:

[0146] ,

[0147] in, Display window Internal belonging to Part of express The mean, express covariance, express and covariance;

[0148] The weighting coefficient Determined by the richness of detail in the source image, a larger local variance implies more detail in the local region, which can be obtained through the following formula:

[0149] ,

[0150] in, , Used to enhance weighting coefficients robustness and stability;

[0151] Pixel information includes pixel values ​​and pixel distribution, reflecting salient information in the source image. Loss function The network focuses on variations in contrast and structure, but exhibits weak constraints on differences in pixel distribution. Therefore, a polarization intensity loss was designed into the network. It can be obtained through the following formula:

[0152] ,

[0153] in, express Norm, and As a weighting factor, , This represents the natural exponential function. A coefficient representing the difference between pixel values ​​in a scaled image. , Indicates the image height. Indicates the image width;

[0154] Based on the multi-scale weighted structural similarity loss function and polarization intensity loss The total loss function of the polarization image fusion network is obtained as follows:

[0155] ,

[0156] in, This represents the total loss of the polarization image fusion network. The weighting parameters represent the relationship between multi-scale weighted structural similarity loss and polarization intensity loss. This represents the multi-scale weighted structural similarity loss. This indicates the loss of polarization intensity.

[0157] This embodiment utilizes a multi-scale weighted structural similarity loss function. It can preserve the structural information of the source image by using polarization intensity loss. It can adaptively constrain the network, selectively extracting and fusing information. (Comprehensive loss function) Combining the characteristics of both, this process ensures the consistency of intensity during fusion while preserving the complementary saliency information of the source images in the fused image. The resulting image fusion is as follows: Figure 4 As shown.

[0158] In step S3, refer to Figure 5 As shown, the salient target detection network includes dual encoders, a polarization guide head (FGH), a cross-modal attention fusion module (CAFM), a cross-level polarization guidance enhancement module (CPGM), and a LawinASPP decoder. The dual encoders employ a ResNet18 backbone network, comprising a five-layer D1 encoder and a five-layer R1 encoder.

[0159] (1) Polarization guide head FGH

[0160] The polarization guide head FGH contains a 1×1 convolutional layer to reduce the feature dimension to 32 channels. Then, through a cross-layer fusion path, the high-level features are upsampled by 2 times by bilinear interpolation and then superimposed element-wise with the low-level features. Feature upsampling superimposed to →The result is then upsampled and superimposed. → Repeat until The final output is a 64-channel polarization semantic feature map, including:

[0161] Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ;

[0162] Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ;

[0163] Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ;

[0164] Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain That is, polarization semantic features .

[0165] (2) Cross-modal attention fusion module (CAFM)

[0166] refer to Figure 6 As shown, CAFM is responsible for fusing image and linear polarization degree image features at each stage. The fused image features include chromaticity distribution and texture patterns, while the linear polarization degree image features characterize the physical properties of polarization through the intensity difference derived from the Stokes vector. To address the key challenge of leveraging cross-sensor complementarity in RGB-PUSOD (RGB-Polarization underwater salient object detection) tasks involving fused images and polarization, CAFM employs a dual-attention fusion mechanism, combining spatial attention (SAA) with channel attention. It processes the input linear polarization degree image features and corresponding level fused image features through spatially alternating attention units (SAA); for example... Figure 7 and Figure 8 As shown, SAA employs a row-column separation computation path. First, it calculates the row-direction global correlation matrix of the feature map to generate row attention weights, which are then applied to the input. Next, it calculates the column-direction correlation matrix of the output features to generate column attention weights for secondary optimization. Subsequently, the weighted linear polarization degree image features / fused image features are concatenated along the channel dimension, and channel attention units (global pooling → fully connected layer → Sigmoid activation) generate channel weights for feature recalibration. Finally, the fused features are output through residual connections. Firstly, using... Multiply To determine common activation regions, spatial attention operations are then performed on the multiplied features to obtain a shared attention map. Finally, the shared attention map is used to... and We perform weighted analysis and then connect the weighted data along the channels to obtain the fused graph:

[0167] ,

[0168] in, This indicates that the cross-modal attention fusion module is in the first... Hierarchical output, Indicates a connection operation. Indicates element-wise multiplication. Note the operations that require attention in the representation space. Indicates the first The output of the layer D1 encoder, Indicates the first The output of layer R1 encoder; to avoid noise caused by suboptimal multimodal fusion, then... Residual operations were performed along the channel dimension. and fusion of image features Linear polarization degree image features The components are spliced ​​together along the channel dimension.

[0169] After splicing, using The filter and ReLU layer are used as the first noise removal operation. An additional layer is added after channel attention. The filter and ReLU layer optimize the final result. Finally, the RGB functionality is further incorporated into the result through residual connections and output.

[0170] To model the full-image dependency on local feature representations with lightweight computation and memory, an innovative Spatial Alternating Attention (SAA) module, which alternates between row and column attention, was designed to balance performance and resources. Figure 7 As shown, the SAA module collects global spatial information sequentially in the horizontal and vertical directions to enhance feature representation. Compared to global spatial attention dependencies, the SAA module significantly reduces the computational resources required.

[0171] The SAA module is equipped with two loops that can extract full-image contextual information from all pixels, thereby generating new features with dense and rich contextual information.

[0172] The row and column attention implementation in the SAA module follows the same approach. For example... Figure 8 As shown, taking the column attention unit as an example, given a local feature map This module first in Two applications The filter generates two feature maps Q and K, and Q and K are resized to... In tensor Every position in space A matrix can be obtained. Similarly, this can also be applied to tensors. Obtain the matrix , Through the above operations, we can obtain the basic matrix operation definition for generating attention maps.

[0173] ,

[0174] in, Representation matrix sum matrix The degree of correlation, that is, the degree of correlation between columns in two dimensions. Tensor operations are defined using batch matrix multiplication, as follows:

[0175] ,

[0176] in, tensor and tensor The degree of correlation between columns in the spatial dimension, that is, the degree of correlation between columns in the three-dimensional features.

[0177] Then, apply to the last dimension of the tensor Energy. Layers are used to generate column attention graphs Applying another to M Filter generation features Batch matrix multiplication is used to transform the features With features Multiplication yields the features enhanced by column attention. Finally, a residual structure is added to avoid the vanishing gradient network degradation problem.

[0178] .

[0179] (3) Cross-level polarization guided enhancement unit CPGM

[0180] The multimodal attention fusion module (CAFM) is responsible for fusing image features and linear polarization degree image features at each stage. Indeed, the fused image features contain a wealth of color and texture information, while the linear polarization degree image features focus on the differences in polarization strength. Fully leveraging the complementary information between the two and effectively utilizing fused image features and linear polarization degree image features for feature fusion is a crucial issue in the RGB-PSOD task. To address this problem, a cross-level polarization guided enhancement unit (CPGM) is designed, combining spatial attention with channel attention to achieve effective cross-modal fusion of fused image features and linear polarization degree image features.

[0181] refer to Figure 9 As shown, the cross-layer polarization-guided enhancement module (CPGM) receives three inputs: the current layer CAFM fused features, the 64-channel polarization semantic features output by PGH, and the previous layer CPGU enhancement features (R1 layer initialized with CAFM output). First, the polarization semantic features and the previous layer enhancement features are activated by Sigmoid to generate a spatial guidance mask. Simultaneously, the feature map is decomposed to the horizontal / vertical directions using a coordinate attention mechanism, and direction-sensitive spatial attention calculations are performed separately. The spatial guidance mask and the directional attention map are multiplied element-wise and fused back to the original features. Finally, noise-suppressed enhancement features are output through channel attention. The inputs to this module are: , and .use and Enhanced guidance .because Includes more texture and appearance information, while Stable polarization guidance information is incorporated, and these are blended together to generate guidance features that include both texture and polarization guidance information. First, a sigmoid activation function is used to obtain... The probability distributions of P and P are then multiplied element-wise by each other's probability distributions. Finally, the results of the multiplications are summed element-wise to obtain the linear polarization degree image features and characteristics. The guiding characteristics determined by the joint fusion The operating formula is as follows:

[0182] ,

[0183] in, This indicates that the cross-level polarization guidance enhancement module is in the first stage. Hierarchical output, This represents the activation function. Indicates the first Hierarchical fusion features.

[0184] To fully utilize the polarization guidance information of the layer preceding RGB To enhance information and save computational costs, feature maps are first separated along the H and W directions. Then add row and column attention layers respectively. Simultaneously, these maps are compared with the feature maps to generate coordinate attention maps. Multiplying them yields the feature enhancement results in the horizontal and vertical directions. Finally, they are added element-wise via residual joins to obtain the final result. Using this method, the dependence of spatial orientation in the feature map can be captured, providing... Add guidance and enhanced information.

[0185] (4) LawinASPP decoder

[0186] The LawinASPP decoder receives features at 5 levels, namely Here, we will follow the LawinASPP method. , , By fusing using LawinASPP, we obtain Low-level features are extracted by upsampling and concatenation. The obtained Low-level information enhancement facilitates the segmentation of the target structure; the LawinASPP decoder can be represented as:

[0187] ,

[0188] in, This represents the output of the LawinASPP decoder. Indicates high-level characteristics, Indicates low-level features, This represents the large window attention space pyramid pooling operation. This represents the convolution operation. Indicates an upsampling operation. This indicates a cross-layer attention operation.

[0189] (5) Loss Function

[0190] The total loss function of the salient object detection network is shown below:

[0191] ,

[0192] in, This represents the total loss of the salient object detection network. Represents the binary cross-entropy loss. Indicates the crossover and union ratio loss;

[0193] The cross-union loss It can be obtained through the following formula:

[0194] ,

[0195] in, Indicates the width of the underwater binocular polarization image. Indicates the height of the underwater binocular polarization image. Indicates the coordinates of the detected target area. Indicates the actual target area coordinates;

[0196] The binary cross-entropy loss It can be obtained through the following formula:

[0197] ,

[0198] in, Represents a logarithmic function.

[0199] The method of using a salient object detection network for object detection includes:

[0200] The linear polarization images are input into a five-layer D1 encoder to obtain... , , , and ;

[0201] Will , , , and The polarization semantic features are extracted by inputting them into the polarization guide head (PGH) to obtain the polarization semantic features. ;

[0202] The fused image is input into the first layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion to obtain the first-level fused features. ;

[0203] The first level of fusion features The input is fed into the second layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features and the first-level fused features are then combined. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the second-level fused features. ;

[0204] The second-level fusion features The input is fed into the third layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are combined with the second-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the third-level fused features. ;

[0205] The third-level fusion feature The input is fed into the fourth layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are then combined with the third-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the fourth-level fused features. ;

[0206] The fourth-level fusion feature The input is fed into the fifth layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are then combined with the fourth-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the fifth-level fused features. ;

[0207] The first level of fusion features Second-level fusion features Third-level fusion features Fourth-level fusion features and fifth-level fusion features The input is fed into the LauinASPP decoder for target detection, resulting in a target region mask.

[0208] like Figure 10 As shown, in both low and high turbidity environments, the method proposed in this embodiment demonstrates better results in images compared to other comparison algorithms. In particular, it provides more reliable and accurate detection results when handling low-contrast, high-turbidity scenes, offering an effective solution for underwater target detection.

[0209] In addition to subjective qualitative analysis of each method from a visual perspective, more objective evaluation metrics are needed to measure the effectiveness of each method. For all test images (including images with all turbidity levels), the performance of each method on these metrics should be calculated. , , as well as The values ​​are shown in Table 1. Bold text indicates that the method performs best on this metric, while underlined text indicates that the method performs second best. It can be seen that the method proposed in this embodiment achieves optimal performance on all salient target detection evaluation metrics. Specifically, the method proposed in this embodiment... The index reached 0.965, demonstrating the method's significant advantage in accuracy for salient target detection; The index reached 0.938, indicating that the method performs excellently in terms of coverage and accuracy in the target area; The M3Net performance achieved an accuracy of 0.968, further demonstrating its superior performance in balancing precision and recall. Finally, in terms of the MAE (Mean Effect) metric, the method proposed in this embodiment demonstrated its superior measurement accuracy with a minimum error of 0.006. Furthermore, M3Net also achieved excellent results in other metrics. The performance of SAM2-UNet is comparable to that of the proposed method. M3Net's performance on [the platform] The performance of BASNet and M3Net and their differences Its performance is second only to the proposed method.

[0210] Table 1 Comparison of the performance of different methods on the evaluation index of saliency target detection.

[0211] In step S4, features are extracted in the salient region using SuperPoint, and SuperGlue is used for matching. After algorithm optimization and constraint verification, reliable matching point pairs are obtained.

[0212] In step S4, traditional image patch-based feature matching algorithms suffer from inaccurate feature point location detection. Some methods employ a network training approach, first detecting feature points and then calculating descriptors. However, this separate training model for feature extraction and descriptor generation not only wastes computational resources but also results in a less streamlined network structure and poor real-time performance. To address this issue, SuperPoint pioneered the simultaneous extraction of feature points and descriptors within a single self-supervised network.

[0213] like Figure 11 As shown, SuperPoint's network design employs an encoder-decoder architecture. The network shares a single encoder to process and reduce the dimensionality of the input image, and then learns task-specific weights through two decoders. One decoder is responsible for feature point detection, and the other is responsible for generating feature descriptors. The decoders learn different network parameters depending on the task. Most network parameters are shared between the two tasks.

[0214] Specifically, the encoder is based on the VGG network architecture and achieves feature dimensionality reduction through stacked convolutional layers. The input image sequentially passes through four feature processing blocks, each containing two convolutional layers and one max-pooling layer, but the pooling operation is omitted in the last block. The InterestPointDecoder performs two convolution operations on the output of the shared encoder, then processes it by channel, removing redundant dimensions before reshaping. The DescriptorDecoder performs convolution decoding on the output of the shared encoder, enlarges its size through bicubic interpolation, and finally performs L2 normalization to normalize the descriptor of each pixel to a unit length.

[0215] In step S5, traditional feature matching methods suffer from insufficient accuracy and robustness in complex scenarios such as changes in illumination, differences in viewpoint, and occlusion. To address the limitations of traditional feature matching, SuperGlue simulates human feature matching, offering a novel perspective on the image feature matching problem. Specifically, it uses an attention-based graph neural network and an optimal transfer method to match feature points. SuperGlue defines the matching task as a graph matching problem involving two sets of sparse image features. By learning the complex relationships between feature points through the network, it can obtain more accurate matching results.

[0216] like Figure 12 As shown, SuperPoint can extract high-quality feature points from images, while SuperGlue can improve the accuracy of feature matching. By leveraging their complementary strengths in feature point detection and feature matching, the combination of the two can achieve more robust and accurate feature point detection and matching.

[0217] Therefore, as Figure 13 As shown, two images to be matched are input. First, the SuperPoint network is used to extract high-quality feature points and feature descriptors. Then, based on these two sets of feature data, the SuperGlue network is used to achieve accurate matching between the two images and output the corresponding matching relationship.

[0218] While the SuperPoint and SuperGlue combination matching method fully leverages the advantages of deep learning technology to improve the accuracy and robustness of image feature extraction and matching, there is still room for improvement in terms of computational complexity due to the inherent complexity of the network itself. It is worth noting that salient object detection possesses certain characteristics; restricting the matching region to salient areas of underwater images can not only reduce computational cost but also improve the robustness and reliability of the matching. For example... Figure 14 As shown, sparse feature matching is performed only in salient regions to obtain reliable feature matching point pairs. and .

[0219] In step S6, based on the dual-target positioning and the feature matching points described in S5, the three-dimensional coordinates, size, and distance of the target are calculated to complete the measurement.

[0220] Based on the camera parameters obtained from binocular calibration, and combined with the principle of binocular triangulation, the three-dimensional coordinates of the matching points can be calculated, thereby constructing the target three-dimensional point cloud of coefficients.

[0221] Dimensional measurement is achieved on the image plane Directional and distance measurements enable camera optical axis measurements. Measurement in direction. For example... Figure 15As shown, the specific steps of the method are as follows:

[0222] (1) Target detection results using the left eye image Construct the minimum bounding rectangle within the target segmentation region, and calculate the two-dimensional coordinates of the four midpoints of the rectangle clockwise. .

[0223] (2) Traverse the set of matching feature points, calculate the distance between the feature points in the left image and the midpoint, and find the four points closest to the four midpoints. Their two-dimensional coordinates are: .

[0224] (3) In sparse target 3D point cloud In the middle, find with Corresponding 3D point coordinates This serves as a key point for calculating the target size.

[0225] (4) Midpoint of the rectangle The point closest to it Since the two sets of points are only very close, there is still some deviation. Direct calculation and distance, and Using distance as the target's size is inaccurate. To calculate the target's size more accurately, first calculate the direction of one set of opposite sides of the rectangle as principal axis 1, and the direction of the other set of opposite sides as principal axis 2, then calculate... and Projection distance on main axis 1 and The projection distance on main axis 2 is used as the final size information of the target.

[0226] (5) Since most targets are planar, the 3D point cloud of sparse targets is calculated. of The average coordinates can be used to preliminarily determine the distance between the underwater target and the camera.

[0227] In summary, the underwater target measurement method based on saliency region sparse matching proposed in this embodiment acquires images at different polarization angles using a polarization camera, and calculates the intensity image (S0) and linear polarization degree image (DoLP) as fusion source images. Secondly, it proposes an underwater polarization image target detection method based on an interactive enhancement paradigm, fusing complementary information from the intensity image and the linear polarization degree image through a polarization image fusion network (PIFNet), and using a saliency target detection network to locate the target region. Next, based on the saliency detection results, sparse feature matching is restricted to salient regions, and efficient and robust feature point matching is achieved by combining SuperPoint and SuperGlue. Finally, based on the matched point pairs and camera parameters, the three-dimensional coordinates of the target are calculated, achieving accurate measurement of size (width, height) and distance. This method provides efficient and reliable technical support for underwater target detection and three-dimensional measurement, and can be applied to fields such as marine engineering and underwater robotics.

[0228] Example 2:

[0229] Based on the same inventive concept as Embodiment 1, this embodiment of the invention discloses an underwater target measurement system with sparse matching of significant regions, comprising:

[0230] The Stokes vector calculation module is used to: calculate intensity images and linear polarization images based on Stokes vectors from the acquired underwater binocular polarization images;

[0231] The image fusion module is used to: perform image fusion using a polarization image fusion network based on the intensity image and the linear polarization image to obtain a fused image;

[0232] A salient target detection module is used to: perform target detection using a salient target detection network based on the linear polarization degree image and the fused image to obtain a target region mask;

[0233] The SuperPoint module is used to: extract feature points within the target area of ​​the underwater binocular polarization image and generate feature descriptors based on the target area mask using the SuperPoint network;

[0234] The image matching module is used to: perform binocular image matching through the SuperGlue network based on the feature points and feature descriptors, and output matching point pairs;

[0235] The distance calculation module is used to calculate the distance from the target to the polarization binocular camera based on the matching point pair and the polarization binocular camera parameters.

[0236] The specific functions of each module are described in the relevant content of Implementation Example 1, and will not be repeated here.

[0237] Example 3:

[0238] This embodiment provides a computer device, including:

[0239] Memory, used to store computer instructions;

[0240] A processor is configured to execute the computer instructions to implement the steps of the underwater target measurement method for sparse matching of significant regions as described in Embodiment 1.

[0241] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0242] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0243] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0244] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0245] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for measuring underwater targets using sparse matching in salient regions, characterized in that, include: Based on the acquired underwater binocular polarization images, intensity images and linear polarization images are calculated using Stokes vectors. Based on the intensity image and the linear polarization degree image, an image fusion network is used to perform image fusion to obtain a fused image. Based on the linear polarization degree image and the fused image, a salient target detection network is used to perform target detection and obtain the target region mask; Based on the target region mask, feature points within the target region of the underwater binocular polarization image are extracted using the SuperPoint network and feature descriptors are generated. Based on the feature points and feature descriptors, stereo image matching is performed using the SuperGlue network to output matching point pairs. Based on the matching point pairs and the polarization binocular camera parameters, the distance from the target to the polarization binocular camera is calculated.

2. The underwater target measurement method using sparse matching of significant regions according to claim 1, characterized in that, The intensity image is obtained using the following formula: ， in, Represents an intensity image. This represents an underwater binocular polarization image with a polarization angle of 0 degrees. This represents an underwater binocular polarization image with a polarization angle of 45 degrees. This represents an underwater binocular polarization image with a polarization angle of 90 degrees. This represents an underwater binocular polarization image with a polarization angle of 135 degrees. The linear polarization degree image is obtained by the following formula: ， in, Represents the linear polarization degree image. This represents the first linearly polarized light. This indicates the second linearly polarized light.

3. The underwater target measurement method using sparse matching of salient regions according to claim 1, characterized in that, The polarization image fusion network includes a first encoder module, a second encoder module, a fusion module, and a decoder module; The first encoder module extracts the brightness features of the intensity image through grouped spatial convolution and dense blocks; The second encoder module extracts the brightness features of the linear polarization image through grouped spatial convolution and dense blocks; The first encoder module and the second encoder module share weights; The fusion module adaptively fuses the brightness features of the intensity image and polarization image through channel attention and spatial attention mechanisms to obtain fused features; The decoder outputs a fused image by reconstructing the fusion features and combining them with the chromaticity information of the intensity image; If the intensity image and linear polarization image are color images, perform the following steps: The intensity image and linear polarization image are converted from the RGB color space to the YCrCb color space, and then input into the first encoder module and the second encoder module respectively. The fused image is converted from the YCrCb color space to the RGB color space.

4. The underwater target measurement method using sparse matching of salient regions according to claim 1, characterized in that, The total loss function of the polarization image fusion network is as follows: ， in, This represents the total loss of the polarization image fusion network. The weighting parameters represent the relationship between multi-scale weighted structural similarity loss and polarization intensity loss. This represents the multi-scale weighted structural similarity loss. Indicates the loss of polarization intensity; The multi-scale weighted structural similarity loss It can be obtained through the following formula: ， in, Indicates the weighting coefficient. Represents the linear polarization degree image. Represents an intensity image. Indicates a fused image. Represents a window. Indicates size is window Down and Local similarity between them Indicates size is window Down and Local similarity between them; Local similarity It can be obtained through the following formula: ， in, Display window Internal belonging to Part of express The mean, Display window Internal belonging to Part of express The mean, express covariance, express covariance, express and covariance, and Indicates a constant; Local similarity It can be obtained through the following formula: ， in, Display window Internal belonging to Part of express The mean, express covariance, express and covariance; The weighting coefficient It can be obtained through the following formula: ， in, , Used to enhance weighting coefficients robustness and stability; The polarization intensity loss It can be obtained through the following formula: ， in, express Norm, and As a weighting factor, , This represents the natural exponential function. A coefficient representing the difference between pixel values ​​in a scaled image. , Indicates the image height. Indicates the image width.

5. The underwater target measurement method using sparse matching of salient regions according to claim 1, characterized in that, The salient target detection network includes a dual encoder, a polarization guide head (FGH), a cross-modal attention fusion module (CAFM), a cross-level polarization guidance enhancement module (CPGM), and a LawinASPP decoder. The dual encoders adopt a ResNet18 backbone network, including a five-layer D1 encoder and a five-layer R1 encoder; The method of using a salient object detection network for object detection includes: The linear polarization images are input into a five-layer D1 encoder to obtain... , , , and ; Will , , , and The polarization semantic features are extracted by inputting them into the polarization guide head (PGH) to obtain the polarization semantic features. ; The fused image is input into the first layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion to obtain the first-level fused features. ; The first level of fusion features The input is fed into the second layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features and the first-level fused features are then combined. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the second-level fused features. ; The second-level fusion features The input is fed into the third layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are combined with the second-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the third-level fused features. ; The third-level fusion feature The input is fed into the fourth layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are then combined with the third-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the fourth-level fused features. ; The fourth-level fusion feature The input is fed into the fifth layer R1 encoder to obtain... ;Will and The input is fed into the cross-modal attention fusion module (CAFM) for feature fusion, and the fused features are then combined with the fourth-level fused features. and polarization semantic features The input is fed into the cross-level polarization guided enhancement module (CPGM) to obtain the fifth-level fused features. ; The first level of fusion features Second-level fusion features Third-level fusion features Fourth-level fusion features and fifth-level fusion features The input is fed into the LauinASPP decoder for target detection, resulting in a target region mask.

6. The underwater target measurement method using sparse matching of salient regions according to claim 5, characterized in that, The , , , and The polarization semantic features are extracted by inputting them into the polarization guide head (PGH) to obtain the polarization semantic features. ,include: Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ; Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ; Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain ; Through a 1×1 convolutional layer After reducing the dimensionality to 32 channels, upsampling is performed, and the sampling results are compared with... By piecing them together, we obtain That is, polarization semantic features .

7. The underwater target measurement method with sparse matching of significant regions according to claim 5, characterized in that, The expression for the cross-modal attention fusion module is as follows: ， in, This indicates that the cross-modal attention fusion module is in the first... Hierarchical output, Indicates a connection operation. Indicates element-wise multiplication. Note the operations that require attention in the representation space. Indicates the first The output of the layer D1 encoder, Indicates the first The output of layer R1 encoder; The expression for the cross-level polarization guidance enhancement module is as follows: ， in, This indicates that the cross-level polarization guidance enhancement module is in the first stage. Hierarchical output, This represents the activation function. Indicates the first Hierarchical fusion features; The expression for the LauinASPP decoder is as follows: ， in, This represents the output of the LawinASPP decoder. Indicates high-level characteristics, Indicates low-level features, This represents the large window attention space pyramid pooling operation. This represents the convolution operation. Indicates an upsampling operation. This indicates a cross-layer attention operation.

8. The underwater target measurement method using sparse matching of significant regions according to claim 5, characterized in that, The total loss function of the salient object detection network is shown below: ， in, This represents the total loss of the salient object detection network. Represents the binary cross-entropy loss. Indicates the crossover and union ratio loss; The cross-union loss It can be obtained through the following formula: ， in, Indicates the width of the underwater binocular polarization image. Indicates the height of the underwater binocular polarization image. Indicates the coordinates of the detected target area. Indicates the actual target area coordinates; The binary cross-entropy loss It can be obtained through the following formula: ， in, Represents a logarithmic function.

9. A sparse matching underwater target measurement system for a significant region, characterized in that, include: The Stokes vector calculation module is used to: calculate intensity images and linear polarization images based on Stokes vectors from the acquired underwater binocular polarization images; The image fusion module is used to: perform image fusion using a polarization image fusion network based on the intensity image and the linear polarization image to obtain a fused image; A salient target detection module is used to: perform target detection using a salient target detection network based on the linear polarization degree image and the fused image to obtain a target region mask; The SuperPoint module is used to: extract feature points within the target area of ​​the underwater binocular polarization image and generate feature descriptors based on the target area mask using the SuperPoint network; The image matching module is used to: perform binocular image matching through the SuperGlue network based on the feature points and feature descriptors, and output matching point pairs; The distance calculation module is used to calculate the distance from the target to the polarization binocular camera based on the matching point pair and the polarization binocular camera parameters.

10. A computer device, characterized in that, include: Memory, used to store computer instructions; A processor for executing the computer instructions to implement the steps of the underwater target measurement method for sparse matching of significant regions as described in any one of claims 1-8.

Images