A non-contact structure deformation detection method based on image processing

Abstract

The application belongs to the technical field of digital images, and specifically discloses a non-contact structure deformation detection method based on image processing, which comprises the following steps: acquiring a reference image in a no-load state, screening feature points and demarcating a target frame based on pixel average gray gradient values; acquiring a deformation image, determining feature point displacement vectors and a target frame overlap area ratio; outputting a positioning point similarity rate through the construction of a similarity evaluation model, dividing rigid displacement and abnormal mutation areas, and then warning and outputting results. The application does not require a target, combines image processing and machine learning to realize multi-point synchronous dynamic monitoring, is suitable for complex environments such as bridges and tunnels, and improves detection efficiency and accuracy, and can accurately identify structure deformation hazards.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of digital images, and particularly relates to a non-contact structural deformation detection method based on image processing. BACKGROUND

[0002] With the increase of time and the change of load conditions, large structures such as bridges and tunnels are prone to displacement, deformation and other damages, so accurate and efficient deformation detection is a core link to ensure the safe operation of bridges and tunnels.

[0003] The existing structural deformation detection generally adopts a contact measurement method. For example, a dial gauge, a level, a total station and the like, which need to set up a scaffold in advance or use an engineering vehicle, not only have a large labor intensity and a high detection cost, but also are limited by the detection environment. For special scenes such as a water-crossing bridge, a line-crossing bridge, a large-span bridge and a special-shaped tunnel, equipment is often unable to be erected, resulting in that detection is difficult to implement. At the same time, such a method has insufficient measurement capability for dynamic deformation of a structure, and is difficult to meet the demand of actual engineering for dynamic performance evaluation. In a non-contact measurement method, although a part of special deflection meters can realize high-precision measurement, they depend on installation of a mark with an LED lamp on a measured structure as a target, which increases the complexity of preliminary preparation work; and are limited by a field of view, and usually can only realize single-point measurement, which cannot meet the demand of multi-point synchronous dynamic monitoring, and is difficult to comprehensively reflect the overall deformation state of a structure. Therefore, there is an urgent need for a non-contact structural deformation detection method which does not depend on a target, is simple to operate, is suitable for complex environments and can realize multi-point dynamic monitoring. SUMMARY

[0004] The purpose of the present application is to overcome the defects in the prior art and provide a non-contact structural deformation detection method based on image processing.

[0005] The present application provides a non-contact structural deformation detection method based on image processing, comprising the following steps:

[0006] S1. Obtain a reference image of a structure in an unloaded state, screen final feature points based on average gray gradient values of each pixel point, and define a target frame in the reference image with the final feature points as the center;

[0007] S2. Obtain a deformation image of the structure under the action of a load, place the deformation image and the reference image in the same coordinate system, determine a displacement vector of the final feature points and a coincidence area ratio of the target frame before and after deformation;

[0008] S3. Construct a similarity evaluation model, input the displacement vector of the final feature points and the coincidence area ratio of the target frame before and after deformation into the similarity evaluation model, and output a similarity rate of each positioning point in the target frame after deformation;

[0009] S4. Selecting N positioning points in the target frame, obtaining similarity values of each positioning point and arranging them in descending order to form a similarity value set;

[0010] S5. Dividing rigid displacement area and abnormal mutation area based on the similarity value set;

[0011] S6. Making early warning judgment on the deformation state of the structure based on the rigid displacement area and the abnormal mutation area;

[0012] S7. Outputting early warning results and abnormal area positioning information.

[0013] Further, the process of screening the final feature points is:

[0014] Traverse the reference image at fixed intervals, and take each traversal point as a candidate feature point;

[0015] Set a fixed size, and expand a rectangular frame around the candidate feature point;

[0016] For each candidate feature point , calculate the gradient in the x and y directions and the gradient norm based on the gradient, and divide the sum of the gradient norms by the area of the rectangular frame to obtain the average intensity of the gray scale change of the entire rectangular frame ;

[0017] Set a gray scale change average intensity threshold , and retain the rectangular frame;

[0018] Take the candidate feature points of the remaining rectangular frames as the final feature points.

[0019] Further, in step S2, the displacement vector of the final feature point is obtained by solving an image deformation model through an iterative optimization algorithm;

[0020] The image deformation model is used to describe the motion of the image target frame centered on the final feature point, and the parameter vector of the image deformation model includes the translation amount of the final feature point and the shape change amount of the target frame;

[0021] The iterative optimization algorithm adjusts the parameter vector to minimize the gray intensity difference between the image target frame before and after deformation;

[0022] The displacement vector of the final feature point is composed of the translation component in the parameter vector after iterative optimization.

[0023] Further, the image deformation model is a first-order shape function model, and the parameter vector to be solved is where u, v are the translation of the final feature point, , represents normal strain, , represents shear strain or rotation.

[0024] Further, the iterative optimization algorithm comprises the following steps:

[0025] a) Given an initial estimate of the image deformation model parameter vector ; ;

[0026] b) Under the current parameter estimate , calculate the objective function F( ) reflecting the intensity difference, and calculate the gradient vector and Hessian matrix H( ) of the objective function at ;

[0027] c) Construct a linear equation system , and solve the equation system to obtain the parameter increment Δ ;

[0028] d) Update the parameter vector: = + Δ ;

[0029] e) Repeat steps b-d until the modulus of the parameter increment Δ is less than the preset tolerance or the change of the objective function value F( ) is less than the preset value, at which time the optimized parameter vector is obtained;

[0030] f) Extract the component representing the translation from the optimized parameter vector , and take it as the displacement vector of the final feature point.

[0031] Further, the similarity evaluation model adopts a similarity evaluation function constructed based on the objective function F( );

[0032] Based on the objective function F( ), the objective function values of the image target frames before and after deformation and their gradient information are calculated to generate numerical features for evaluating similarity;

[0033] The numerical features, the displacement vector of the final feature point, and the coincidence area ratio of the target frames before and after deformation are combined to serve as the input feature vector of the similarity evaluation model.

[0034] Further, the construction process of the similarity evaluation model is:

[0035] Collect sample data sets of the plurality of structures under different load conditions, the sample data sets including the overlap area ratio of the target frame before and after deformation and the displacement vector of the final feature point before and after deformation;

[0036] Use the objective function F( ) to process the image target frame before and after deformation, and extract the intermediate numerical results or final numerical results generated by the objective function F( ) in the process of solving the displacement vector as numerical features;

[0037] The numerical features at least include: the function value of the objective function F( ) under the optimal parameters or the norm of the gradient vector of the objective function F( ) with respect to the deformation parameters or at least one of the Hessian matrix of the objective function F( ) with respect to the deformation parameters;

[0038] Record the overlap area ratio of the target frame before and after deformation and the displacement vector of the final feature point before and after deformation corresponding to the numerical features;

[0039] Mark the first mapping relationship between the overlap area ratio of the target frame before and after deformation and the similarity rate, and mark the second mapping relationship between the displacement vector of the final feature point before and after deformation and the similarity rate by artificial experts;

[0040] Based on the first mapping relationship and the second mapping relationship, the overlap area ratio of the target frame before and after deformation, the displacement vector of the final feature point before and after deformation, and the numerical features extracted from the objective function F( ) are input into the neural network model for iterative training to obtain a similarity evaluation model with the input of the overlap area ratio of the target frame before and after deformation, the displacement vector of the final feature point before and after deformation, and the numerical features, and the output of the similarity rate.

[0041] Further, in the step S5, the positioning points corresponding to the first k high similarity values are divided into a maximum vector group, and the positioning points corresponding to the last m low similarity values are divided into a minimum vector group;

[0042] The first k positioning points are sequentially connected to form a rigid displacement area, and the last m positioning points are sequentially connected to form an abnormal mutation area;

[0043] The rigid displacement area is used to define the reference range of normal deformation and filter overall motion interference;

[0044] The abnormal mutation region is used to define the spatial characteristics of local abnormalities and structural deformation hazard identification.

[0045] Further, in the step S6, the process of pre-warning judgment on the structural deformation state is:

[0046] The deformation amount of the rigid displacement region is calculated as the reference deformation amount of the structure;

[0047] The deformation amount of each positioning point in the abnormal mutation region is calculated;

[0048] The deformation amount threshold of the abnormal mutation region is set based on the reference deformation amount of the structure;

[0049] The system traverses each positioning point in the abnormal mutation region, compares its deformation amount with the deformation amount threshold, and pre-warns the part exceeding the deformation amount threshold.

[0050] Further, the pre-warning result in the step S7 includes at least one of pre-warning level, abnormal region position coordinates or deformation amount value;

[0051] The abnormal region positioning information includes the boundary coordinates and the region center position coordinates of the abnormal mutation region.

[0052] Compared with the prior art, the present application has the following advantages:

[0053] The present application realizes non-contact, multi-point synchronous dynamic monitoring of structural deformation by combining image processing and machine learning, and through comprehensive analysis of displacement vector and image coincidence degree and other information, a similarity evaluation model is constructed to realize full-process automatic detection from image acquisition to pre-warning output.

[0054] The present application can automatically screen out points with significant texture features and easy matching as final feature points by calculating the average gray gradient intensity of the fixed rectangular region around the candidate points, effectively avoiding errors caused by matching in texture-poor areas, and ensuring the basic quality of subsequent displacement calculation.

[0055] The present application combines intermediate numerical features generated during displacement calculation with displacement vector and coincidence area ratio to reflect matching quality from multiple dimensions, and through machine learning model training, can learn the deep mapping relationship between matching quality and multiple features under complex working conditions, so that the similarity evaluation model has stronger generalization ability and higher judgment accuracy, and can effectively identify potential error points. Attached Figure Description

[0056] The following figures are for illustrative purposes only and are not intended to limit the scope of the invention, wherein:

[0057] Figure 1 : Schematic diagram of the detection method of the present invention;

[0058] Figure 2 : Schematic diagram of image deformation and displacement vector;

[0059] Figure 3 Flowchart for anomaly assessment. Detailed Implementation

[0060] To make the objectives, technical solutions, design methods, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0061] This invention provides a non-contact structural deformation detection method based on image processing, which is used to monitor the displacement of multiple target points on large structures such as bridges, tunnels, and buildings in real time, and can realize long-distance, multi-point dynamic observation.

[0062] like Figure 1 As shown, it includes the following steps:

[0063] S1. Use a high-definition camera to capture reference images of large structures such as bridges, tunnels, and buildings under no-load conditions. Extract the gray-level gradient values ​​of each pixel in the image through image processing algorithms, calculate the average gray-level gradient of the area around each pixel, set a threshold to filter out the final feature points, and delineate a target box of a fixed size in the reference image with each final feature point as the center.

[0064] S2. After the load is applied to the structure, take a picture of its deformation, align the deformation picture with the reference picture, calculate the displacement vector of each final feature point through image registration technology, and calculate the area ratio of the overlapping area of ​​each target box before and after deformation.

[0065] S3. Construct a similarity evaluation model based on machine learning, taking the ratio of the displacement vector of each feature point to the overlapping area of ​​its corresponding target box as the input feature, and outputting the similarity rate of each localization point within the target box;

[0066] S4. N localization points are uniformly selected within each target box, and sorted in descending order according to the similarity rate output by the similarity evaluation model to form a set of similarity values;

[0067] S5. According to the distribution of the similarity value set, the positioning points corresponding to the first k high similarity rates are divided into a rigid displacement region, and the positioning points corresponding to the last m low similarity rates are divided into an abnormal mutation region;

[0068] S6. A reference deformation range is established according to the deformation amount of the rigid displacement region, and threshold value judgment is performed on the deformation amount of each point in the abnormal mutation region, and if the threshold value is exceeded, a pre-warning is triggered;

[0069] S7. An output is performed on a pre-warning level, position coordinates of an abnormal region, a deformation amount value and boundary information thereof.

[0070] In step S1, the process of screening the final feature points is as follows: in the reference image, a fixed step is traversed in the reference image, each traversed point is taken as a candidate feature point, a rectangular frame with a fixed size is constructed with each candidate point as the center, the gradients of all pixels in the rectangular frame in the x and y directions are calculated through the Sobel operator, and then the gradient norm of each pixel is calculated, and the average gray level change intensity is obtained after summation and division by the area of the rectangular frame. A threshold value is set, and only the candidate points with are retained as the final feature points. The calculation formula of the average gray level change intensity is as follows:

[0071] ;

[0072] wherein w and h are the width and height of the target frame in the image;

[0073] is the gray level gradient norm of the candidate feature point in the X direction and the Y direction;

[0074] To achieve high-precision structural deformation measurement, in step S2, a first-order shape function is used as an image deformation model, and the model parameters are solved through an iterative optimization algorithm to extract the displacement vector of the final feature point. As shown in Figure 2 , the gray level information of the feature point in the reference image under the no-load state is , and the gray level information of the feature point in the reference image under the load state is .

[0075] The specific implementation process is as follows:

[0076] A first-order shape function model is established to describe the motion behavior of the image target frame centered on the final feature point during the deformation process. The parameter vector of the model is defined as , wherein u and v represent the translation amounts of the feature point in the x and y directions, , Indicates positive strain. , It represents shear strain or rotation, thus enabling comprehensive capture of the local deformation characteristics of the target area.

[0077] The parameter vector in this model is solved using an iterative optimization algorithm. The algorithm begins with an analysis of the parameter vector. Assign initial estimate In each iteration, based on the current parameter estimate... Calculate the objective function F( ) used to measure the grayscale difference between the reference image and the deformed image. ), and further solve for the gradient vector ∇F( of the objective function at the current position. ) and Hessian matrix H( Next, construct a system of linear equations H( )Δ =- ∇F( The parameter increment Δ is obtained by numerical methods. The parameter vector is updated using this increment, i.e. = +Δ Repeat this process until the magnitude of the parameter increment is less than the preset tolerance or the change in the objective function value tends to stabilize. At this point, the algorithm is considered to have converged, and the optimized parameter vector is obtained. .

[0078] From the optimized parameter vector The translational motion components u and v are extracted and used as the displacement vectors of the final feature points. This method achieves sub-pixel level displacement calculation accuracy, significantly improving the accuracy of deformation measurement and the ability to detect minute deformations. Furthermore, the optimization process is stable and computationally efficient, making it suitable for real-time or near-real-time processing requirements in practical engineering.

[0079] In this embodiment, to achieve accurate evaluation of image matching quality during structural deformation, a similarity evaluation model for outputting a similarity rate reflecting matching reliability is constructed in step S3. This similarity evaluation model comprehensively utilizes numerical features, displacement vectors, and regional overlap information during deformation calculation, ultimately outputting a similarity rate that reflects matching reliability. The specific implementation process is as follows:

[0080] The model's basic input features include the final feature point displacement vectors obtained directly from image registration, the ratio of overlapping regions before and after bounding box deformation, and deep numerical features derived from the iterative solution of the displacement vectors. These numerical features are processed through the objective function F( ) when processing the target frame before and after the deformation, including: the final function value of the objective function under the optimal parameter, the norm of the gradient vector of the deformation parameter, and the eigenvalue or norm of the Hessian matrix, etc. These features reflect the convergence, local sensitivity and reliability of the matching from different angles, and together with the displacement and coincidence degree information form a comprehensive feature vector as the input of the similarity evaluation model.

[0081] A large number of sample image data covering various structure types such as bridges and tunnels and different load conditions are required to be collected, each sample containing the target frame overlap area ratio, the final displacement vector of the feature point, and the numerical features calculated by the objective function F( ). According to the prior knowledge, the technical experts label the samples to establish the first mapping relationship between the overlap area ratio and the true similarity rate, and the second mapping relationship between the displacement vector and the true similarity rate. On this basis, the numerical features, displacement vectors and overlap area ratios are jointly used as inputs, and the similarity rate labeled by experts is used as the target output, which is input into the neural network model for supervised iterative training. The trained model can learn the nonlinear mapping relationship from complex multidimensional features to the final similarity rate. When processing new detection images, the system automatically extracts the above features and forms a feature vector, which is input into the model to obtain the similarity rate prediction value of each positioning point. Through the fusion of deep numerical information in the optimization process, the judgment ability of the matching quality is significantly enhanced, and the model has excellent generalization performance, which can adapt to complex and variable actual working conditions, greatly improving the accuracy and reliability of the deformation detection results.

[0082] In this embodiment, after the similarity rate of each positioning point is calculated and sorted, the deformation area is analyzed and warned, and the specific implementation process is as follows:

[0083] According to the similarity rate set of N positioning points in each target frame, the N positioning points are arranged in descending order from high to low in value, the first k positioning points with the highest similarity rate are classified into a maximum vector group, representing the part of the structure body that occurs synergistic and consistent deformation, and the last m positioning points with the lowest similarity rate are classified into a minimum vector group, representing the area that is likely to occur abnormal deformation. Wherein, k and m represent a proportion of the total number N, for example, in the embodiment, k = 0.3N and m = 0.15N, k and m are rounded up respectively. In the spatial distribution, the first k positioning points are sequentially connected to form a rigid displacement area, which is used to define the spatial reference range of the normal deformation of the structure body under the action of the load, and effectively filters out the measurement errors caused by the overall rigid body displacement or environmental interference; correspondingly, the last m positioning points are sequentially connected to form an abnormal mutation area, which is used to accurately define the spatial distribution characteristics of the local deformation anomaly, and provides a key target for identifying potential damage or hidden danger of the structure. On this basis, the system intelligently warns and judges the deformation state of the structure body, takes the rigid displacement area as the reference, calculates the average deformation of all positioning points in the area as the reference deformation of the entire structure under this working condition. The actual deformation of each positioning point in the abnormal mutation area is calculated one by one, the deformation threshold of the abnormal area is automatically set according to the reference deformation, and the deformation of each positioning point in the abnormal mutation area is compared with the threshold in real time. Once the deformation of a certain point exceeds the threshold, the warning judgment is triggered immediately.

[0084] In the above, the result of the warning judgment is converted into specific output information, and the warning result not only includes the warning level, but also includes the accurate abnormal area position coordinates, the specific deformation value, and the detailed boundary coordinates and area center position coordinates of the abnormal mutation area. This greatly facilitates engineers to quickly locate the problem area, assess the risk level, and develop corresponding maintenance and reinforcement strategies, thereby realizing the full-process automation from image acquisition to intelligent diagnosis to precise warning, and significantly improving the efficiency and reliability of structural health monitoring. In the embodiment, since the reference image and the deformation image have been placed in the same coordinate system in step S2, each pixel point in the image corresponds to a unique coordinate value (x, y), so when calculating the coordinates of the final feature points and the positioning points, it is based on the unified coordinate system. In step S4, N positioning points are uniformly selected in the target frame centered on each final feature point, that is, the coordinates of each positioning point have been determined, and the coordinates of these positioning points include the abnormal area position coordinates in the output information; when calculating the deformation value, the first-order shape function model is solved by an iterative optimization algorithm to obtain the displacement vector of the final feature point, and the shape function model not only gives the translation of the final feature point, but also describes the displacement of any point in the target frame. For a certain positioning point (x, y) in the target frame, its displacement vector can be obtained by the following formula:

[0085] = + ;

[0086] = + ;

[0087] in, Let x be the component of the displacement vector of the positioning point (x, y) in the x-direction. Let y be the component of the displacement vector of the positioning point (x,y) in the y-direction;

[0088] The deformation value of the positioning point (x, y) is: ;

[0089] in,( , (x, y) represents the coordinate offset of the positioning point (x, y) relative to the center of the target box;

[0090] In the above, the first m positioning points in the abnormal mutation region are connected in sequence. These positioning points form a polygonal region after being connected. The vertex coordinates of this polygon are the boundary coordinates. The center position coordinates of the region are obtained by extracting the centroid of this polygonal region.

[0091] In this embodiment, step S7 is further used to output an anomaly severity assessment value. The anomaly severity assessment value is a quantitative assessment of the overall risk level of the anomalous mutation region. Specifically, as shown... Figure 3 The flowchart for anomaly assessment shown includes the following calculation process:

[0092] Input all localization points of the abnormal mutation region and perform parallel computation:

[0093] (a) The absolute deformation of each abnormal location point, that is, the deformation value of any location point (x,y);

[0094] (b) Based on the deformation data of the rigid displacement region, calculate its reference deformation, which is the average value of the deformation of all positioning points in the rigid displacement region.

[0095] For each abnormal location point, the relative difference between its absolute deformation and the reference deformation is calculated to reflect the degree of deviation of the abnormal point from the normal deformation of the overall structure. Specifically, the relative difference is calculated as follows: take the difference between the absolute deformation of an abnormal point and the reference deformation, divide it by the reference deformation, and then normalize it.

[0096] The calculated relative difference value is compared with a preset fixed threshold value to determine whether the deviation is significant; if the relative difference value is greater than the threshold value, the point is marked as an abnormal point and the exceeding value is calculated, otherwise the point is marked as normal;

[0097] The exceeding values of all the marked abnormal points are summed to obtain the final abnormality degree evaluation value, and the greater the abnormality degree evaluation value, the greater the exceeding amount of the abnormal mutation region relative to the overall deformation, and the higher the potential risk of the structure.

[0098] The above has described the embodiments of the present application, the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles, practical applications, or technical improvements in the market of the embodiments, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

Claims

1. An image processing-based non-contact structure deformation detection method, characterized by, The method comprises the following steps: S1. Obtain a reference image of the structure in an unloaded state, screen final feature points based on average gray gradient values of each pixel point, and define a target frame in the reference image with the final feature points as the center; S2. Obtain a deformation image of the structure under the action of the load, place the deformation image and the reference image in the same coordinate system, determine displacement vectors of the final feature points and a coincidence area ratio of the target frame before and after deformation; S3. Construct a similarity evaluation model, input the displacement vectors of the final feature points and the coincidence area ratio of the target frame before and after deformation into the similarity evaluation model, and output a similarity rate of each positioning point in the target frame after deformation; S4. Select N positioning points in the target frame, obtain similarity values of the positioning points and arrange the similarity values in descending order to form a similarity value set; S5. Divide a rigid displacement area and an abnormal mutation area based on the similarity value set; S6. Perform early warning judgment on the deformation state of the structure based on the rigid displacement area and the abnormal mutation area; S7. Output early warning results and abnormal area positioning information. The process of screening the final feature points is as follows: traverse the reference image at fixed intervals, take each traversal point as a candidate feature point; set a fixed size, expand a rectangular frame around the candidate feature point with the candidate feature point as the center; for each candidate feature point calculating the gradient in x, y direction and gradient norm based on the gradient, summing the gradient norm and dividing by the area of the rectangular frame to get the average intensity of the gray scale change of the whole rectangular frame ; Setting a gray scale change average intensity threshold , retaining a rectangular frame; take the candidate feature points of the remaining rectangular frames as the final feature points.

2. The non-contact structural deformation detection method based on image processing according to claim 1, wherein, In the step S2, the displacement vectors of the final feature points are obtained by solving an image deformation model through an iterative optimization algorithm; The image deformation model is used to describe the motion of the image target frame centered on the final feature point, and the image deformation model parameter vector The translation amount of the final feature point and the shape change amount of the target frame are included. The iterative optimization algorithm minimizes the difference between the gray intensity of the image target frame before and after deformation by adjusting the parameter vector the gray intensity of the image target frame before and after deformation The displacement vector of the final feature point is composed of the translation component in the iteration-optimized parameter vector .

3. The non-contact structural deformation detection method based on image processing according to claim 2, wherein, The image deformation model is a first-order shape function model, and a parameter vector to be solved is where u, v are translation amounts of the final feature points, 、 denotes normal strain, 、 denotes shear strain or rotation.

4. The non-contact structural deformation detection method based on image processing according to claim 3, wherein, The iterative optimization algorithm comprises the following steps: a) Given the image deformation model parameter vector initial estimate ; b) in the current parameter estimate Next, a target function F( ) for reflecting the difference in the gray scale intensity is calculated, and a gradient vector ∇F( ) and a Hessian matrix H( ) of the target function at the point x are calculated. c) constructing a system of linear equations H )Δ = -∇F ), and solving said system of equations for the parameter increment Δ ; d) updating the parameter vector: = + Δ ;​​ e) repeat steps b-d until the modulus of the parameter increment Δ is less than a preset tolerance or the change in the objective function value F( ) is less than a preset value, at which point the optimized parameter vector is obtained; f) extracting from the optimized parameter vector a component representing the translation amount as a displacement vector of the final feature point.

5. The non-contact structural deformation detection method based on image processing according to claim 4, wherein, The similarity evaluation model adopts a similarity evaluation function constructed based on an objective function F( ). Based on the target function F( ), the target function value of the image target frame before and after deformation and its gradient information are calculated to generate numerical characteristics for evaluating the similarity; combine the numerical features, the displacement vectors of the final feature points and the coincidence area ratio of the target frame before and after deformation, and use them as input feature vectors of the similarity evaluation model.

6. The non-contact structural deformation detection method based on image processing according to claim 5, wherein The construction process of the similarity evaluation model is as follows: collect sample data sets of various structures under different load conditions, the sample data sets comprising the coincidence area ratio of the target frame before and after deformation and the displacement vectors of the final feature points before and after deformation; using the target function F( ) to process the image target frame before and after deformation, and extracting intermediate numerical results or final numerical results generated by the target function F( ) in the process of solving the displacement vector as the numerical features; The numerical features at least include: a function value of the objective function F( ) at the optimal parameters or a norm of a gradient vector of the objective function F( ) with respect to the deformation parameters or at least one of a Hessian matrix of the objective function F( ) with respect to the deformation parameters. record the coincidence area ratio of the target frame before and after deformation and the displacement vectors of the final feature points before and after deformation corresponding to the numerical features; mark a first mapping relationship between the coincidence area ratio of the target frame before and after deformation and the similarity rate and a second mapping relationship between the displacement vectors of the final feature points before and after deformation and the similarity rate through artificial experts; The ratio of the coincident regions before and after deformation of the target frame, the displacement vectors of the final feature points before and after deformation, and the numerical features extracted from the target function F( ) are input into a neural network model for iterative training based on the first mapping relationship and the second mapping relationship, to obtain a similarity evaluation model with the input being the ratio of the coincident regions before and after deformation of the target frame, the displacement vectors of the final feature points before and after deformation, and the numerical features, and the output being the similarity rate.

7. The non-contact structural deformation detection method based on image processing according to claim 6, wherein, In the step S5, divide the positioning points corresponding to the first k high similarity values into a maximum vector group and the positioning points corresponding to the last m low similarity values into a minimum vector group; sequentially connect the first k positioning points to form a rigid displacement area and sequentially connect the last m positioning points to form an abnormal mutation area in spatial distribution; the rigid displacement area is used to define a reference range of normal deformation and filter overall motion interference; the abnormal mutation area is used to define spatial features of local abnormalities and structure deformation hazard identification.

8. The non-contact structural deformation detection method based on image processing according to claim 7, wherein, In the step S6, the process of performing early warning judgment on the deformation state of the structure is as follows: calculate a deformation amount of the rigid displacement area as a reference deformation amount of the structure; calculate a deformation amount of each positioning point in the abnormal mutation area; set a deformation amount threshold of the abnormal mutation area based on the reference deformation amount of the structure; The system traverses each positioning point in the abnormal mutation region, compares the deformation amount with the deformation amount threshold, and gives a warning for the part exceeding the deformation amount threshold.

9. The non-contact structural deformation detection method based on image processing according to claim 8, wherein, The warning result in the step S7 includes at least one of a warning level, an abnormal region position coordinate or a deformation amount value; The abnormal region positioning information includes boundary coordinates and region center position coordinates of the abnormal mutation region.

Images