Underwater structure damage identification method and system based on optical polarization imaging

Abstract

The invention provides an underwater structure damage identification method and system based on optical polarization imaging, relates to the field of deep learning, and solves the technical problem of low crack detection precision caused by serious scattering and lack of physical prior guidance in a high-turbidity underwater environment in the prior art. The method comprises the following steps: acquiring light intensity images of an underwater target area at a plurality of preset polarization angles; calculating a multi-dimensional physical quantity representing a polarization state based on the light intensity image, and constructing multi-channel input data including the multi-dimensional physical quantity; inputting the multi-channel input data into a deep learning detection model, and outputting damage information; a risk level is assessed based on the damage information. The method is used in an underwater structure damage identification process.

Description

Technical Field

[0001] This application relates to the field of deep learning, and in particular to a method and system for underwater structure damage identification based on optical polarization imaging. Background Technology

[0002] In hydraulic engineering, underwater structures such as dams are subjected to long-term scouring, corrosion, and stress from water flow, making them prone to sub-millimeter-level cracks. Failure to detect these cracks in time can lead to leakage or even dam failure. Traditional detection methods rely on manual diving or sonar, which suffer from low efficiency, high subjectivity, high safety risks, and extremely low signal-to-noise ratios in turbid water with turbidity >100 NTU. While recent research has applied deep learning, such as the YOLO series, or polarization imaging to underwater detection, it remains difficult to distinguish between genuine cracks and interference from algae, sediment, etc. Therefore, this application provides a method and system for underwater structural damage identification based on optical polarization imaging. Summary of the Invention

[0003] This application provides an underwater structure damage identification method based on optical polarization imaging, which solves the technical problem of low crack detection accuracy in high-turbidity underwater environments due to severe scattering and lack of physical prior guidance. This application also provides an underwater structure damage identification system based on optical polarization imaging.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for underwater structure damage identification based on optical polarization imaging is provided, including: Acquire light intensity images of underwater target areas at multiple preset polarization angles; Based on the light intensity image, multidimensional physical quantities characterizing the polarization state are calculated, and multi-channel input data containing the multidimensional physical quantities is constructed; the multidimensional physical quantities include Stokes vectors and polarization difference images; The multi-channel input data is input into a deep learning detection model, which is obtained by fusing a physical attention mechanism in the YOLOv7-AC architecture. The physical attention mechanism is an attention generation mechanism driven by external polarization imaging physical quantities to guide the deep learning detection model to focus on structural damage-sensitive regions. Spatial modulation of intermediate feature maps is performed based on physical attention and global attention mechanisms; the intermediate feature maps are multi-channel feature tensors generated by deep learning detection models in backbone networks or feature pyramid networks. Based on the spatially modulated intermediate feature map, the deep learning detection model outputs damage information, and the risk level is assessed based on the damage information; the damage information includes damage location and damage parameters.

[0005] Based on the above technical solutions, in the underwater structure damage identification method based on optical polarization imaging provided in this application, traditional visual methods are difficult to effectively detect tiny cracks on the structure surface in high-turbidity underwater environments. This is mainly due to the limitations of strong scattering, low contrast, and background interference, which cause damage features to be submerged. AI models are easily misled by noise, resulting in high rates of missed detections and false alarms. To solve this problem, this application proposes an intelligent identification scheme that deeply integrates the physical characteristics of optical polarization with an improved deep learning architecture. This scheme first acquires light intensity images through multi-angle polarization imaging, calculates the Stokes vector and polarization difference image (PDI), and transforms the polarization information with clear physical meaning into multi-channel inputs. Then, it constructs a detection model based on YOLOv7-AC and introduces a physical attention mechanism, that is, it uses PDI to generate a spatial attention mask as an "optical prior" to guide the model to focus on the real damage area, realizing "seeing scattering but not seeing cracks". At the same time, it combines a global attention mechanism to perform dual-path spatial modulation on the intermediate feature map, taking into account both data adaptability and physical interpretability. Finally, it outputs damage information containing location and parameters, which is used to quantify the risk level. This technical solution overcomes the performance bottleneck of pure data-driven methods in harsh underwater environments, significantly improving detection accuracy, robustness, and engineering applicability, and providing a highly reliable and deployable intelligent monitoring method for the safe operation and maintenance of critical infrastructure such as dams and bridges.

[0006] In conjunction with the first aspect above, in one possible implementation, the physical attention mechanism generates a first spatial attention mask based on a polarization difference image, and uses the first spatial attention mask to modulate the intermediate feature map in the YOLOv7-AC architecture.

[0007] In conjunction with the first aspect above, in one possible implementation, the deep learning detection model includes a feature fusion module, a polarization feature enhancement module, and a multi-scale crack detection module; The feature fusion module is used to receive multi-channel input data and generate an initial feature map through convolutional layers, batch normalization, and nonlinear activation functions. The polarization feature enhancement module is used to introduce a global attention mechanism; the global attention mechanism includes a channel attention submodule and a spatial attention submodule. The multi-scale crack detection module is used to expand the structure of the backbone network or feature pyramid network and introduce a physically guided attention mechanism.

[0008] In conjunction with the first aspect above, in one possible implementation, the spatial modulation of the intermediate feature map based on the physical attention mechanism and the global attention mechanism includes: The physical attention mechanism generates a first spatial attention mask based on a polarization difference image calculated from a multi-angle polarization intensity image. The global attention mechanism generates a second spatial attention mask based on the intermediate feature map itself by modeling the dependency between channels and spatial dimensions. The deep learning detection model further includes a learnable fusion module, which is used to weightedly fuse the first spatial attention mask and the second spatial attention mask to obtain a comprehensive attention mask, and then use the comprehensive attention mask to perform element-wise multiplication with the intermediate feature map to spatially modulate the intermediate feature map.

[0009] In conjunction with the first aspect above, in one possible implementation, the assessment of risk level based on damage information includes: The equivalent length L and spatial location of the crack are extracted from the damage information. In conjunction with material degradation patterns, stress concentration effects, and prior structural sensitivity, a comprehensive risk index (IRI) is calculated, defined as follows: ; in, The reference length is given by m, the material degradation coefficient (m>1), and β is the proportionality coefficient. [0,1], For the spatial orientation angle of the damage Under the material brittleness parameter k, the stress concentration factor in the damaged region , The length of half a crack. , The radius of curvature of the crack tip is . The principal stress direction For structural sensitivity weighting fields; Mapping IRIs to discrete risk levels: ; in, , For a preset threshold, and < .

[0010] In conjunction with the first aspect above, in one possible implementation, the structural sensitivity weight field includes: The structural sensitivity weight field is composed of prior knowledge, and its expression is: ; in, For spatial points The distance to the nearest critical component, which includes load-bearing beams, anchorage zones, or stress concentration areas; λ is the attenuation coefficient, used to characterize the rate at which structural sensitivity decreases with distance.

[0011] In conjunction with the first aspect described above, in one possible implementation, the method for generating the first spatial attention mask includes:

[0012] The polarization difference image is processed through at least one lightweight convolutional layer, and a first spatial attention mask belonging to the interval [0,1] is generated using the Sigmoid activation function.

[0013] In conjunction with the first aspect above, in one possible implementation, the method for generating the second spatial attention mask includes: The channel attention submodule performs global pooling on the intermediate feature map and processes it through a multilayer perceptron to generate channel weights, and then performs channel dimension weighting on the intermediate feature map based on the channel weights. The spatial attention submodule aggregates the intermediate feature map by channel dimension, performs spatial convolution with a large convolution kernel to generate spatial weights, and then performs spatial dimension weighting on the channel-weighted feature map based on the spatial weights. A second attention mask is generated by weighting feature maps based on channel dimension and spatial dimension.

[0014] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the multidimensional physical quantity includes: Based on the light intensity image, calculate the Stokes vector, including the total light intensity. Horizontal-vertical polarization difference polarization difference Right-handed to left-handed circular polarization difference ; Based on the Stokes vector, using the formula PDI= The polarization difference image (PDI) is calculated.

[0015] Secondly, this application provides an underwater structure damage identification system based on optical polarization imaging, comprising: an acquisition module, a detection module, and an evaluation module; wherein, the acquisition module is used to acquire light intensity images of an underwater target area at multiple preset polarization angles; calculate multidimensional physical quantities characterizing the polarization state based on the light intensity images, and construct multi-channel input data containing the multidimensional physical quantities; the detection module is used to input the multi-channel input data into a deep learning detection model, the deep learning detection model being obtained based on the YOLOv7-AC architecture with a fused physical attention mechanism; spatially modulate the intermediate feature map based on the physical attention mechanism and the global attention mechanism; the evaluation module is used to output damage information from the deep learning detection model based on the spatially modulated intermediate feature map, and evaluate the risk level based on the damage information.

[0016] This application provides a method and system for underwater structural damage identification based on optical polarization imaging. This method can efficiently and accurately identify minute cracks and other damage on the surface of concrete or steel structures in complex underwater environments such as high turbidity and low light conditions, and achieve automated quantitative assessment of damage location, geometric parameters, and risk level. Specifically, the method acquires light intensity images of underwater targets through multi-angle polarization imaging, and extracts physically meaningful polarization features by combining Stokes vectors and polarization difference images (PDI), effectively suppressing water scattering noise and enhancing surface details. Furthermore, PDI is used as a physical prior to drive the physical attention mechanism in a deep learning model, co-modulating feature representation with the global attention mechanism, significantly improving the ability to perceive real damage areas. Finally, based on the detection results, the method integrates material degradation patterns, stress concentration effects, and structural sensitivity priors to calculate the Integrated Risk Index (IRI), achieving intelligent monitoring throughout the entire process from "visible" to "accurate judgment and scientific assessment." This system can be deployed on remotely operated vehicles (ROVs) or autonomous underwater vehicles (AUVs), and is suitable for regular inspections, hazard identification, and emergency response of key water conservancy and marine infrastructure such as dams, bridge piers, subsea tunnels, and nuclear power plant cooling towers. It significantly reduces the safety risks and costs of manual diving inspections, improves the intelligence, standardization, and resilience of infrastructure operation and maintenance, and enables accurate detection of water bodies to determine whether turbidity is present, and performs measurements based on the degree of turbidity.

[0017] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0018] Figure 1 A system architecture diagram of an underwater structure damage identification system based on optical polarization imaging is provided for embodiments of this application; Figure 2 A flowchart illustrating an underwater structure damage identification method based on optical polarization imaging, provided in an embodiment of this application; Figure 3This is a schematic diagram of the detection process of a deep learning detection model provided in an embodiment of this application. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The underwater structure damage identification method based on optical polarization imaging provided in this application embodiment can be applied to an underwater structure damage identification system based on optical polarization imaging, such as... Figure 1 As shown, the system includes an acquisition module, a detection module, and an evaluation module; The acquisition module is used to acquire light intensity images of underwater target areas at multiple preset polarization angles; calculate multidimensional physical quantities characterizing polarization state based on the light intensity images, and construct multi-channel input data containing multidimensional physical quantities; The detection module is used to input multi-channel input data into the deep learning detection model, which is based on the YOLOv7-AC architecture and incorporates the physical attention mechanism. The deep learning detection model performs spatial modulation on the intermediate feature map based on the physical attention mechanism and the global attention mechanism. The assessment module is used to output damage information from a deep learning detection model based on spatially modulated intermediate feature maps, and to assess the risk level based on the damage information.

[0021] To address the technical problem of low crack detection accuracy in high-turbidity underwater environments due to severe scattering and lack of physical prior guidance, this application provides an underwater structural damage identification method based on optical polarization imaging. The method includes: acquiring light intensity images of an underwater target area at multiple preset polarization angles; calculating multidimensional physical quantities characterizing the polarization state based on the light intensity images and constructing multi-channel input data containing these multidimensional physical quantities; inputting the multi-channel input data into a deep learning detection model, which is based on the YOLOv7-AC architecture and incorporates a physical attention mechanism; spatially modulating intermediate feature maps based on the physical attention mechanism and a global attention mechanism; outputting damage information based on the spatially modulated intermediate feature maps; and assessing the risk level based on the damage information. By deeply coupling the Stokes vector physical properties of polarization imaging with the improved global attention mechanism of the YOLOv7-AC model, this application achieves high-precision, adaptive extraction and multi-scale identification of underwater dam cracks in turbid water environments, applicable to dam safety maintenance, hazard investigation, emergency response, and structural health monitoring scenarios.

[0022] like Figure 2 As shown in the figure, an underwater structure damage identification method based on optical polarization imaging provided in this application includes: S201. Acquire light intensity images of the underwater target area at multiple preset polarization angles.

[0023] In some implementations, a remotely operated vehicle (ROV) inspects the underwater section of the dam along a gridded path (0.3m spacing) at a speed of 2m / s. A polarization camera acquires multi-angle polarization images at 40FPS. This application acquires images at four angles: 0°, 45°, 90°, and 135°. There are typically two implementation methods: Single sensor + rotating polarizer: Every 4 frames form a complete polarization observation (actual effective sampling rate ≈ 10 groups / second). Multi-sensor / splitting plane polarization camera: can acquire 4 angles simultaneously in a single frame (at this time 40FPS=40 sets / second).

[0024] In this application, polarization image acquisition can be achieved in two ways: First, a single sensor combined with a rotating polarizer offers low cost and high resolution, but requires four frames to synthesize a single set of polarization data, resulting in an effective sampling rate of approximately 10 sets / second. Second, a split-focus plane polarization camera (such as a multi-sensor integrated solution) can simultaneously acquire images at four polarization angles (0°, 45°, 90°, and 135°) within a single frame, achieving real-time sampling of 40 sets / second. The latter is particularly suitable for scenarios involving ROV high-speed inspections at 2 m / s, effectively avoiding multi-frame misalignment caused by motion, improving the accuracy of polarization information calculation and system robustness, and is more suitable for the high-reliability detection requirements in dynamic underwater environments.

[0025] S202. Calculate multidimensional physical quantities representing polarization state based on light intensity images, and construct multi-channel input data containing multidimensional physical quantities.

[0026] Among them, the multidimensional physical quantities include Stokes vectors and polarization difference images.

[0027] It should be noted that constructing multi-channel input data containing the aforementioned multi-dimensional physical quantities refers to converting light intensity images at different polarization angles obtained from underwater target areas using optical polarization imaging technology into multi-dimensional physical quantities that reflect the polarization characteristics of the target object's surface, and integrating these physical quantities into a multi-channel data format for use as input to a deep learning model. Specific steps include: A series of light intensity images are acquired by photographing the target area at multiple preset polarization angles (such as 0°, 45°, 90°, and 135°). Then, a Stokes vector is calculated based on these light intensity images. This vector contains information such as total light intensity, horizontal-vertical polarization difference, diagonal polarization difference, and right-hand-left-hand circular polarization difference. Furthermore, a polarization difference image (PDI) is calculated based on the Stokes vector to quantify the degree of linear polarization at different locations.

[0028] The multidimensional physical quantities calculated above, including the components of the Stokes vector and the polarization difference image, are combined with the original light intensity image to form a multi-channel input data structure. For example, if an RGB image with polarization features is used, a 7-channel data structure may be formed (RGB image occupies 3 channels, Stokes vector (…)…) , , (Both PDI and PDI occupy one channel), enabling the deep learning model to simultaneously utilize color information, intensity information, and polarization characteristics for more accurate structural damage identification and analysis. This multi-channel input data is then fed into the deep learning detection model for training or prediction.

[0029] S203. Input the multi-channel input data into the deep learning detection model. The deep learning detection model is obtained by fusing the physical attention mechanism in the YOLOv7-AC architecture.

[0030] Among them, the physical attention mechanism is an attention generation mechanism driven by external polarization imaging physical quantities, used to guide deep learning detection models to focus on structural damage-sensitive regions.

[0031] The physical attention mechanism is based on polarization difference images, generates a first spatial attention mask, and uses the first spatial attention mask to modulate the intermediate feature map in the YOLOv7-AC architecture; It should be noted that the essence of the physical attention mechanism is to transform the optical physics (PDI) revealed by polarization imaging into a learnable attention signal, which is then injected into the deep learning model as a "physical prior" to achieve "seeing scattering and seeing cracks that cannot be seen".

[0032] S204. The deep learning detection model spatially modulates the intermediate feature map based on the physical attention mechanism and the global attention mechanism; the intermediate feature map is a multi-channel feature tensor generated by the deep learning detection model in the backbone network or feature pyramid network.

[0033] S205. Based on the spatially modulated intermediate feature map, the deep learning detection model outputs damage information, and the risk level is assessed based on the damage information.

[0034] The damage information includes the location of the damage and the damage parameters.

[0035] Based on the above technical solutions, the underwater structure damage identification method based on optical polarization imaging provided in this application addresses the challenge that traditional visual detection methods struggle to effectively identify minute cracks in high-turbidity underwater environments due to severe scattering and low contrast. Furthermore, existing deep learning models lack physical interpretability and are susceptible to noise interference, leading to high rates of missed detections and false alarms. To solve this core problem, this application proposes an integrated technical solution of "polarization imaging—physical attention—risk quantification." First, a polarization camera mounted on an ROV simultaneously acquires light intensity images at four angles (0°, 45°, 90°, and 135°) at 40FPS, ensuring complete polarization observation without motion artifacts even during high-speed inspections at 2m / s. Second, multi-dimensional physical quantities (such as PDI) are constructed based on Stokes vectors and injected as multi-channel inputs into the improved YOLOv7-AC model. The key innovation lies in introducing a physical attention mechanism, transforming PDI into a first-space attention mask to guide the model to focus on real damage areas with strong polarization responses, thus "using optical laws to tell the AI ​​where to look." Then, combined with a global attention mechanism for dual-path feature modulation, the sensitivity and positioning accuracy of micro- and multi-scale cracks are significantly improved. Finally, crack parameters are extracted from the detection results and the fracture mechanics and structural prior calculations are integrated to calculate the Integrated Risk Index (IRI), achieving a closed loop from "detection" to "assessment." This solution not only overcomes the bottleneck of imaging turbid water bodies but also achieves high-precision, robust, and interpretable intelligent monitoring, meeting the stringent requirements for safe operation and maintenance of critical infrastructure such as dams.

[0036] In one possible implementation of this application embodiment, the above-mentioned S202 can be specifically described as follows: Based on the light intensity image, calculate the Stokes vector, including the total light intensity. Horizontal-vertical polarization difference Diagonal polarization difference Right-handed to left-handed circular polarization difference ; in, , , , These represent the light intensities measured when the polarizer angles are 0°, 90°, 45°, and 135°, respectively. The intensity of right-handed circularly polarized light. This represents the intensity of left-handed circularly polarized light.

[0037] It should be noted that total light intensity represents the total energy of incident light and does not include polarization information; it is used for normalization or illumination compensation. The horizontal-vertical polarization difference indicates the dominance of linear polarization in the horizontal direction; if positive, the horizontal polarization is stronger. The diagonal polarization difference represents the polarization characteristics in the diagonal direction and determines the direction of the principal polarization axis. The right-handed to left-handed circular polarization difference indicates the direction of rotation of the light wave.

[0038] Based on the Stokes vector, using the formula PDI= The polarization difference image (PDI) is calculated.

[0039] It should be noted that a PDI image is a single-channel grayscale image. The higher the pixel value, the stronger the linear polarization of that point. In an underwater environment, scattered light (from the water body) is usually unpolarized or weakly polarized, resulting in a low PDI value. Reflected light from structural surfaces (such as the edges of concrete cracks) has strong polarization characteristics, leading to a high PDI value. Therefore, PDI can effectively enhance the features of structural surfaces, suppress background scattering noise, and significantly improve the visibility of micro-cracks.

[0040] Based on the above technical solutions, traditional methods are limited by underwater optical scattering, turbidity interference, and computational resource constraints, making it difficult to achieve sub-millimeter level crack accuracy detection. This application utilizes Stokes vectors (… , , , By deeply fusing multidimensional polarization information with the adaptive deep learning framework of YOLOv7-AC, a physically guided intelligent detection system for high-turbidity water environments was constructed. The total light intensity was calculated by acquiring light intensity images at four polarization angles (0°, 90°, 45°, and 135°). Horizontal-vertical polarization difference Diagonal polarization difference and circular polarization difference This process generates a polarization difference image (PDI). It fully utilizes the physical property that light retains its polarization characteristics when reflected at different media interfaces, effectively distinguishing between light reflected from structural surfaces (strong polarization) and light scattered from water (weak polarization), significantly improving the signal-to-noise ratio. As a key physical prior, PDI not only enhances the visibility of minute cracks but also provides interpretable and robust input features for subsequent deep learning models. This method overcomes the bottleneck of traditional visual inspection being "unclear and difficult to identify" in turbid water, achieving a paradigm shift from "data-driven" to "physical augmentation." Its results are not only applicable to water conservancy facilities such as dams and reservoirs but can also be extended to the intelligent monitoring of critical infrastructure such as underwater bridge structures and nuclear power plant cooling towers. It responds to the core requirements of the Sustainable Development Goals (SDG9) regarding improving infrastructure resilience and intelligent operation and maintenance, possessing broad technological promotion value and engineering application prospects.

[0041] In one possible implementation of this application embodiment, the above-mentioned S203 can be specifically described as follows, such as Figure 3 As shown; The deep learning detection model includes a feature fusion module, a polarization feature enhancement module, and a multi-scale crack detection module; The feature fusion module receives multi-channel input data and generates an initial feature map through convolutional layers, batch normalization, and non-linear activation functions.

[0042] For example, suppose the input is an RGB image containing a building surface and a corresponding polarization difference (PDI) image. The feature fusion module processes these two types of input separately and fuses them together through additional convolution operations to form a comprehensive feature representation.

[0043] The polarization feature enhancement module is used to introduce a global attention mechanism; the global attention mechanism includes a channel attention submodule and a spatial attention submodule.

[0044] It should be noted that when detecting cracks on the surface of an underwater object, channel attention may emphasize features that are particularly useful for identifying cracks (such as edge or texture information), while spatial attention helps to highlight the specific location where cracks may exist.

[0045] A multi-scale crack detection module is used to expand the structure of the backbone network or feature pyramid network and introduce a physically guided attention mechanism to enhance the feature response to potential structural damage areas.

[0046] It should be noted that, in order to effectively detect cracks of various sizes, the multi-scale crack detection module may combine deep features at low resolution (for large crack detection) and shallow features at high resolution (for small crack detection). Physically guided attention mechanisms can utilize physical quantities (such as the previously mentioned polarization information) to improve the response to real cracks and reduce false alarms.

[0047] Based on the above technical solutions, traditional visual methods struggle to effectively detect minute cracks in hydraulic structures such as dams in high-turbidity underwater environments. This is primarily due to limitations such as strong scattering interference, low image contrast, and blurred damage features, leading to high false negative rates and poor robustness. To address this critical technical bottleneck, this invention proposes an intelligent detection architecture that deeply couples the physical characteristics of polarized light imaging with an improved YOLOv7-AC model. This scheme utilizes Stokes vectors (… , , The system generates a polarization difference image (PDI) as a physical prior input feature fusion module, which collaborates with the RGB image to construct an initial feature map rich in structural information. The polarization feature enhancement module introduces a global attention mechanism (GAM), using channel and spatial attention to adaptively enhance crack-related edge and texture features and focus on key areas. The multi-scale crack detection module further combines a physically guided attention mechanism to accurately respond to complex morphologies ranging from millimeter-level microcracks to macroscopic cracks at different scales. This technical solution not only overcomes the limitations of turbid water on imaging quality but also achieves dual enhancement through "physical laws + data-driven" approaches, significantly improving detection accuracy, generalization ability, and real-time performance. Therefore, it is particularly suitable for high-reliability scenarios such as dam safety maintenance, hazard investigation, emergency response, and structural health monitoring (SHM), providing a feasible and scalable technical path for intelligent operation and maintenance of underwater infrastructure.

[0048] In one possible implementation of this application embodiment, the above-described S204 body can be specifically described as follows: Spatial modulation of intermediate feature maps is performed based on physical attention and global attention mechanisms. The specific modulation process is as follows: The physical attention mechanism generates a first spatial attention mask based on polarization difference images; Specifically, the method for generating the first-space attention mask is as follows: The polarization difference image is processed through at least one lightweight convolutional layer, and a first spatial attention mask belonging to the interval [0,1] is generated using the Sigmoid activation function.

[0049] The global attention mechanism is based on the intermediate feature map itself, and generates a second spatial attention mask by modeling the dependency between channels and spatial dimensions. Specifically, the method for generating the second-space attention mask is as follows: The channel attention submodule performs global pooling on the intermediate feature map and processes it through a multilayer perceptron to generate channel weights, and then performs channel-dimensional weighting on the intermediate feature map based on the channel weights. It should be noted that the original intermediate feature map is weighted and summed according to the channel weights to emphasize important feature channels.

[0050] The spatial attention submodule aggregates the intermediate feature maps by channel dimension and performs spatial convolution with a large convolution kernel to generate spatial weights. Based on the spatial weights, it performs spatial dimension weighting on the channel-weighted feature maps. It should be noted that spatial weights are used to further weight the feature map after channel weighting with spatial dimensions.

[0051] A second attention mask is generated by weighting feature maps based on channel dimension and spatial dimension.

[0052] It should be noted that, assuming the intermediate feature map is a multi-channel image, the channel attention submodule may find that certain channels are particularly important for crack identification and therefore assign higher weights to these channels. The spatial attention submodule, on the other hand, may notice that certain specific regions are more likely to contain cracks than other regions and thus assign higher spatial weights to these regions.

[0053] The deep learning detection model also includes a learnable fusion module, which is used to weightedly fuse the first spatial attention mask and the second spatial attention mask to obtain a comprehensive attention mask, and then use the comprehensive attention mask to perform element-wise multiplication with the intermediate feature map to perform spatial modulation on the intermediate feature map.

[0054] Based on the above technical solutions, in high-turbidity underwater environments, crack features are weak and easily masked by scattering noise. A single attention mechanism struggles to balance physical priors with data adaptability, leading to insufficient robustness and a high false alarm rate in the detection model. To address this issue, this invention proposes a dual-path spatial modulation strategy that integrates physical attention and global attention (GAM) mechanisms. The physical attention mechanism generates a first spatial attention mask based on polarization difference images (PDI), directly incorporating optical physics principles—highlighting potential damage regions with strong polarization responses—providing clear interpretability and anti-interference capabilities. The global attention mechanism, on the other hand, starts from the intermediate feature map itself, emphasizing key feature channels (such as edges and textures) for crack identification through channel attention, and then focusing on high-probability damage locations through spatial attention, achieving data-driven context awareness. The two are weighted and integrated through a learnable fusion module to form a comprehensive attention mask that conforms to both physical principles and task semantics, enabling precise modulation of the feature map. This scheme effectively overcomes the generalization defects of purely data-driven methods in low signal-to-noise ratio scenarios, while avoiding the insufficient adaptability of purely physical methods to complex backgrounds. This significantly improves the model's sensitivity and positioning accuracy for micro and multi-scale cracks in turbid water, meeting the urgent need for highly reliable and robust intelligent monitoring of critical infrastructure such as dams.

[0055] In one possible implementation of this application embodiment, the above-described S205 body can be specifically described as follows: Extract the equivalent length L and spatial location of the crack from the damage information. In conjunction with material degradation patterns, stress concentration effects, and prior structural sensitivity, the comprehensive risk index IRI is calculated. IRI is defined as: ; Wherein, the equivalent length L is the actual crack length or area equivalent to the length of a straight line. The reference length is used for normalization, and m is the material degradation coefficient, where m>1; For the spatial orientation angle of the damage The stress concentration factor in the damaged region under the material brittleness parameter k; β is the proportionality coefficient, β [0,1] is used to control the weight of the stress concentration term on the total risk, avoiding excessive amplification; For structural sensitivity weighting fields; It should be pointed out that, It is a function characterizing the degree of stress concentration under the combined effect of damage direction and material properties. ;in, The length of half a crack. , The radius of curvature of the crack tip is . The principal stress direction is the direction of pure tension or pure compression within the material. In structural damage assessment, it is the key physical basis for determining whether a crack is in a high-risk propagation state. when hour, , At its maximum, the crack tends to propagate along the direction of the principal tensile force, posing a high risk; when and When vertical, , The cracks showed almost no tendency to expand.

[0056] The structural sensitivity weight field is composed of prior knowledge, and its expression is: ; in, For spatial points The distance to the nearest critical component, which includes load-bearing beams, anchorage zones, or stress concentration areas; λ is the attenuation coefficient, used to characterize the rate at which structural sensitivity decreases with distance.

[0057] The closer the damage is to the critical component, the larger W is and the higher the risk; the farther the damage is from the critical component, the smaller W is and the lower the risk.

[0058] It should be noted that the attenuation scale coefficient is preset according to the structure type. For concrete hydraulic structures, the value is 0.5 meters; for steel bridges, the value is 0.3 meters. This value is determined based on the stress diffusion range in structural mechanics and engineering practice experience, and can also be further optimized through historical failure data or finite element simulation.

[0059] This application presents an IRI (Infrastructure Responsibility Index) as a mathematical expression for a comprehensive risk index used to quantify the potential hazard of damage (such as cracks) detected in underwater structures (such as dams, bridges, and tunnels). It integrates four dimensions: materials science, fracture mechanics, structural engineering, and location sensitivity.

[0060] Mapping IRIs to discrete risk levels: ; in, , For a preset threshold, and < .

[0061] Based on the above technical solutions, traditional single-dimensional analysis methods often fail to fully reflect the complexity of potential hazards when assessing crack damage in underwater structures such as dams, bridges, or tunnels. Therefore, this application proposes the concept of an Integrated Risk Index (IRI), which comprehensively quantifies the risk of damage such as cracks by integrating material degradation patterns, stress concentration effects, and prior information on structural sensitivity. Equivalent length L and spatial location are also considered. As basic input parameters, information on the crack size and location is provided. Combined with the material degradation coefficient *m*, the additional risks caused by material aging can be adjusted. Next, The stress concentration factor takes into account the interaction between the crack direction and the brittle properties of the material, especially when the crack direction is close to the principal stress direction, its propagation risk increases significantly. This fracture mechanics-based method can accurately capture the probability and trend of crack propagation. Furthermore, a structural sensitivity weight field is introduced. This method adjusts risk assessments based on the distance of cracks from critical components. It considers not only the impact of local stress environments but also the overall structural characteristics, emphasizing the importance of critical areas. By setting appropriate attenuation coefficients λ for different structural types, it can flexibly adapt to the specific requirements of different engineering structures. Integrating these factors into an IRI (Intensive Risk Index) and mapping it to discrete risk levels achieves a shift from quantitative analysis to qualitative evaluation, facilitating decision-makers to quickly understand the severity of damage and formulate corresponding maintenance plans. The advantage of this method lies in providing a multi-dimensional and systematic framework that considers both physical factors and practical considerations in structural safety assessments. This approach not only improves the accuracy of risk assessments but also identifies key areas for subsequent repair work, helping to optimize resource allocation, improve maintenance efficiency, and reduce potential safety hazards.

[0062] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

Claims

1. A method for underwater structure damage identification based on optical polarization imaging, characterized in that, include: Acquire light intensity images of underwater target areas at multiple preset polarization angles; Based on the light intensity image, calculate the multidimensional physical quantities characterizing the polarization state, and construct multi-channel input data containing the multidimensional physical quantities; The multidimensional physical quantities include Stokes vectors and polarization difference images; The multi-channel input data is input into a deep learning detection model, which is obtained by fusing a physical attention mechanism in the YOLOv7-AC architecture. The physical attention mechanism is an attention generation mechanism driven by external polarization imaging physical quantities to guide the deep learning detection model to focus on structural damage-sensitive regions. The deep learning detection model spatially modulates the intermediate feature map based on physical attention and global attention mechanisms; the intermediate feature map is a multi-channel feature tensor generated by the deep learning detection model in the backbone network or feature pyramid network. Based on the spatially modulated intermediate feature map, the deep learning detection model outputs damage information, and the risk level is assessed based on the damage information; the damage information includes damage location and damage parameters.

2. The underwater structure damage identification method based on optical polarization imaging according to claim 1, characterized in that, The physical attention mechanism generates a first spatial attention mask based on a polarization difference image, and uses the first spatial attention mask to modulate the intermediate feature map in the YOLOv7-AC architecture.

3. The underwater structure damage identification method based on optical polarization imaging according to claim 1, characterized in that, The deep learning detection model includes a feature fusion module, a polarization feature enhancement module, and a multi-scale crack detection module; The feature fusion module is used to receive multi-channel input data and generate an initial feature map through convolutional layers, batch normalization, and nonlinear activation functions. The polarization feature enhancement module is used to introduce a global attention mechanism; the global attention mechanism includes a channel attention submodule and a spatial attention submodule. The multi-scale crack detection module is used to expand the structure of the backbone network or feature pyramid network and introduce a physically guided attention mechanism.

4. The underwater structure damage identification method based on optical polarization imaging according to claim 3, characterized in that, The deep learning detection model spatially modulates the intermediate feature map based on physical attention and global attention mechanisms, including: The physical attention mechanism generates a first spatial attention mask based on a polarization difference image calculated from a multi-angle polarization intensity image. The global attention mechanism generates a second spatial attention mask based on the intermediate feature map itself by modeling the dependency between channels and spatial dimensions. The deep learning detection model further includes a learnable fusion module, which is used to weightedly fuse the first spatial attention mask and the second spatial attention mask to obtain a comprehensive attention mask, and then use the comprehensive attention mask to perform element-wise multiplication with the intermediate feature map to spatially modulate the intermediate feature map.

5. The underwater structure damage identification method based on optical polarization imaging according to claim 1, characterized in that, The risk level assessment based on damage information includes: The equivalent length L and spatial location of the crack are extracted from the damage information. In conjunction with material degradation patterns, stress concentration effects, and prior structural sensitivity, a comprehensive risk index (IRI) is calculated, defined as follows: ； in, The reference length is given by m, the material degradation coefficient (m>1), and β is the proportionality coefficient. [0,1], For the spatial orientation angle of the damage Under the material brittleness parameter k, the stress concentration factor in the damaged region , The length of half a crack. , The radius of curvature of the crack tip is . The principal stress direction For structural sensitivity weighting fields; Mapping IRIs to discrete risk levels: ； in, , For a preset threshold, and < .

6. The underwater structure damage identification method based on optical polarization imaging according to claim 5, characterized in that, The structural sensitivity weight field includes: The structural sensitivity weight field is composed of prior knowledge, and its expression is: ； in, For spatial points The distance to the nearest critical component, which includes load-bearing beams, anchorage zones, or stress concentration areas; λ is the attenuation coefficient, used to characterize the rate at which structural sensitivity decreases with distance.

7. The underwater structure damage identification method based on optical polarization imaging according to claim 4, characterized in that, The method for generating the first spatial attention mask includes: The polarization difference image is processed through at least one lightweight convolutional layer, and a first spatial attention mask belonging to the interval [0,1] is generated using the Sigmoid activation function.

8. The underwater structure damage identification method based on optical polarization imaging according to claim 4, characterized in that, The method for generating the second spatial attention mask includes: The channel attention submodule performs global pooling on the intermediate feature map and processes it through a multilayer perceptron to generate channel weights, and then performs channel dimension weighting on the intermediate feature map based on the channel weights. The spatial attention submodule aggregates the intermediate feature map by channel dimension, performs spatial convolution with a large convolution kernel to generate spatial weights, and then performs spatial dimension weighting on the channel-weighted feature map based on the spatial weights. A second attention mask is generated by weighting feature maps based on channel dimension and spatial dimension.

9. The underwater structure damage identification method based on optical polarization imaging according to claim 1, characterized in that, The method for obtaining the multidimensional physical quantities includes: Based on the light intensity image, calculate the Stokes vector, including the total light intensity. Horizontal-vertical polarization difference polarization difference Right-handed to left-handed circular polarization difference ; Based on the Stokes vector, using the formula PDI= The polarization difference image (PDI) is calculated.

10. An underwater structure damage identification system based on optical polarization imaging, operating based on the underwater structure damage identification method based on optical polarization imaging according to any one of claims 1-9, characterized in that, It includes a detection module, as well as an acquisition module and an evaluation module connected to it; The acquisition module is used to acquire light intensity images of an underwater target area at multiple preset polarization angles; Based on the light intensity image, calculate the multidimensional physical quantities characterizing the polarization state, and construct multi-channel input data containing the multidimensional physical quantities; The detection module is used to input the multi-channel input data into the deep learning detection model, which is obtained based on the physical attention mechanism fused in the YOLOv7-AC architecture. Spatial modulation of intermediate feature maps is performed based on physical attention mechanism and global attention mechanism; The evaluation module is used to output damage information from a deep learning detection model based on the spatially modulated intermediate feature map, and to evaluate the risk level based on the damage information.

Images