System for visual analysis of mega-sphere architectural images

Abstract

The application discloses a kind of visual analysis systems for giant ball building appearance defect detection, to solve the problems of low efficiency, incomplete coverage, poor accuracy of traditional image recognition, and risk disposal without standardization.The system dynamically optimizes the shooting distance, flight speed and shooting frequency through the unmanned aerial vehicle in two stages of ascending and horizontal, to obtain clear original images of the building surface.After preprocessing such as filtering, denoising, distortion correction and reflection detection by the visual analysis module, cracks, corrosion and deformation defects are detected.The risk disposal module divides the risk level according to the defect characteristics and automatically matches the resources, and after repair, the unmanned aerial vehicle is used for re-inspection to verify whether the acceptance threshold is met, and a qualified signal is generated if the threshold is met.The system is optimized for the curved surface structure of giant ball building, significantly improves the detection efficiency and defect identification accuracy, and provides full-process automation support for the operation and maintenance of giant ball building.

Description

Technical Field

[0001] This invention relates to the field of computer vision and building structure monitoring technology, specifically a visual analysis system for giant spherical buildings, which addresses key technical issues such as spherical geometric distortion correction, defect risk assessment, and dynamic classification. Background Technology

[0002] Giant spherical structures (such as large spherical storage tanks and spherical dome buildings) are characterized by complex curved structures, large surface areas, and special inspection environments. Manual inspection of their appearance defects (such as cracks, corrosion, and deformation) suffers from low efficiency, high cost, and incomplete coverage. Traditional image acquisition methods are prone to insufficient image clarity and severe distortion due to improper shooting distance and angle. Furthermore, the lack of dedicated preprocessing and fusion algorithms for spherical curved surfaces in post-processing leads to low defect identification accuracy. Simultaneously, the lack of standardized procedures for defect risk level classification and resource allocation hinders efficient end-to-end inspection and repair.

[0003] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0004] The purpose of this invention is to provide a visual analysis system and method for images of giant spherical buildings, which solves the problems of poor image acquisition quality, low defect identification accuracy, and low risk handling efficiency in the prior art. It realizes the full-process automated management from image acquisition to defect repair, and improves detection efficiency and reliability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A visual analysis system for images of giant spherical buildings includes an image acquisition module, a visual analysis module, and a risk classification and resource disposal module;

[0007] The image acquisition module dynamically adjusts the drone's shooting distance, flight speed, number of shots, and ascent altitude to obtain clear, original image data of the giant spherical building's surface; the acquired information is then sent to the visual analysis module.

[0008] The visual analysis module is used to receive the raw image data from the acquisition module, preprocess the raw image data, including filtering and denoising, distortion detection, and reflection detection; detect the values ​​corresponding to the parameters in the preprocessed image; if the values ​​are not within the preset threshold, perform defect detection on the image to obtain the type of defect; classify the defect type and send it to the risk classification and resource disposal module.

[0009] The risk classification and resource disposal module is used to receive defects from the visual analysis module, classify the defects by risk level, allocate corresponding resources according to the risk level, and finally confirm whether the repair effect meets the acceptance threshold through drone re-inspection, and generate a qualified signal.

[0010] In a preferred embodiment of the present invention, the image acquisition module operates as follows:

[0011] When adjusting the drone's shooting distance, the sharpness index C is determined by referring to resolution N, contrast J, sharpness T, brightness U, and pixel density G; according to the formula... ;in These are weighting coefficients used to balance the impact of each parameter on the final sharpness; , , , and These are the baseline values ​​for resolution, contrast, sharpness, brightness, and pixel density, respectively; the sharpness index C is greater than the threshold. At that time, increase the shooting distance until the resolution index is reached. , This represents the margin for the sharpness index; if it falls below the threshold, the distance is shortened until the sharpness index exceeds the threshold.

[0012] When determining the drone's flight speed and the number of shots, the actual effective width of a single image, ϵ×W×(1-p), is calculated based on the image pixel width W, the lens visual conversion coefficient ϵ, and the preset overlap rate e. The number of shots, f, is then determined by combining this with the circumference L of the horizontal orbital trajectory. ;

[0013] Flight speed is determined by the blur level X based on the blur pixel offset A, dynamic contrast decay rate F, motion blur direction entropy O, temporal resolution matching degree P, and motion blur frequency response decay value Y, according to the formula: X = ,in This is a weighting coefficient; when the ambiguity is below a threshold, the speed is increased until the ambiguity decreases. , This is the margin for ambiguity. When the ambiguity is greater than the threshold, the speed is reduced until the ambiguity is less than the threshold.

[0014] The vertical pixels of the image captured by the camera at the rising height To determine, the actual physical vertical length covered by a single image on the surface of the giant spherical building is B = ϵ × The actual vertical length of a single image and the height between layers are B×(1-e);

[0015] In a preferred embodiment of the present invention, the visual analysis module operates as follows:

[0016] The filtering and denoising process calculates noise based on the pixel dimension and uses a 3×3 neighborhood median filter, replacing the current pixel value with the median of the neighboring pixels. The filtering intensity is adaptively adjusted: when the value of i increases, the window is expanded or the iteration is increased to improve the filtering intensity; when the value of i decreases, the number of filtering iterations is reduced to weaken the filtering intensity. Here, i is the noise evaluation index, i=U / U0, where U is the noise standard deviation of the current region image and U0 is the preset baseline noise standard deviation.

[0017] Distortion detection determines whether distortion exists by using the polar stretching coefficient and angular distortion rate, and then by considering the projection center deviation (Δζ, Δη) and radius coefficient. Correction is performed in terms of the angular deformation coefficient Δρ;

[0018] Projection center deviation correction: Calculate the deviation (Δζ, Δη) using three feature points, measure the actual coordinates (ζ, η) of the corresponding points in the image, and then perform a translation correction on the coordinates. The correction formula is as follows: ; ;

[0019] Projection radius coefficient correction: Calculate the actual projected radius q of each of the five feature points, then calculate the theoretical geocentric angle t; the ratio is... The theoretical coefficient under ideal equal area conditions By comparison, the deviation value Δm is determined. Then, scale q for all projected points proportionally, and the corrected formula is: ;

[0020] Angular distortion coefficient correction: Using a feature line with a known included angle, measure the actual included angle θ in the projected image, and calculate... Where ρ is the theoretical included angle of 90°, the rotation correction formula for each pixel is: ;

[0021] Reflectivity detection is based on average grayscale value and peak brightness. When the average grayscale value of the image is greater than 100%, the reflection is detected. And peak brightness > At that time, determine the image reflection;

[0022] When there is reflection, the drone is used to re-capture the reflection areas at different locations. When there is reflection at the top, the flight altitude is reduced so that the angle between the camera optical axis and the normal is 30°. When there is reflection at the side, the drone is moved 15° up, down, left and right to take a total of 8 images, and the image with the highest clarity index is selected. When there is reflection at the bottom, the drone is controlled to move along the bottom circumference of the giant spherical building surface and the shooting angle is changed.

[0023] In a preferred embodiment of the present invention, the visual analysis module further includes:

[0024] Crack detection: Detect the product ratio Z of the length and depth of crack pixels in the image, that is, the ratio of the actual length Q of the crack to the allowable length multiplied by the ratio of the actual depth D of the crack to the allowable depth . Normalize the allowable length and allowable depth of the crack. Through the formula: Z = ; When Z > 0.01, it is determined as a crack;

[0025] Rust detection: Adopt integrated color space conversion, screen areas with hue H = 8° - 20°, saturation S ≥ 0.8 and color difference ΔE ≤ 8, and determine it as rust;

[0026] Deformation detection: Based on the standard template, detect and determine using the similarity index NCC; Through the formula , y is the pixel gray value of the template image, is the average gray value of the template pixels, v is the pixel gray value of the window image, is the average gray value of the window image pixels; Windows with NCC values < 0.8 are marked as deformation areas.

[0027] As a preferred implementation manner of the present invention, the specific operation process of the risk classification and resource disposal module is as follows:

[0028] Crack: Divide according to the product ratio Z of the length and depth of crack pixels. 0.01 < Z ≤ 0.025 is low risk, 0.025 < Z ≤ 0.1 is medium risk, and Z > 0.1 is high risk;

[0029] Rust: Divide according to the percentage of rust pixels in the total pixels of the image. The proportion of rust pixels ≤ 5% is low risk, 5% < the proportion of rust pixels ≤ 15% is medium risk, and the proportion of rust pixels > 15% is high risk;

[0030] Deformation: Divide according to the NCC value of the similarity index. 0.8 ≤ NCC < 1.0 is low risk, 0.6 ≤ NCC < 0.8 is medium risk, and NCC < 0.6 is high risk.

[0031] As a preferred implementation manner of the present invention, the risk classification and resource disposal module further includes:

[0032] In terms of resource disposal, adopt corresponding deployment plans for different risk levels:

[0033] Low-risk defects are responsible for by a basic team of 2 - 3 people, equipped with conventional detection tools such as cameras and tape measures;

[0034] Medium-risk defects are handled by a professional team of 4 - 6 people, carrying special repair equipment such as crack repair tools and rust cleaning equipment, and the on-site disposal needs to be completed within 7 working days;

[0035] For high-risk defects, a technical expert team of 6-8 people, including structural engineers and materials experts, will be dispatched and required to arrive on site within 48 hours. Resource mobilization will prioritize the nearest resource pool and be allocated according to the shortest time path, without being restricted by economic cost factors.

[0036] During the repair and acceptance process, acceptance thresholds are defined according to the defect type: cracks must meet the length-to-depth product ratio Z≤0.01, corrosion must meet the pixel ratio ≤1%, and deformation must meet the similarity index NCC≥0.9. After confirming compliance through drone re-inspection, a qualified signal is generated.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. More precise image acquisition: By dynamically adjusting the drone's shooting distance, speed, number of shots, and ascent altitude, clear and complete images of the giant spherical building's surface can be obtained, providing high-quality data for subsequent analysis;

[0039] 2. More accurate defect detection: Through pretreatment and specialized detection methods, the type and extent of defects such as cracks, corrosion, and deformation can be accurately identified, reducing errors;

[0040] 3. The risk classification is reasonable and the resource allocation is optimized. It can match the corresponding teams and equipment according to different risk levels, improve the efficiency of handling, and ensure that the repair meets the standards through re-inspection.

[0041] 4. Adaptable to the curved structure and other features of giant spherical buildings, it can achieve all-round, high-precision monitoring, providing strong support for the operation and maintenance of large, irregularly shaped buildings. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to illustrate the main idea of ​​the present invention.

[0043] Figure 1 This is the overall system block diagram of the present invention. Detailed Implementation

[0044] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0045] like Figure 1 As shown, a visual analysis system for images of giant spherical buildings includes an image acquisition module, a visual analysis module, and a risk classification and resource disposal module.

[0046] The image acquisition module uses a drone to acquire original images of the building's exterior via a combination of horizontal circling and vertical ascent. During the acquisition process, the shooting distance is adjusted according to the sharpness index, the flight speed is adapted according to the blur parameter, and the number of shots is determined in combination with the pixel width to finally obtain a clear original image that meets the requirements.

[0047] The visual analysis module preprocesses the acquired raw images, specifically including: first, filtering and denoising the images; then, performing distortion detection (based on polar stretching coefficient and angular distortion rate) and correction on the denoised images; performing reflection detection (based on average gray value and peak brightness) on the distorted images, and re-acquiring images that are determined to have reflections; and finally, using a multi-view fusion algorithm to fuse the preprocessed images.

[0048] After preprocessing, a dedicated algorithm is used to detect cracks (based on dilated convolutional chains and morphological features), corrosion (based on color space thresholds and color difference analysis), and deformation (based on template similarity NCC index) in parallel.

[0049] The risk classification and resource disposal module classifies the risk level according to the defect characteristics (cracks are classified according to the length-depth product ratio, corrosion is classified according to the pixel ratio, and deformation is classified according to the NCC value). Based on the risk level, it allocates corresponding resources (basic team, professional group, expert team) to plan routes. Finally, it uses drones to re-inspect and confirm whether the repair effect meets the acceptance threshold, and generates a qualified signal.

[0050] The image acquisition module uses drones to collect raw image data of the giant spherical building's exterior, specifically including:

[0051] The flight path of a drone is divided into two phases: ascent and horizontal.

[0052] During the horizontal phase, the quality of data collection needs to be ensured by adjusting the drone's shooting distance, flight speed, and number of shots; the specific steps are as follows:

[0053] The drone's shooting distance is adjusted by analyzing image sharpness metrics, including resolution N, contrast J, sharpness T, brightness U, and pixel density G.

[0054] Let the actual image sharpness index be C, and the preset sharpness index threshold be... According to the formula: ;in These are weighting coefficients used to balance the impact of each parameter on the final sharpness; , , , and These are the baseline values ​​for resolution, contrast, sharpness, brightness, and pixel density, set based on mainstream camera parameters and the scene of giant spherical building inspection.

[0055] when If the image sharpness is deemed sufficient for the application, the shooting distance is increased until the desired sharpness is achieved. , This represents a margin (5%–10%) for the sharpness index; when At this time, it is necessary to shorten the shooting distance to improve the sharpness index, until... ;

[0056] Determining the flight speed and the number of shots:

[0057] Based on the camera parameters, the pixel width W of the image is known. When the drone flies around the surface of the giant spherical building at different heights, it will form a corresponding circumference L. The actual physical width covered by a single image on the surface of the giant spherical building is ϵ×W, where ϵ is the lens visual conversion coefficient; it is the calibration coefficient (unit: mm / pixel) of the camera lens to convert the pixel size into the actual physical size. It is calculated by taking pictures of a standard calibration board of known size (such as 200mm×200mm) and the ratio of the actual size to the pixel size. The calibration error is ≤±0.01mm / pixel. Before the drone takes off, the ϵ value is determined by a ground calibration experiment: the calibration board is fixed on a plane with the same material as the surface of the giant spherical building, the drone takes pictures at a preset distance (such as 5m), the pixel size of the calibration board in the image is measured, and the ϵ value is calculated.

[0058] The default overlap rate between adjacent images is p (p=20%); the actual effective width of a single image is ϵ×W×(1-p); the number of shots is f= The calculation result is rounded up to ensure complete coverage;

[0059] The flight speed of a drone is determined by the blurriness of images captured during flight, specifically including:

[0060] By analyzing the blur parameters of the image, including:

[0061] Motion blur pixel offset A: The pixel offset caused by motion blur (unit: pixels), calculated by the image edge detection algorithm;

[0062] Dynamic contrast decay rate F: The ratio of image contrast under motion to static contrast (value 0-1).

[0063] Motion blur direction entropy O: describes the degree of disorder in the distribution of motion blur directions (value 0-1), the larger the entropy value, the more dispersed the directions;

[0064] Temporal resolution matching degree P: The matching coefficient between the drone's flight speed and the camera's frame rate (value ranges from 0 to 1). When P=1, there is no blur.

[0065] Motion blur frequency response attenuation value Y: the attenuation ratio of high-frequency components in the image frequency domain (value ranges from 0 to 1).

[0066] Calculate the actual blurriness X of the image based on the above blurriness parameters, and then use the formula... ,in These are weighting coefficients used to balance the influence of each parameter on the final ambiguity;

[0067] The preset ambiguity threshold is When X≤ If the image blurriness is determined to meet the application requirements, the flight speed is increased until the desired blurriness is achieved. , This is a margin for ambiguity (5%–10%); when > At this point, it is necessary to reduce the flight speed to reduce ambiguity until... ≤ ;

[0068] The trajectory of the drone during the horizontal phase is as follows: the radius is the sum of the distance from the building surface to the center of the circle (the radius of the circle at that altitude) plus the shooting distance from the drone to the building surface, and the drone performs horizontal circling shooting; as the altitude increases, the trajectory radius also decreases.

[0069] The height of the ascent phase is measured by the vertical pixels of the image captured by the camera. To confirm, the specific steps are as follows:

[0070] The actual physical vertical length covered by a single image on the surface of a giant spherical building is B = ϵ × The preset overlap rate is e, and the actual vertical length of a single image and the height between layers are B×(1-e). During the ascent phase, the drone camera must always be perpendicular to the surface of the giant spherical building.

[0071] As the drone ascends vertically to another altitude level, its shooting distance increases, and the image sharpness index captured at that current distance is compared to... By comparing and dynamically adjusting the drone's shooting distance, when... Maintain the current distance, if It is necessary to shorten the shooting distance to improve the sharpness index until the sharpness index is greater than the sharpness index threshold;

[0072] When shooting the top, the vertex area needs to be covered. The maximum upward height is the radius of the giant spherical structure plus the shooting distance and the additional coverage height. The additional coverage height is set to... This is used to compensate for curvature deviations in the top arc structure, ensuring the integrity of the top.

[0073] Each time the drone ascends to a new altitude, it recalculates the current shooting distance and repeats the horizontal phase motion trajectory described above.

[0074] Images of all the giant spherical structures captured by drones are recorded as the original image data of the giant spherical structures.

[0075] The visual analysis module includes a preprocessing unit and a defect pattern recognition unit;

[0076] The preprocessing unit is used to preprocess the raw image data, including filtering and denoising, distortion detection, reflection recognition, and multi-view fusion algorithms. The specific operation steps are as follows:

[0077] Noise standard deviation calculation: The noise denoising unit constructs the noise standard deviation U based on the pixel dimension x and a preset constant term; the pixel dimension x refers to the spherical latitude coordinates of the giant spherical building surface corresponding to the image pixel (with the center of the sphere as the origin, the equator as 90°, and the poles as 0°), which is calculated by matching the UAV GPS positioning data with the three-dimensional model of the giant spherical building.

[0078] Median filtering: A 3×3 neighborhood median filtering algorithm is used to replace the current pixel value with the median value of the pixels in the neighborhood.

[0079] Adaptive denoising control: Based on the dimensional distribution characteristics of noise in the spherical image, a preset deviation threshold is established. Dynamically adjust the value of i: when the value of i increases to i≥U+ (This means that noise caused by pixel latitude changes in this area is significant), the system automatically increases the filtering strength, expands the filtering window strength, or increases the number of filtering iterations; when the value of i decreases to i≤U- (This means that the pixel latitude change in this area is gradual and the noise features are weak.) The system automatically reduces the filtering intensity and the number of filtering iterations to balance the denoising effect and detail preservation. Here, i is the noise evaluation index, i=U / U0, U is the noise standard deviation of the current area image (calculated based on the gray value of 3×3 neighboring pixels), and U0 is the preset baseline noise standard deviation.

[0080] Detecting geometric distortion: For the denoised image of the giant spherical building, detect its polar stretching coefficient and angular distortion rate. The specific operation method is as follows:

[0081] Select a small area at the top or bottom of the giant spherical building and calculate its actual physical area; then match the area to the photograph and calculate the area in the photograph. Divide the area in the photograph by the actual area to obtain the stretching coefficient; set a threshold (0.8 to 1.2). If the stretching coefficient is greater than 1.2, it indicates significant stretching distortion; if the stretching coefficient is less than 0.8, it indicates compression distortion.

[0082] On the surface of the giant spherical building, the meridians and parallels are originally perpendicular at 90°. Find the intersection of these two lines in the photo and measure the actual angle in the photo. The ratio obtained by (photo angle - 90°) ÷ 90° is the angle distortion rate. Set a threshold (5%), then an angle that should be 90° will become less than 85° or more than 95° in the photo, which is considered to be an angle distortion.

[0083] The method for correcting image distortion of giant spherical buildings is as follows:

[0084] The parameters that affect the distortion of images of giant spherical buildings include: projection center deviation (Δζ, Δη) and projection radius coefficient. Angular deformation coefficient Δρ;

[0085] Projection center deviation: Select three feature points with known precise geographic coordinates (such as the sphere's apex or pre-defined marker points) on the surface of the giant spherical building; based on the difference between the theoretical projected coordinates and the actual image coordinates of the feature points, calculate the average value (Δζ, Δη) of all feature point differences using the least squares method; measure the actual coordinates (ζ, η) of the corresponding points in the image; and perform translation correction on the coordinates using the formula... , Calculate the corrected projection plane coordinates to eliminate distortion caused by center offset;

[0086] Projection radius coefficient: Five feature points are selected on the surface of the giant spherical building from the projection center to the edge. The actual projection radius q (pixel distance from the point in the image to the center) of each point is calculated; the theoretical geocentric angle t (based on geographic coordinates) is calculated as follows: The theoretical coefficient under ideal equal area conditions By comparison (where R is the radius of the giant spherical building), determine the deviation value Δm = Then, scale q for all projected points proportionally, and correct the formula as follows: ; to make the revised Strictly equal to the theoretical coefficient This eliminates distortion caused by the projection radius coefficient;

[0087] Angular distortion coefficient: Select a characteristic line (orthogonal meridian) with a known included angle on the surface of the giant spherical building, measure its actual included angle θ in the projected image, and calculate... (ρ is the theoretical included angle of 90°); Rotation correction is performed on each pixel: The pixel position is adjusted based on the corrected angle to eliminate distortion caused by the angle deformation coefficient;

[0088] Reflection detection unit: Detects the average grayscale value of the giant spherical building image after distortion processing. Peak brightness The specific operating steps are as follows:

[0089] For the denoised image of the giant spherical building, the average grayscale value and peak brightness are detected, with a preset threshold for the average grayscale value. Peak brightness threshold is When the measured image > and > Determine image reflection;

[0090] When an image has a reflection problem, the specific steps to solve it are as follows:

[0091] The geographic coordinates (longitude, latitude, and altitude) of the drone at the time of the shot are obtained by parsing the metadata of the reflected image, and these coordinates are defined as the drone reshoot positioning point;

[0092] Control the drone to re-capture images of the same reflective area from different angles; when the reflective point is located at the top of the giant spherical building's surface, control the drone to lower its flight altitude so that the angle between the camera's optical axis and the normal to the giant spherical building's surface is adjusted to 30° for shooting; if the reflective point is located on the side of the giant spherical building's surface, control the drone to move 15° left and right and 15° up and down, capturing 2 images from each direction for a total of 8 images, and obtain the image sharpness index of the acquired images. The higher the image sharpness index, the clearer the image, and select the clearest image as the image of this area of ​​the giant spherical building's surface; if the reflective point is located at the bottom of the giant spherical building's surface, control the drone to move along the circumference of the bottom of the giant spherical building's surface, changing the shooting angle for shooting.

[0093] The preprocessed image is then fused from multiple perspectives. The specific steps are as follows:

[0094] Mismatched points are removed: rotation-invariant feature points are extracted using the SuperPoint deep neural network. Based on the radius of the giant sphere and camera parameters, the maximum allowable pixel displacement corresponding to the design arc length is calculated. A 2-pixel displacement threshold is strictly enforced in the equatorial region and relaxed to 5 pixels in the polar region. Mismatched points with displacement exceeding the limit are removed using the RANSAC algorithm.

[0095] Homography matrix calculation: Calculate the homography matrix between adjacent images. , where g is the plane homography matrix and j is the spherical rotation correction matrix, which projects the local image onto a three-dimensional reference plane with the center of the sphere as the origin, thus solving the image stretching problem in the equatorial region;

[0096] Overlapping region fusion: For any pixel p within the overlapping region, S1 is the distance from p to the left boundary of the left image, S2 is the distance from p to the right boundary of the right image, and the weight of the left image is... The weights in the right figure are A preset threshold for the difference between the pixel values ​​of two images is used. When the difference between the pixel values ​​of the two images exceeds the threshold, the pixel value of the image with the larger weight is directly selected for weighted fusion. ,in , The pixel values ​​at corresponding positions in the two images are used to generate a fused image without obvious texture stretching or distortion.

[0097] The defect pattern recognition unit is used to detect cracks, corrosion, and deformation in parallel through dedicated channels on the pre-processed image. The operation steps are as follows:

[0098] Crack detection: For preprocessed images of giant spherical buildings, detect the length-depth product ratio Z of crack pixels in the image (i.e., the ratio of the actual crack length Q to the allowable crack length). The ratio of the actual crack depth D to the allowable crack depth, multiplied by the ratio of the crack depth D to the allowable crack depth. The ratio of the allowable length and depth of the crack is normalized using the formula: Z = (The allowable length of the crack is 200mm and the depth is 5mm). When Z>0.01, it is judged as a crack.

[0099] Corrosion Detection: The preprocessed image of the giant spherical building is processed using a corrosion detection integrated color space conversion unit. Corrosion features are filtered through preset hue and saturation dual thresholds. Specifically, in the corresponding color space, areas that meet the preset hue H value range of 8° to 20°, saturation S≥0.85, and lightness V≥0.7 are selected. In the CIE Lab color space, areas with a color difference ΔE≤8 from the standard values ​​(L*=65, a*=35, b*=45) are further selected. When the image meets all of the above requirements, it is determined to be a corrosion area.

[0100] Deformation detection: Similarity index detection is performed on the preprocessed image. The specific operation is as follows:

[0101] Acquire a high-resolution image of the giant spherical building in its deformation-free state (a high-resolution image of the building at the time of manufacture). Extract a 100×100 pixel mesh texture region with uniform and defect-free surface texture from the image as a standard template. Slide the standard template across the image in 5-pixel increments and calculate the similarity index between each window and the template. y represents the grayscale value of a pixel in the template image. v is the average grayscale value of the template pixels, and v is the grayscale value of the window image pixels. The NCC value is the average grayscale value of the window image pixels; the closer the NCC value is to 1, the higher the similarity; the preset NCC threshold is 0.8, and windows with NCC values ​​below 0.8 are marked as deformed areas.

[0102] The risk classification and resource disposal module includes a risk classification unit and a resource mobilization unit;

[0103] Risk classification unit: Based on the image recognition results, the defects are classified into risks. The specific operation is as follows:

[0104] Cracks are classified into risk levels based on the length-depth product ratio Z of the crack pixels in the image: low-risk cracks (0.01 < Z ≤ 0.025), medium-risk cracks (0.025 < Z ≤ 0.1), and high-risk cracks (Z > 0.1).

[0105] Corrosion is classified into risk levels based on the percentage of corroded pixels in the total number of pixels in the image: low-risk corrosion (pixel percentage ≤ 5%), medium-risk corrosion (pixel percentage 5% to 15%), and high-risk corrosion (pixel percentage > 15%).

[0106] Deformation is classified into risk levels based on the similarity index in the image: low risk deformation (0.8≤NCC<1.0), medium risk deformation (0.6≤NCC<0.8), and high risk deformation (NCC<0.6).

[0107] Resource Allocation Unit: Based on the risk level and defect information output by the risk classification unit, formulate and execute corresponding resource allocation plans. The specific operations are as follows:

[0108] Establish repair principles: prioritize repairing high-risk defects, followed by medium-risk defects, and then low-risk defects. If multiple defects have the same risk level, prioritize repairing defects located in critical areas (the connection area between the bottom of the giant spherical building and the outer surface, and the projection area of ​​the main load-bearing frame on the surface). If defects of the same level appear in the same location, prioritize repairing cracks, then deformation, and finally corrosion.

[0109] Receive information such as risk level, defect type, and defect location output by the risk classification unit;

[0110] Based on the risk level and corresponding resource allocation rules, determine the type and scale of resource allocation;

[0111] At low risk levels, a basic testing team (2-3 people, including operators and surveyors) and standard testing tools (such as high-definition cameras and measuring tapes) are required.

[0112] At the medium-risk level, a professional maintenance team with standard configuration (4-6 people, including corrosion prevention personnel and crack repair technicians) and specialized maintenance equipment (such as crack repair tools and rust removal equipment) will complete on-site handling within 7 working days;

[0113] In cases of high risk, a team of technical experts (6-8 people, including structural engineers and materials experts) will arrive on site within 48 hours and develop a response plan.

[0114] Based on the risk level, resource mobilization routes are planned. When a high-risk level occurs, resources are directly mobilized from the nearest resource pool without considering economic factors, taking the shortest route (closer distance, no congestion). For medium-risk levels, resources and routes occupied by high-risk resources are avoided, and other resources and routes are used (these resources and routes can be used after the high-risk resources have been mobilized). For low-risk levels, idle or reserve resources are used to reduce costs, and the operation is initiated after high or medium-risk levels occur.

[0115] After the repair is completed, based on the defect characteristics, the acceptance thresholds are defined according to the type (crack product ratio ≤ 0.01, rust pixel ratio ≤ 1%, deformation similarity index ≥ 0.9); images of the repaired surface of the giant spherical building are collected by drone, and the images are inspected. If the inspection results meet the acceptance thresholds, a repair acceptance signal is generated.

[0116] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the scope of the invention as defined in the claims. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The invention is defined by the claims and their equivalents.

Claims

1. A visual analysis system for mega-sphere architectural images, characterized by, The image acquisition module, the visual analysis module, the risk classification and resource disposal module; The image acquisition module acquires clear original image data of the surface of the giant ball building by dynamically adjusting the shooting distance, the flight speed, the shooting frequency and the rising height of the unmanned aerial vehicle; and sends the collected information to the visual analysis module; The visual analysis module receives the original image data of the image acquisition module, pre-processes the original image data, including filter denoising, distortion detection and reflection detection; detects the value corresponding to the parameter of the pre-processed image; if the value is not within the preset threshold, detects the defect of the image to obtain the type of the defect; The type of the defect is sent to the risk classification and resource disposal module; The risk classification and resource disposal module receives the defect of the visual analysis module, classifies the risk level of the defect and allocates the corresponding resource according to the risk level, and finally confirms whether the repair effect meets the acceptance threshold through the unmanned aerial vehicle re-inspection to generate a qualified signal; The specific operation process of the image acquisition module adjusting the shooting distance of the unmanned aerial vehicle is as follows: When adjusting the shooting distance of the UAV, the reference resolution N, the contrast J, the sharpness T, the brightness U and the pixel density G are used to determine the definition index C according to the formula ; wherein is a weight coefficient, used to balance the influence of each parameter on the final definition; , , , and are the resolution reference value, the contrast reference value, the sharpness reference value, the brightness reference value and the pixel density reference value respectively; when the definition index C is greater than a threshold value , the shooting distance is increased until the definition index , is the margin of the definition index; when the definition index is lower than the threshold value, the distance is shortened until the definition index is greater than the threshold value. The specific operation process of the image acquisition module determining the flight speed and the shooting frequency of the unmanned aerial vehicle is as follows: When the flight speed of the unmanned aerial vehicle and the number of photographing times are determined, the actual effective width of a single photo is calculated as e×W×(1-p) according to the image pixel width W, the lens visual conversion coefficient e and the preset overlap rate e, and the number of photographing times f is determined in combination with the circumference length L of the horizontal wrap-around track ; The flying speed determines the blurring degree X according to the blurring pixel offset A, the dynamic contrast attenuation rate F, the motion blurring direction entropy O, the time resolution matching degree P, and the motion blurring frequency response attenuation value Y according to the formula: X= , wherein is a weight coefficient; the speed is increased when the blurring degree is lower than a threshold value until the blurring degree , is a margin of the blurring degree, and the speed is decreased when the blurring degree is greater than a threshold value until the blurring degree is less than the threshold value; The specific operation process of the image acquisition module determining the rising height of the unmanned aerial vehicle is as follows: The ascending height is determined by the vertical pixels of the image taken by the camera The actual physical vertical length of the single image covering the surface of the giant ball building is The preset overlapping rate is e, and the actual vertical length of the single image and the height between layers is B x (1-e).

2. A system for visual analysis of mega-sphere construction images according to claim 1, characterized in that, The specific operation process of the visual analysis module filter denoising is as follows: Filter denoising calculates the noise condition according to the pixel dimension, adopts 3*3 neighborhood median filter to replace the current pixel value with the median value of the neighborhood pixels; The filter strength is adjusted adaptively, when the value of i increases, the window is expanded or the iteration is increased to improve the filter strength, when the value of i decreases, the filter iteration times are reduced to weaken the filter strength; Wherein, i is the noise evaluation index, i=U / U0, U is the noise standard deviation of the current region image, U0 is the preset reference noise standard deviation.

3. A system for visual analysis of mega-sphere construction images according to claim 1, characterized in that, The specific operation process of the visual analysis module distortion detection is as follows: The distortion detection judges whether distortion exists through polar stretch coefficient and angle distortion rate, and then corrects from projection center deviation (Δζ, Δη), radius coefficient , and angle deformation coefficient Δρ. Projection center deviation correction: calculate the deviation (Δζ, Δη) through 3 feature points, measure the actual coordinates (ζ, η) of the corresponding points in the image, and perform translation correction on the coordinates. The correction formula is: ; ; Projection radius coefficient correction: the actual projection radius q of each point is calculated by 5 feature points, and the theoretical geocentric angle t is calculated, and the ratio is the theoretical coefficient under the ideal equal-area condition In comparison, the deviation value Δm is determined , and the q of all projection points is scaled in proportion, and the correction formula is: ; Angle deformation coefficient correction: through the known angle of the characteristic line, measure its actual angle θ in the projection image, calculate Where ρ is the theoretical angle 90°, and the rotation correction formula for each pixel point is: .

4. A system for visual analysis of mega-sphere construction images according to claim 1, characterized in that, The operation process of the visual analysis module distortion reflection detection is as follows: Reflectivity detection is based on average grayscale value and peak brightness. When the average grayscale value of the image is greater than 100%, the reflection is detected. And peak brightness > At that time, determine the image reflection; When there is reflection, the unmanned aerial vehicle is used to re-collect the reflection area at different positions, when there is reflection on the top, the flight height is reduced to make the angle between the camera optical axis and the normal line be 30°; when there is reflection on the side, 8 images are taken by shifting up, down, left and right by 15°, and the image with the highest definition index is selected; when there is reflection on the bottom, the unmanned aerial vehicle is controlled to shift along the bottom circumference of the giant ball building surface to change the shooting angle.

5. A system for visual analysis of mega-sphere construction images according to claim 1, wherein, The visual analysis module further comprises: Crack detection: detect the length-depth product ratio Z of crack pixels in the image, that is, the ratio of the actual length Q of the crack to the allowable length , multiplied by the ratio of the actual depth D of the crack to the allowable depth , normalize the allowable length and allowable depth of the crack, and pass through the formula: Z= ; when Z>0.01, it is determined that there is a crack; Rust detection: integrated color space conversion is adopted to screen the area with hue H=8°-20°, saturation S≥0.8 and color difference ΔE≤8, and determine it as rust; Deformation detection: take the standard template as the benchmark, detect the similarity index NCC judgment; through the formula , y is the pixel gray value of the template image, is the average gray value of the template pixel, v is the pixel gray value of the window image, is the average gray value of the window image; the window with NCC value <0.8 is marked as the deformation area.

6. A system for visual analysis of mega-sphere construction images according to claim 1, wherein, The specific operation process of the risk classification and resource disposal module is as follows: Crack: according to the length-depth product ratio Z of crack pixels, 0.01 Rust: according to the percentage of rust pixels in the total image pixels, rust pixels account for ≤5% is low risk, 5% Deformation: According to the similarity index NCC value, 0.8<=NCC<1.0 is low risk, 0.6<=NCC<0.8 is medium risk, and NCC<0.6 is high risk.

7. A system for visual analysis of mega-sphere construction images according to claim 1, characterized in that The risk classification and resource handling module further comprises: In terms of resource handling, corresponding allocation schemes are adopted for different risk levels: Low-risk defects are handled by a basic team of 2-3 people, equipped with cameras and tape detection tools; Medium-risk defects are handled by a professional team of 4-6 people, carrying crack repair tools, rust cleaning equipment, and maintenance equipment, and need to be completed within 7 working days; High-risk defects are handled by a team of 6-8 technical experts including structural engineers and material specialists, who are required to arrive at the scene within 48 hours, with the resource mobilization prioritizing the nearest resource library and the shortest time path, regardless of economic cost factors; During repair acceptance, the acceptance threshold is clearly defined according to the defect type: the crack needs to satisfy the length-depth product ratio Z<=0.01, the rust needs to satisfy the pixel ratio <=1%, the deformation needs to satisfy the similarity index NCC>=0.9, and the qualified signal is generated after the UAV re-inspection confirms compliance. ​

Images