A method for detecting wall surface damage structure based on polarization imaging

Abstract

The application provides a wall damage structure detection method based on polarization imaging, and relates to the field of image recognition, and the steps include: collecting single-channel gray digital images of a damaged wall in four polarization directions of 0°, 45°, 90° and 135° to construct a polarization imaging data set; taking the 90° polarization direction image as input, combining a scale smoothing kernel and a non-Gaussian anisotropic kernel to obtain a multi-scale polarization structure enhancement feature map; generating a polarization angle feature and a polarization degree feature based on a polarization intensity image and constructing a local structure reconstruction tensor; using the polarization degree feature to construct a multi-directional geometric perception kernel to extract a damage response feature and fuse the damage response feature with the polarization intensity image to output a damage probability map; and integrating the above steps into a unified wall damage structure detection model, updating model parameters through an iterative optimization training strategy and realizing accurate detection of a damage area.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and in particular to a wall surface damage structure detection method based on polarization imaging. BACKGROUND

[0002] As the core enclosure structure in civil infrastructure and industrial buildings, the structural integrity of building walls directly affects the safety performance and service life of buildings. Wall damage detection technology plays a key role in building health monitoring, preventive maintenance and safety evaluation. In particular, in the non-destructive detection system based on optical imaging, the accurate extraction and identification of wall damage features are of great significance for discovering early diseases, assessing damage degree and developing maintenance strategies. With the increasing complexity of modern building structures and the diversity of materials, the damage patterns of walls under the coupling effects of environmental erosion, load changes and material aging show weak, multi-type and anisotropic distribution characteristics. It is urgent to develop more robust and discriminative detection methods to achieve high-precision identification and quantitative evaluation of wall damage structures.

[0003] Publication No. CN210629548U discloses a curtain wall damage detection device for remote communication, which constructs a piezoelectric sensing device through a plurality of piezoelectric elements arranged in a uniform lattice. After detecting the vibration signal or sound of wall damage, it is converted into an electrical signal for alarm. Publication No. CN113592869A discloses a building curtain wall glass breakage image recognition method and alarm system, which cascades feature extraction nodes, curtain wall damage area-offset extraction nodes, area attribute search nodes and area attribute prediction nodes to obtain a curtain wall damage area prediction unit. By using a multi-layer structure AI network unit and a curtain wall damage trigger prediction unit with combined attention feature layers, the identification of glass curtain wall damage is realized.

[0004] The prior art does not fully consider the anisotropic response characteristics of wall damage in the polarization physical field, ignores the coupling mechanism of polarization state and surface geometry, the synergistic enhancement relationship between multi-directional features and the dynamic normalization modeling requirements, resulting in the problems of insufficient damage feature extraction, insufficient noise suppression capability and low reliability of probabilistic output in the actual detection environment with diverse materials and complex environmental conditions. SUMMARY

[0005] The application provides a wall damage structure detection method based on polarization imaging, aiming at the problem of insufficient detection accuracy caused by weak damage characteristics, multiple types and anisotropy in a complex building wall environment, and proposes a multi-stage processing mechanism that fuses multi-scale polarization feature enhancement, local structure reconstruction and multi-directional geometric perception, including constructing a multi-scale polarization structure enhancement feature map based on a non-Gaussian anisotropic kernel and a weighted fusion strategy, generating a local structure reconstruction tensor by coupling a gradient intensity image and a polarization structure enhancement feature map, and realizing damage area fusion decision by combining a polarization degree constraint geometric perception kernel and a dynamic normalization factor, so as to realize unified representation and high-precision detection of wall damage physical characteristics and geometric morphology.

[0006] A wall damage structure detection method based on polarization imaging, the specific method is:

[0007] S1, based on , , and polarization direction acquisition of single-channel gray image of damaged wall, and construction of wall damage image dataset;

[0008] S2, taking polarization direction imaging as input, respectively performing twice convolution of scale smoothing kernel and non-Gaussian anisotropic kernel, and outputting multi-scale polarization structure enhancement feature map;

[0009] S3, taking , , polarization image to generate polarization intensity image, generating polarization angle feature and polarization degree feature based on the polarization intensity image, and obtaining polarization angle gradient term based on the polarization angle feature, and combining the multi-scale polarization structure enhancement feature map to obtain local structure reconstruction tensor;

[0010] S4, constructing four-direction polarization geometric perception kernel based on the polarization degree feature, extracting damage response feature from the local structure reconstruction tensor, and performing decision feature fusion based on the second polarization intensity image and the third polarization intensity image in S3 to obtain a damage area map;

[0011] S5, integrating the processing procedures of S2 to S4 into a wall damage structure detection model based on polarization imaging, and the model realizes multi-scale feature enhancement, structure reconstruction and damage detection;

[0012] S6, inputting the wall damage image dataset into the model for detection, and updating the model parameters through an iterative optimization training strategy to obtain damage in the image.

[0013] Preferably, for construction of the wall damage image dataset based on polarization imaging, first, a transmission axis fixed at , , and A linear polarizer with a specific orientation ensures... The polarization direction was perpendicular to the ground baseline to suppress the horizontal polarization component from the mirror reflection of the wall surface and enhance the vertical polarization signal caused by surface roughness and cracks. Secondly, the shooting environment was chosen during the day with ample natural light, and LED lights with polarizers were used to ensure that the incident light contained sufficient polarization information. Additionally, data was collected from the damaged wall, and the camera was stably aimed at the target area for shooting, ultimately obtaining... , , and A single-channel grayscale digital image with polarization direction, where the grayscale value of each pixel corresponds to the position in... , , and The reflection intensity under polarization is uniformly cropped to 778×778 for each single-channel grayscale digital image under each polarization direction to construct a standard polarization-based imaging dataset of wall damage images.

[0014] Preferably, in step S2, the polarization-imported wall damage image is compared with the scale smoothing kernel. Perform convolution to obtain the basic scale feature set. ;

[0015] A non-Gaussian anisotropic kernel aligned with the 90° polarization axis and odd-symmetric about the vertical direction is constructed. The construction process first determines the center coordinate position of the kernel function, which is automatically calculated based on the kernel size parameter. The mathematical structure of the kernel function includes linear coordinate components and exponentially decaying components. The linear components generate odd-symmetric response characteristics in the vertical direction, while the exponential components achieve spatial decay control through the radial distance squared term. The scale adaptability of the kernel function is achieved through a width control factor, which maintains a quantitative relationship with the kernel size parameter.

[0016] Set of basic scale features Its corresponding non-Gaussian anisotropic kernel Perform convolution operations to output an enhanced set of directional structural features. Determine the feature weights for each scale through model training. Fuse the enhanced set of directional structural feature maps with the scale feature weights to output a multi-scale polarization structure enhanced feature map.

[0017] Further, in view of the problem that the multi-scale characteristics and the non-obvious directional response of the wall surface damage structure under the polarization imaging condition, the application provides a feature enhancement mechanism combining multi-scale smoothing and anisotropic kernel convolution, first, taking the 90° polarization direction image as input, the image is convolved by different scale smoothing kernels to construct a basic scale feature set to stably represent the structure distribution state of the wall surface under different spatial scales, then a symmetric non-Gaussian anisotropic kernel is constructed in the vertical polarization direction, by introducing the vertical direction differential structure and the horizontal direction Gaussian suppression, the kernel function produces directional enhancement response to the damage edge, while suppressing the interference of continuous normal texture, further, the basic scale feature and the corresponding anisotropic kernel are convolved to obtain an enhanced directional structure feature set, finally, the directional features of different scales are fused through learnable weights to form a multi-scale polarization structure enhancement feature map, which can adaptively highlight the texture mutation features of the multi-scale damage area and keep the response consistency in the vertical direction, effectively improving the perception accuracy and robustness of the subsequent module to defects.

[0018] Preferably, in the S3 step, the wall surface damage structure images of different polarization directions are combined to obtain three polarization intensity images.

[0019] According to the three polarization intensity images, the polarization angle feature is constructed by nonlinear arctangent function operation, and the polarization degree feature is constructed by root formula;

[0020] Based on the polarization angle feature, the gradient is calculated in the horizontal and vertical directions to obtain the polarization angle gradient term;

[0021] Based on the multi-scale polarization structure enhancement feature map, the horizontal direction decomposition and the vertical direction decomposition are performed to obtain the horizontal direction decomposition feature and the vertical direction decomposition feature, the feature matrix is constructed based on the horizontal direction decomposition feature and the vertical direction decomposition feature, and the determinant is calculated, the ratio of the polarization degree feature and the polarization angle gradient term is combined to obtain the local structure reconstruction tensor.

[0022] Further, in view of the problem that the anisotropic response of the damaged area of the wall surface in the polarization physical field is insufficiently coupled with the geometric structure, the application proposes a structure reconstruction mechanism fusing the polarization physical quantity and the multi-scale geometric feature, first constructs three polarization intensity images based on the original images of three polarization directions to provide basic data support for subsequent physical feature extraction, then analyzes the polarization angle feature and the degree of polarization feature through the arctangent function and the square root of the sum of squares operation respectively, quantizes the optical polarization state into calculable physical parameters, and accurately describes the orientation characteristics and ordered changes of the surface microstructure, further calculates the spatial gradient of the polarization angle feature to obtain the polarization angle gradient term reflecting the geometric mutation characteristics, simultaneously decomposes the multi-scale polarization structure enhancement feature map into horizontal and vertical direction components to construct a second-order tensor matrix representing the local geometric structure, and finally generates a local structure reconstruction tensor through the determinant operation combined with the degree of polarization feature and the polarization angle gradient term, effectively unifies the polarization physical characteristics and the geometric structure features, retains the anisotropic response mode of the damaged area, enhances the structure expression ability of the edge and texture, and provides high signal-to-noise ratio input features for subsequent multi-directional perception.

[0023] Preferably, in the S4 step, four directional polarization geometric perception kernels are constructed based on the degree of polarization feature as the polarization constraint;

[0024] The local structure reconstruction tensor is convolved based on the four directional polarization geometric perception kernels respectively to obtain the damaged response features of the corresponding directions;

[0025] The four directional damaged response features are spliced in the channel dimension to form the fusion damaged response feature;

[0026] A normalization factor is dynamically constructed based on the local energy mean of each channel of the fusion damaged response feature, an exponential decay function is used in the numerator part to perform nonlinear mapping of the response value of each channel in the probability interval, the square sum and square root operation of the second and third polarization intensity are introduced in the denominator part to form the normalization term, and the damaged area map is generated through the probabilistic conversion.

[0027] Further, for multi-directional geometric perception of wall surface damage and output of damaged area and detection of damage category, first, a four-direction polarization geometric perception kernel is constructed based on a polarization degree feature to form a perception kernel group that selectively responds to damage with a specific orientation; then, the local structure reconstruction tensor is respectively convolved with each direction perception kernel to extract damage response features with direction specificity, and a multi-directional response tensor is generated through channel dimension fusion to comprehensively capture the spatial distribution and direction pattern of the damaged area; further, a dynamic normalization factor is calculated based on the local energy of each channel response value, the direction response value is mapped to a probability interval through an exponential decay function, and a multiplication operation is used to strengthen the consistency of multi-directional response; finally, a second polarization intensity image and a third polarization intensity image are introduced to output a final damaged area image; the probability map not only retains the spatial distribution characteristics of multi-directional geometric response, but also ensures the explainability of the optical mechanism through polarization physical constraints, significantly improving the accuracy and robustness of damage recognition in complex environments.

[0028] Preferably, in the S5 step, the S2 step is represented as a non-Gaussian polarization feature extraction unit, and the wall surface damage image dataset obtained by polarization imaging is processed through the non-Gaussian polarization feature extraction unit to obtain a multi-scale polarization structure enhanced feature map;

[0029] The S3 step is represented as a polarization feature fusion debugging unit, and the multi-scale polarization structure enhanced feature map is subjected to polarization debugging through the polarization feature fusion debugging unit to obtain a local structure reconstruction tensor;

[0030] The S4 step is represented as a multi-directional polarization geometric perception unit, and the local structure reconstruction tensor damage response feature is output through the multi-directional polarization geometric perception unit;

[0031] Steps S2 to S4 construct a wall surface damage structure detection model based on polarization imaging.

[0032] Further, the wall surface damage structure detection model based on polarization imaging first passes through a non-Gaussian polarization feature extraction unit, which can suppress background interference and enhance damage direction structure features, improving the discrimination of feature expression; then passes through a polarization feature fusion debugging unit for deep fusion of multi-scale features and polarization angle and polarization degree, ensuring the integrity and accuracy of local structure reconstruction; finally, passes through a multi-directional polarization geometric perception unit to extract response features from different directions and perform decision fusion to improve the robustness and stability of damage area detection; in summary, the method has significant advantages in feature enhancement, structure reconstruction and damage recognition accuracy, and can realize efficient detection of wall surface damage in complex scenes.

[0033] In the above technical solution, the present application provides technical effects and advantages:

[0034] The application constructs a wall damage structure detection method based on polarization imaging, proposes multi-scale polarization feature enhancement, local structure reconstruction and multi-directional geometric perception mechanism, and aims at the problem that the traditional optical detection method has insufficient response to weak damage and anisotropic characteristics in complex wall environment. First, a non-Gaussian anisotropic kernel and a multi-scale weighted fusion strategy are introduced, and a multi-scale polarization structure enhancement feature map is constructed combined with the polarization physical constraint. This step can effectively enhance the texture mutation characteristics of the damage area, break through the limitation of the sensitivity of the traditional method to the polarization direction, realize the structured expression of multiple types of damage, and significantly improve the discriminability and environmental adaptability of feature extraction.

[0035] The application establishes a polarization intensity image and generates a local structure reconstruction tensor by coupling polarization degree features and polarization angle features, and then jointly modulates multi-scale geometric features and polarization physical quantities, introduces a determinant operation and a ratio constraint to construct a structure reconstruction mechanism. This mechanism effectively unifies the surface geometric shape and the polarization optical characteristics, ensures the complete capture of the anisotropic response of the damage area and the cooperative suppression of noise interference, thereby improving the physical interpretability of feature expression and the ability to characterize complex damage patterns.

[0036] The application realizes damage area fusion decision based on a four-direction polarization geometric perception kernel and a dynamic normalization factor, and generates a probabilistic output combined with a multi-directional response polarization intensity image constraint. This design realizes the cooperative optimization of direction-specific response and polarization physical mechanism, significantly enhances the accuracy of the probability map in spatial distribution and the robustness to material diversity and illumination changes, and effectively improves the recognition accuracy and reliability of wall damage detection in real scenes.

[0037] The application constructs a wall damage structure detection model based on polarization imaging. First, a non-Gaussian polarization feature extraction unit is used to suppress background interference and enhance damage direction structure features, improve the discriminability of feature expression. Then, a polarization feature fusion adjustment unit is used to realize deep fusion of multi-scale features and polarization angle and polarization degree, ensure the integrity and accuracy of local structure reconstruction. Finally, a multi-directional polarization geometric perception unit is used to extract response features from different directions and make decision fusion, improve the robustness and stability of damage area detection. In summary, the method of the application has significant advantages in feature enhancement, structure reconstruction and damage recognition accuracy, and can realize efficient detection of wall damage in complex scenes. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of a wall damage structure detection method based on polarization imaging provided by the application.

[0039] Figure 2It is a non-Gaussian polarization feature extraction unit structure diagram provided by the application.

[0040] Figure 3 It is a polarization feature fusion debugging unit structure diagram provided by the application.

[0041] Figure 4 It is a multi-directional polarization geometric perception unit structure diagram provided by the application.

[0042] Figure 5 It is a structure diagram of a wall surface damage structure identification model based on polarization imaging provided by the application.

[0043] Figure 6 It is a wall surface damage structure diagram based on polarization imaging provided by the application.

[0044] Figure 7 It is a wall surface damage structure detection diagram provided by the application. Specific implementation method

[0045] The application proposes a detection method for wall surface damage structure based on polarization imaging, aiming at the problem of insufficient detection accuracy caused by the weak nature, multi-type nature and anisotropy of the damage features in the complex building wall surface environment, and proposes a multi-stage processing mechanism that fuses multi-scale polarization feature enhancement, local structure reconstruction and multi-directional geometric perception, including constructing a multi-scale polarization structure enhancement feature map based on a non-Gaussian anisotropic kernel and a weighted fusion strategy, generating a local structure reconstruction tensor by coupling a gradient intensity image and a polarization structure enhancement feature map, realizing damage area fusion decision by combining a polarization degree constrained geometric perception kernel and a dynamic normalization factor, so as to realize unified representation and high-precision detection of the physical characteristics and geometric morphology of the wall surface damage.

[0046] S1, based on , , and Four polarization directions, collect the corresponding single-channel gray digital image of the damaged wall surface, and construct a wall surface damage image dataset based on polarization imaging.

[0047] The construction of the wall surface damage dataset based on polarization imaging can be divided into three steps, first, install a linear polarizer with transmission axis fixed in , , and Direction in front of the polarization camera, ensure The polarization direction is perpendicular to the ground baseline to suppress the horizontal polarization component from the mirror reflection of the wall surface and enhance the vertical polarization signal caused by surface roughness and cracks; secondly, the shooting environment is selected during the daytime with sufficient natural light, and LED lights with polarizers are used to ensure that the incident light contains sufficient polarization information; thirdly, the damaged wall is sampled, and the camera is stably aimed at the target area for shooting, ultimately obtaining... , , and A single-channel grayscale digital image with polarization direction, where the grayscale value of each pixel corresponds to the position in... , , and Reflection intensity under polarization.

[0048] S2. Based on the wall damage image obtained from polarization imaging, a basic scale feature set is obtained. A non-Gaussian anisotropic kernel is designed. The basic scale feature set and the non-Gaussian anisotropic kernel are then subjected to weighted convolution calculation to output a multi-scale polarization structure enhancement feature map. .

[0049] Furthermore, in step S2, a multi-scale polarization structure enhancement feature map is obtained, the process of which is as follows: Figure 2 As shown, the specific steps are as follows:

[0050] In this embodiment, the set of convolution kernel side lengths is defined as follows: Define the equivalent scale smoothing kernel as Wall damage images based on scale smoothing kernel of the polarization imaging Perform convolution to obtain the basic scale feature set. The mathematical model for the basic scale feature set is as follows:

[0051] ;

[0052] in, This is represented as an image of wall damage with a polarization angle of 90 degrees. Represented as convolution, Represented as an equivalent scale smoothing kernel, Represented as the set of convolution kernel side lengths, Represented as a set of features at the basic scale;

[0053] In this embodiment, convolution using an equivalent scale smoothing kernel can stably suppress noise and smooth local fluctuations, thereby improving the stability and reliability of basic features in the subsequent feature extraction stage of wall damage images in polarization imaging.

[0054] The non-Gaussian anisotropic kernel is constructed with 90° polarization axis alignment and odd symmetry about the vertical direction. The center coordinate position of the kernel function is determined first in the construction process, which is automatically calculated according to the kernel size parameter. The mathematical structure of the kernel function includes linear coordinate components and exponential decay components. The linear components generate odd symmetry response characteristics in the vertical direction, and the exponential components realize spatial decay control through the square term of the radial distance. The scale adaptability of the kernel function is realized by the width control factor, which maintains a quantitative relationship with the kernel size parameter.

[0055] In this embodiment, the basic scale feature set can only reflect the multi-scale overall structure information, and it is difficult to accurately highlight the damage texture mutation in the polarization direction. Therefore, a non-Gaussian anisotropic kernel is introduced to construct an odd symmetry structure in the vertical direction and apply Gaussian suppression to the horizontal direction, so that it has a high response to the edge of the polarization direction and maintains suppression to the continuous normal texture, thereby avoiding the weakening of the directional information caused by the isotropic smoothing kernel and enhancing the expression ability of the polarization direction structure. The mathematical model of the non-Gaussian anisotropic kernel is:

[0056] ;

[0057] wherein, is the inner coordinate of the non-Gaussian anisotropic kernel, , is the center position of the kernel, is the scale-related Gaussian width control factor;

[0058] In this embodiment, represents the convolution kernel side length set, and the value is set to , which is used to realize multi-scale structure response extraction. Different scales can respectively perceive the texture and edge features of fine-grained, small and medium-sized damage areas, and enhance the adaptability to different structure sizes. The kernel center position is , which is adaptively set with the scale, which helps to maintain the symmetry of the kernel function and make the response in the vertical direction, i.e. the polarization sensitive direction, more balanced and accurate. The scale-related Gaussian width control factor is , which is used to adjust the diffusion range of the kernel. A smaller value focuses on details and enhances them, and a larger value is suitable for extracting fuzzy or weak edge structures, realizing flexible adaptation to different structure scales. Designing a non-Gaussian anisotropic kernel can effectively enhance the structure features in the vertical direction of the polarization image and improve the clarity and recognition of cracks, peeling and other structures in the wall damage image.

[0059] The basic scale feature set is convolved with the corresponding non-Gaussian anisotropic kernel to output the enhanced directional structure feature set. Further, according to the weight, the multi-scale polarization structure enhancement feature map is output by fusing the enhanced directional structure features of different scales. ;

[0060] In the embodiment, the non-Gaussian anisotropic kernel is applied to the basic feature map set, and the structural mutation information of the polarization direction is explicitly mapped into the feature space through a convolution operation, thereby further highlighting the significant response of the damaged area in the corresponding direction, suppressing the continuous distribution of normal texture, and enhancing the directional structure feature map after enhancement. The mathematical model of the enhanced directional structure feature map is:

[0061] ;

[0062] wherein, denotes the enhanced directional structure feature map set, denotes the basic scale feature set, denotes convolution, denotes a non-Gaussian anisotropic kernel, denotes a convolution kernel edge length set;

[0063] The enhanced directional structure feature map set is fused based on a learnable weight to form a multi-scale polarization structure enhanced feature map , and the mathematical model is:

[0064] ;

[0065] wherein, denotes each scale feature weight, which is learned in the training stage, denotes point multiplication.

[0066] In the embodiment, as each scale feature weight, which dynamically changes in the model training stage, and can automatically adjust the influence degree of each scale enhanced feature according to the actual contribution of different scale damaged textures in the polarization wall surface damaged image. Each scale weight , and satisfies ; by weighting and fusing the directional enhanced feature maps under three scales, a multi-scale polarization structure enhanced feature map PEC is formed, so that the damaged area has consistent high response in the multi-scale space, and the accuracy of the subsequent damage recognition module is further improved.

[0067] S3, according to the wall surface damaged structure image of different polarization directions, a polarization intensity image is constructed, polarization angle features and polarization degree features are generated, a polarization angle gradient term is obtained based on the polarization angle features, and a local structure reconstruction tensor is obtained by combining the multi-scale polarization structure enhanced feature map.

[0068] Further, in the S3 step, the local structure reconstruction tensor is constructed, and the flowchart is as shown in Figure 3 The specific steps of constructing the local structure reconstruction tensor are as follows:

[0069] The wall damage structure images of different polarization directions are combined to obtain polarization intensity images;

[0070] In this embodiment, the wall damage structure images of three polarization directions are combined to obtain three polarization intensity images, and the mathematical model of the polarization intensity images is:

[0071] ;

[0072] wherein, , , respectively represent the first polarization intensity image, the second polarization intensity image and the third polarization intensity image, , , respectively represent the wall damage structure images of the polarization directions , , ;

[0073] In this embodiment, the three polarization directions are set as , , , which can capture the mirror reflection information of horizontal and vertical damage, detect the oblique cracks and material peeling, and capture the polarization response difference to the greatest extent while reducing the data amount; through the design of the parameters, three polarization intensity images are constructed to completely capture the damage characteristics in different directions and enhance the physical coupling of the characteristics.

[0074] According to the three polarization intensity images, a polarization angle feature is constructed through nonlinear combination, and a degree of polarization feature is constructed through root formula;

[0075] In this embodiment, the polarization state distribution of the wall damage area has significant direction sensitivity, and the polarization angle of the intact area presents a random distribution characteristic; therefore, an arctangent form polarization state solving function is combined with a polarization intensity image ratio constraint to construct a polarization angle feature, and the mathematical model is:

[0076] ;

[0077] wherein, represents the polarization angle feature, describes the spatial distribution of the angle between the incident light wave vibration direction and the reference axis, and can reflect the orientation characteristics of the micro-geometric structure of the wall;

[0078] In this embodiment, the second polarization intensity image and the third polarization intensity image are subjected to square sum and square root calculation, and the first polarization intensity image is combined through calculation of the intensity ratio to construct a degree of polarization feature, and the mathematical model is:

[0079] ;

[0080] in, This indicates the polarization degree characteristic, reflecting the ordered changes in the wall's microstructure and the degree of anisotropy in light scattering. It normalizes the incident energy and quantifies the degree of polarization state modulation caused by damage. It is directly related to physical mechanisms.

[0081] Polarization adjustment is performed based on polarization angle and polarization degree characteristics;

[0082] In this embodiment, the polarization angle feature The gradients are calculated in the horizontal and vertical directions to obtain the polarization angle gradient term. The mathematical model is as follows:

[0083] ;

[0084] in, The term representing the polarization angle gradient directly reflects the abrupt changes in the micro-geometric structure of the wall surface, exhibiting enhanced specificity for damaged edges and immunity to noise in uniform regions.

[0085] In this embodiment, the feature map is enhanced based on the multi-scale polarization structure. Polarization adjustment is performed, and by combining the polarization degree characteristics and the polarization angle gradient term, the local structure reconstruction tensor is obtained. The mathematical model is as follows:

[0086] ;

[0087] in, Represents the local structural reconstruction tensor. Represents a determinant. Represents multi-scale polarization structure enhancement feature maps Horizontal decomposition features, Represents multi-scale polarization structure enhancement feature maps Vertical decomposition features, This indicates a local window.

[0088] In this embodiment, The central zero value ignores uniform regions and focuses on edge variations. The symmetrical weights on both sides are more sensitive to the edges of horizontal cracks. The weights in the middle row are doubled to enhance the response in the principal gradient direction. Similarly, Local window The size is set to 7×7 to avoid signal aliasing. With the above parameter settings, the constructed local structure reconstruction tensor significantly amplifies the anisotropic difference between the damaged area and the intact background, and provides input features with higher signal-to-noise ratio for the subsequent multi-directional polarization ensemble sensing kernel.

[0089] S4. Based on the polarization degree features, construct polarization geometric sensing kernels in four directions, extract damage response features from the local structure reconstruction tensor, and perform decision feature fusion based on the second polarization intensity image and the third polarization intensity image to obtain a damage probability map.

[0090] Furthermore, in step S4, a damage probability map is obtained, the process of which is as follows: Figure 4 As shown, the steps to obtain the damage probability map are as follows:

[0091] Based on the polarization degree characteristics as polarization constraints, a polarization geometric sensing kernel with four directions is constructed;

[0092] In this embodiment, polarization geometric sensing kernels are constructed at four angles by combining polarization degree features, resulting in polarization geometric sensing kernels in four directions. The specific mathematical model for constructing the polarization geometric sensing kernels is as follows:

[0093] ;

[0094] in, , Indicates the first Polarization geometric sensing kernels in each direction, Indicates the first The angles corresponding to each direction This indicates the attenuation control term. Represents the polarization constraint term, when When the term is non-zero, This represents the constraint threshold.

[0095] In this embodiment, Through four-directional positive interactive compensation, an unbiased response to damage features at any angle is formed; based on the image resolution, settings are... Matching the physical scale, it effectively suppresses high-frequency noise from the wall's own texture while preserving precise boundary information of the damaged area; polarization characteristics It reflects the orderliness of the wall surface and sets up... The algorithm prioritizes capturing high-confidence damaged regions and efficiently eliminates non-damaged interference, providing high-confidence candidate regions for subsequent processing, thus improving algorithm efficiency and meeting performance requirements. Through the above parameter settings, the constructed polarization geometry sensing kernel... This enables precise capture of the anisotropic polarization response of the damaged area of ​​the wall, while suppressing interference from the material itself and improving algorithm efficiency.

[0096] Damage response features are extracted from the local structure reconstruction tensor based on polarization geometric sensing kernels in four directions;

[0097] In this embodiment, the mathematical model is:

[0098] ;

[0099] wherein, represents the breakage response feature of the th direction;

[0100] In this embodiment, the breakage response features of the four directions are merged in the channel dimension to obtain the fused breakage response feature , The fused breakage response feature more comprehensively reflects the geometric shape and directionality information of the breakage region in the spatial distribution, not only improving the coverage and detection sensitivity of the complex breakage edge, but also providing more accurate and multi-dimensional input for subsequent decision feature fusion combined with the polarization intensity image, thereby significantly enhancing the accuracy and robustness of the final generated breakage probability map.

[0101] Based on the fused breakage response feature , decision feature fusion is performed combined with the second polarization intensity image and the third polarization intensity image to obtain a breakage probability map.

[0102] In this embodiment, the mathematical model for obtaining the breakage probability map is:

[0103] ;

[0104] wherein, represents the breakage probability map, represents a dynamic normalization factor, , represents the th channel of the fused breakage response feature, .

[0105] In this embodiment, the introduction of the dynamic normalization factor enables the model to maintain a stable response distribution under different wall materials, lighting conditions and imaging noise, reducing false positives and false negatives. By fusing the response features of the four directions, the final probability map more completely reflects the geometric shape of the breakage region in the spatial distribution, improving the coverage ability of complex breakage patterns. The breakage probability map as the final output form can provide intuitive probability threshold reference for the automatic detection system or subsequent intelligent decision module.

[0106] S5, the processing procedures of steps S2 to S4 are integrated into a unified wall breakage structure recognition model based on polarization imaging, which can simultaneously complete multi-scale polarization feature enhancement, local structure reconstruction and breakage probability output;

[0107] Further, in the S5 step, the wall breakage structure recognition model based on polarization imaging is constructed, and the process is as follows Figure 5As shown, the specific steps are as follows:

[0108] In this embodiment, the S2 step is a non-Gaussian polarization feature extraction unit. The wall damage image dataset obtained by polarization imaging is processed by the non-Gaussian polarization feature extraction unit to obtain a multi-scale polarization structure enhanced feature map, and the mathematical model is:

[0109]

[0110] Among them, indicates the non-Gaussian polarization feature extraction unit, indicates the wall damage image with a polarization angle of 90 degrees, indicates the multi-scale polarization structure enhanced feature map;

[0111] In this embodiment, the Gaussian polarization feature extraction unit The directional damage structure such as wall crack and peeling is highlighted, the significant region is adaptively filtered through the polarization degree threshold, and the parameter adjustment flexibility is stronger under different working conditions, so that the multi-scale polarization structure enhanced feature map more in line with the actual damage characteristics is obtained.

[0112] In this embodiment, the S3 step is a polarization feature fusion debugging unit. The multi-scale polarization structure enhanced feature map is subjected to polarization debugging by the polarization feature fusion debugging unit to obtain a local structure reconstruction tensor, and the mathematical model is:

[0113]

[0114] Among them, indicates the polarization feature fusion debugging unit, indicates the multi-scale polarization structure enhanced feature map, indicates the local structure reconstruction tensor;

[0115] In this embodiment, the polarization feature fusion debugging unit The intensity information of different polarization directions can be jointly expressed, and the polarization angle feature and the polarization degree feature can be constructed in a nonlinear manner to realize accurate description of the geometric characteristics of the damage area. At the same time, the unit can adaptively debug the enhanced feature map, reduce the interference caused by changes in light and differences in material, and ensure that the local structure reconstruction result is more stable and robust, thereby providing more reliable input features for subsequent damage response feature extraction.

[0116] In this embodiment, the S4 step is a multi-directional polarization geometric perception unit. The local structure reconstruction tensor damage response feature is processed by the multi-directional polarization geometric perception unit to obtain a damage probability map, and the mathematical model is:

[0117] ​​ ;

[0118] wherein, is represented as a multi-directional polarization geometric perception unit, is represented as a damage probability map, is represented as a local structure reconstruction tensor;

[0119] In this embodiment, the multi-directional polarization geometric perception unit The four directional perception kernels constructed by the polarization degree feature fully capture the anisotropic response patterns of the damaged area in each direction, ensuring that cracks and peeling of different orientations can be effectively activated; then the multi-directional response features extracted from the local structure tensor through convolution operation convert the abstract texture coherence into concrete directional intensity spectrum, effectively distinguishing the structural edges of real damage from random noise; finally, the high-precision and low-missed detection damage area map is generated by fusing the decision response features.

Claims

1. A method for detecting a wall surface damage structure based on polarization imaging, characterized by, The method comprises the following steps: S1, acquiring a single-channel grayscale image of a damaged wall surface based on 0°, 45°, 90° and 135° polarization directions, and constructing a wall surface damage image dataset; S2, constructing a scale smoothing kernel and a non-Gaussian anisotropic kernel, The set of convolution kernel side lengths is set as , and the equivalent scale smoothing kernel is constructed as , The mathematical model of the non-Gaussian anisotropic kernel is: ; wherein, is a non-Gaussian anisotropic kernel, , is a kernel center position, is a scale-dependent Gaussian width control factor, and the input is imaged at a 90° polarization direction, and the output is a multi-scale polarization structure enhancement feature map after two convolutions of scale smoothing kernel and non-Gaussian anisotropic kernel, respectively. S3, combining wall surface damage structure images of 0°, 45° and 90° polarization directions to obtain corresponding three polarization intensity images, and the mathematical model of the polarization intensity image is: ; wherein, 、 、 respectively represent a first polarization intensity image, a second polarization intensity image and a third polarization intensity image, 、 、 respectively represent a wall surface damage structure image with polarization direction 、 、 , polarization angle features and polarization degree features are generated based on the polarization intensity images, and a polarization angle gradient term is obtained based on the polarization angle features, combined with the multi-scale polarization structure enhancement feature map, to obtain a local structure reconstruction tensor; S4, constructing a polarization geometric perception kernel mathematical model of four polarization directions: ; wherein, , denotes a polarization geometric perception kernel of the denotes an angle corresponding to the denotes an attenuation control term, denotes a polarization constraint term, which is nonzero when denotes a constraint threshold, based on the polarization geometric perception kernel, a damage response feature is extracted from the local structure reconstruction tensor, and a decision feature fusion is performed based on the second polarization intensity image and the third polarization intensity image in S3, to obtain a damage region.​​​ S5, integrating the processing procedures of steps S2 to S4 into a wall surface damage structure detection model based on polarization imaging, and the model realizes multi-scale feature enhancement, structure reconstruction and damage detection; S6, inputting the wall surface damage image dataset into the model for detection, updating the model parameters through an iterative optimization training strategy, and obtaining a damage area in the image.

2. The method for detecting a wall damage structure based on polarization imaging according to claim 1, characterized in that, A linear polarizer with a transmission axis fixed at 0°, 45°, 90° and 135° is installed in front of the polarization camera to ensure that the 90° polarization direction is perpendicular to the ground reference line; the shooting environment is selected in daylight with sufficient natural light, and a LED lamp with a polarizer is configured; a damaged wall is collected, the camera is stably aligned with the target area for shooting, and finally single-channel grayscale digital images of four polarization directions are obtained, and a wall surface damage image dataset based on polarization imaging containing 0°, 45°, 90° and 135° polarization directions is created.

3. The method for detecting a wall damage structure based on polarization imaging according to claim 2, characterized in that, Wall surface damage image based on scale smoothing kernel for the polarization imaging performing convolution to obtain a base scale feature set ; A non-Gaussian anisotropic kernel aligned with the 90° polarization axis and odd symmetric about the vertical direction is constructed. The construction process first determines the central coordinate position of the kernel function, which is automatically calculated according to the kernel size parameter. The mathematical structure of the kernel function includes a linear coordinate component and an exponential decay component. The linear coordinate component produces an odd symmetric response characteristic in the vertical direction, and the exponential decay component realizes spatial decay control through a radial distance square term. The scale adaptability of the kernel function is realized by a width control factor, which maintains a quantitative relationship with the kernel size parameter; Collecting basic scale feature sets A non-Gaussian anisotropic kernel corresponding thereto Performing a convolution operation outputs an enhanced directional structure feature map set, the weight of each scale feature is determined through model training, and the enhanced directional structure feature map set and the weight of the scale feature are fused to output a multi-scale polarization structure enhanced feature map.

4. The method for detecting a wall damage structure based on polarization imaging according to claim 3, characterized in that, The wall surface damage structure images of different polarization directions are combined to obtain three polarization intensity images. According to the three polarization intensity images, the polarization angle feature is constructed by a nonlinear inverse tangent function operation, and the polarization degree feature is constructed by a root formula; Based on the polarization angle feature, the gradient in the horizontal and vertical directions is calculated to obtain the polarization angle gradient term; Based on the multi-scale polarization structure enhancement feature map, horizontal direction decomposition and vertical direction decomposition are performed to obtain horizontal direction decomposition features and vertical direction decomposition features. Based on the horizontal direction decomposition features and the vertical direction decomposition features, a feature matrix is constructed and the determinant is calculated. The local structure reconstruction tensor is obtained by combining the ratio of the polarization degree feature and the polarization angle gradient term.

5. The method for detecting a wall damage structure based on polarization imaging according to claim 4, characterized in that, Based on the polarization degree feature as a polarization constraint, a polarization geometric perception kernel of four polarization directions is constructed; Based on the polarization geometric perception kernel of four polarization directions, convolution operations are performed on the local structure reconstruction tensor to obtain damage response features in corresponding directions; The damage response features of four polarization directions are spliced in the channel dimension to form a fused damage response feature. The local energy mean of each channel based on the fusion damage response feature is dynamically constructed to form a normalization factor, an exponential decay function is used to map the response value of each channel to a probability interval in the numerator part, and the square root operation of the sum of the squares of the second and third polarization intensities is introduced to form a normalization term, and the damage area is generated through the probabilistic conversion.

6. The method for detecting a wall damage structure based on polarization imaging according to claim 5, characterized in that, The S2 step is represented as a non-Gaussian polarization feature extraction unit, which processes the wall damage image dataset obtained by polarization imaging to obtain a multi-scale polarization structure enhancement feature map; The S3 step is represented as a polarization feature fusion debugging unit, which performs polarization debugging on the multi-scale polarization structure enhancement feature map to obtain a local structure reconstruction tensor; The S4 step is represented as a multi-directional polarization geometric perception unit, which outputs the damage area by decision fusion based on the local structure reconstruction tensor damage response feature. Steps S2 to S4 construct a wall damage structure detection model based on polarization imaging.

7. The method of claim 6, wherein the method is a method of detecting a wall damage structure based on polarization imaging. The single-channel gray digital image dataset based on 0°, 45°, 90° and 135° polarization directions is input into the wall damage structure detection model based on polarization imaging, and the model parameters are updated through an iterative optimization training strategy, and finally the wall damage structure area is detected.

Images